Understanding the spatio-temporal variability of vegetation phenology is necessary to study the timing of recurring biological events over large areas and different types of crops. Currently, with the long-term and frequent satellite observations (e.g., Sentinel-1, Sentinel-2 an Sentinel-3), it is possible to monitor the vegetation phenological characteristics such as the start and end of the growing season, maximum peak, duration, etc. Over time, several sophisticated software packages have been developed to quantify vegetation cycles using time-series of vegetation indices products. Well-known examples include BFAST [1] and TIMESAT [2]; both of them are freely available to the scientific community.

In this study, we present a new time series toolbox that serves as alternative to these existing toolboxes. The So-called Decomposition and Analysis of Time Series Software (DATimeS) is an image processing GUI toolbox written in MATLAB. DATimeS has been recently developed by the research group Laboratory for Earth Observation at the Image Processing Laboratory (IPL) of the University of Valencia (UV). The toolbox enables to analyze different complex and advanced time series tasks (1) for reproducing spatially continuous fields from discontinuous data, i.e. gap-filling methods have been implemented using advanced machine learning regression algorithms (e.g. Gaussian Processes Regression) and harmonic analyses (Fourier Transform); and (2) for the detection of complex spatial patterns of the phenological parameters, a double sigmoidal function with a non-linear least square regression using a sliding window approach has been developed.

From a practical perspective, all these methods require minimal user intervention to optimize the algorithm parameters that define the model architecture. A valuable feature of DATimeS is that time series data should not be evenly spaced over a specific period. Another advantage is that intuitively complete sentinel images, specific region of interests and single pixels can be processed in multiple formats. As an application, here we show first DATimeS results of seasonal trends. For that purpose, LAI and NDVI time series derived from Sentinel-2 and Sentinel-3 were analyzed on seasonal phenology trends. As a next step, harvest trends and disturbances can be detected and quantified.
In conclusion, DATimeS pretends to be a user-friendly and powerful time series toolbox that allows to analyze any kind of temporal data. The toolbox will become freely available to the community as part of the ARTMO (http://ipl.uv.es/artmo/) software framework.

[1] Verbesselt, J., Hyndman, R., Newnham, G., and Culvenor, D. (2010). Detecting Trend and Seasonal Changes in Satellite Image Time Series. Remote Sensing of Environment, 114(1), 106-115.
[2] Jönsson, P., and Eklundh, L. (2004). TIMESAT - a program for analysing time-series of satellite sensor data, Computers and Geosciences 30, 833-845.

Area A - Board 1

Multilabel classification of optical multispectral and Synthetic Aperture Radar (SAR) imagery has been applied in Remote Sensing for many years, especially for land use and landcover mapping and in many other research projects for the delineation of objects and analysis of spatio-temporal patterns. In the real world however, very often there is only one target class with corresponding labels based on in-situ observations. A negative class is neither asked for nor are labels for any other semantic class provided and a decision boundary between multiple classes cannot be drawn. This is when the application of one class classification algorithms (OCC) is required. In this study, we tested the plausability of OCC algorithms such as One-Class Support Vector Machine (OC-SVM), Isolation Forest (IF), Lokal Outlier Factor (LOF) and Elliptic Envelope (EE) for the operational deployment in our agricultural monitoring platform ag|knowledge. The platform uses multitemporal Remote Sensing imagery to create object-based time series for each agricultural parcel. It further provides the possibility to register agricultural parcels in a spatial database, labeled with the current crop type as training data. Thus, the training set and the classification model can be updated constantly in an online learning manner. The purpose of the system is to detect outliers in the training data as many „ground truth“ and in-situ measurements might be contaminated with erroneous sample data. It then predicts whether new and unlabeled parcels belong to the trained class or not. These predictions are done for new growing seasons and new observation years of the target crop, thus, the model transferability between different time frames and years is crucial. Based on Sentinel-1 and Sentinel-2 data, time series signatures of SAR and Vegetation Index features for different crops were learned by the model and its transferability was tested on consecutive years. To ensure the same dimensionality, i.e. same number of observations over the years (or growing seasons), dimensionality reduction through temporal aggregation of statistical features was conducted. We can conclude that the IF performed best for outlier detection and the OC-SVM with a radial basis function kernel achieved the most satisfying results for novelty detection. Best results were yielded based on fused Sentinel-1 and Sentinel-2 data in terms of algorithm precision and performance whereas the best choice for the temporal aggregation were bi-weekly statistical features. Critical for accurate predictions is the consideration of different growth stages and varieties of the respective crop type.

Area A - Board 2

Many studies on crop type classification from Remote Sensing imagery exist with a different set and number of classes. The classification accuracy often varies and is highly dependent on the quality of the training data, the set of features and the number of classes. Usually, a large number of classes reduces the classification accuracy and causes a higher inter-class confusion and entropy of the posterior class probabilities, especially for classes with a similar phenology and similar spectral characteristics. In this study we demonstrate different strategies to optimize the training of classification algorithms used for crop type discrimination with multispectral Sentinel-2 time series data. We addressed the question of how many categories, i.e. semantic classes, are useful and necessary to generate a well-performing classification model without a too large loss of categorical information. We examined the semantic resolution requirements based on a large, imbalanced, labeled parcel set of 74 crop type classes. We tested the regrouping and the aggregation of categories based on their phenological characteristics, crop growth stages, agronomic practices, geographical location and biophysical parameters such as leaf orientation and crop canopy height. We assessed similar classes through similarity metrics such as Euclidean Distance, Cosine Similarity and Pearson Correlation and permuted different classes in order to find the optimum combination through reduction of categories, e.g. winter wheat, winter barley, winter rye and durum were grouped within a new category of winter cereals. This semantic generalization leads to the development of a new set of classes with a higher inter-class separability as well as to an overall increase of recall and precision for certain crop classes. Some classes however needed to be assigned with weights due to their predominance in the training data set and should not be generalized. Furthermore, we tested different feature selection and sampling strategies from the provided training data and evaluated the classification performance with two different machine learning classifiers, a Gradient Boosting Trees (GBT) and a Multilayer Preceptron (MLP) algorithm. Both are capable to cope with non-linear class separation problems based on high dimensional data with a high semantic resolution. The analysis contributed to the establishment of a pre-trained model for the crop type classification through an analytical and cloud-based agricultural monitoring platform ag|knowledge.

Area A - Board 3

The key for the effective biomass production from the short rotation coppice (SRC) is the agrotechnical supervision of plantations allowing for a control of plantation growth and readiness for harvesting, and a fast response to sudden changes in plants' condition caused by diseases or temporal flooding. Due to the very high variability of the structure of plants over time, a successful monitoring of energy crops has to be based on systematic data acquisitions delivering up-to-date and complex information of crop characteristics. In this study we propose a fusion of remote sensing data: high resolution multispectral imagery from Sentinel-2, radar data from Sentinel-1 and low-altitude (UAS) RGB and hyperspectral imagery for the evaluation of plantation extent, possible crop losses, plant vigor, height and density, as well as the estimation of nutrient content. A complex analysis of multitemporal set of Sentinel-2 and Sentinel-1 images was performed to evaluate crop condition for the selected test plantations of energy willow (Salix). Vegetation indices computed from the S-2 images were used to study interannual variations in the vegetation phenology and plant vigor. For the on-site verification of the results, as well as for the detailed analysis of crop condition over the areas where a detailed description of a plantation was needed, hyperspectral measurements obtained from a hyperspectral camera mounted on UAV (Unmanned Aerial Vehicle) were used. The hyperspectral measurements carried out for several willow varieties showed that they had different content of leaf pigments, mainly chlorophyll, at the same moment of observation. Extracted spectral indices for texturally homogeneous crops revealed the areas of senescent plants. Complementary RGB photos permitted to extract a 3D information on crop height and identify crop losses. Additionally, for the areas characterized by homogenous vegetation indices calculated from S-2 data spatial statistics of indices based on UAS data were computed to better understand the relation between satellite measurements and local field conditions, as well as other crop characteristics. The methodology presented in this study can be used for general inspections of SRC on spatially large areas and locally as a tool for agrotechnical supervisions of crop growth conditions and potential biomass losses for a given field.

Area A - Board 4

The unprecedented combination of Sentinel-2 (S2) and Landsat-8 (LC8) provides an opportunity to monitor the land surface at high resolution. Different sensors have different spectral, spatial and temporal characteristics, and the combined observations need to be carefully interpreted in order to produce inferences of land surface parameters.

In this contribution, we present a sensor-agnostic approach for the optimal combination of observations from different sensors, with an emphasis on S2 and LC8. The method has been developed under the EC-funded H2020 MULTIPLY project. The approach is based on the inversion of physical models that describe the photon-soil-canopy-leaf interaction based on radiative transfer theory. The inversion of these models is hampered by the ill-posed nature of the problem, which results in large uncertainties in the retrieved parameters. These can be somehow mitigated by using relevant prior information.

We show what forms this prior information can take (from user-defined parameter a priori distributions, to dynamic parameter models), and how the models and the prior information can be optimally combined with the observations to produce inferences on land surface parameters such as leaf area index (LAI), leaf pigments (e.g. chlorophyll, carotenoids, etc.) or leaf components (moisture, dry matter, ...). The approach is general, and a discussion on pragmatic decisions to make it applicable to large volumes of data available from S2 and LC8 will be given.

We will demonstrate the approach on monitoring croplands in Germany over a whole growing season. In this case, we will use the combined PROSPECT+SAIL radiative transfer (RT) models (for leaf and canopy properties, respectively) to interpret the surface reflectance. The surface reflectance from S2 and LC8 has already been pre-processed with the SIAC atmospheric correction scheme, which ensures radiometric consistency between LC8 and S2 observations, as well as uncertainty estimates of surface reflectance. The RT models are used through emulators (Gómez-Dans, Lewis, & Disney, 2016) for computational efficiency. Temporal dynamic models are used for some of the parameters (e.g. LAI and chlorophyll), and the results of the inversion will be compared with a multi-parameter set of ground measurements. Some final discussion on the implications of these methods and their use in data assimilation approaches to crop monitoring will also be provided.

Area A - Board 5

Rice is a major food crop for the people of the world in general and Asians in particular; nearly 90% of the world's rice is produced and consumed in this region. With an increasing world population and a significant imbalance between rice supply and demand since the last few years, the situation is prone to create tensions in food markets that could lead to world food price crises, as in 2008 when the price of rice more than doubled in only seven months. In this context of price instability and threatened food security, tools to timely monitor rice production are highly needed by governments and decision makers. Accurate information is needed on the spatial distribution of rice fields, growth status, risk occurrence and annual production projections. Moreover, rice agriculture is strongly linked to environmental issues, from water resource to methane emissions. For these reasons, long term inter-annual monitoring is also required in order to study the environmental impacts of rice cropping density and cultural practices.
The potential of satellite SAR data for rice monitoring has been widely demonstrated in past studies, but without systematic and cost effective data, applications were difficult to achieve. With the dense time series of Sentinel-1 accessible worldwide and at no cost, rice monitoring should be posible, if robust and fully validated methods are available.
The objective of this paper is to develop tools for monitoring the rice cultivation systems, by generating products including map of rice planted area, map of rice start-of-season, and maps of phenological stage, and map of rice cropping intensity, together with rice crop parameters such as rice varieties (long or short cycle), and plant height. The information to be provided is necessary for the estimation of crop production, and for the management of rice ecosystems at the regional scale.
This paper presents our study using Sentinel-1A &B, IW data over the Mekong Delta (300 km x 300 km), at 10m resolution and repeat cycle of every 12 days from October 2014 and every 6 days since October 2016 to observe the temporal evolution of rice cultivation. In this region, one of the world’s major rice regions, rice is grown with 2- 3 crops a year, with crop calendar, rice varieties and water management changing from one region to another, and even from one field to the next. In this context, physical based methods (rather than statistical based ones) are required to deal with the diversity of rice field conditions.
The first step is to understand the temporal variation of the backscatter Sentinel-1 backscatter of rice fields, at VH and VV polarizations, and at incidence angle from 29° to 46°. For this purpose, in-situ data have been collected on 60 fields during 2 years, for the 5 rice seasons. It was found that backscatter time series of rice fields show very specific temporal behavior, with respect to other land use land cover types. The temporal and polarization variations of the rice backscatter have been interpreted with respect to physical interaction mechanisms to relate the backscatter dynamics to the key phenological stages, when the plants change its morphology and biomass. Because the same trend of temporal curves was observed over 5 rice seasons, it was possible to derive a specific ‘mean’ curve to be used in the methodology developed for detecting rice phenology, and deriving information such as the date of sowing, the rice varieties of long and short duration cycle, or plant height, at each SAR acquisition date.
The algorithms have been developed and applied to the entire Mekong delta. Product validation based on independent ground data provides good agreement with the reference data sets: 98% in rice/non-rice accuracy, the sowing dates RMSE of about 4 days, plant height RMSE of 7.8 cm, the long/short variety map has 91.7% accuracy and for phenology, only one season has been processed with good detection rate of 98%.
The methods will be discussed regarding their applications to other rice ecosystems (rainfed, upland rice.), and for a diversity of cultural practices. Finally, the approach to integrate the rice products derived from Sentinel-1 data in process-based models for rice production estimation and methane emission estimation will be exposed

Area A - Board 6

Approx. 10% of the global land surface is covered by agriculture and underlies high temporal dynamics through annual and inter-annual variations in crop type and areal extent. Time series of satellite images are therefore needed at a high spatial and temporal resolution specifically to improve monitoring of crop type, plant health, agricultural management and yield. The choice of the spatial resolution is a trade-off between amount of data and cost as well as loss of fields, which disappear through lowering spatial resolution. To resolve in field heterogeneities of crop growth and thereby open the promising field of precision agriculture with its high environmental and economic potentials sensor resolution should be at a fraction of the sizes of the agricultural fields. Field sizes vary considerably globally depending on the level of mechanization of the agriculture and on the cultural background.
The sensors of the Sentinel-2 satellite mission as part of the Copernicus Program, developed for land surface monitoring and land use cover change, have a spatial resolution in the decisive wavelengths for plant growth monitoring between 10 m and 20 m. They create rasterized images of the land surface in which distinct objects like agricultural fields are split into pure field member pixels and mixed pixels at the borders of the objects. As soon as mixed pixels predominate the assessment of in-field heterogeneity becomes impossible. No in depth, real world analysis on the exact fraction of fields as well as areas lost through Sentinel-2’s spatial resolution currently exists.
The paper presents a systematic assessment of rasterized vector field maps at spatial resolutions ranging from 5 m (possibly improved future Sentinel-2), 10 m (current Sentinel-2, VIS-NIR) to 20 m (current Sentinel-2 SWIR). Bavaria, Germany (circa 70 000 km²) was chosen as a representative Central European agricultural landscape. The vector field data set contains accurate information on the field boundaries, field sizes and crop types for approx. 1.9 million fields in Bavaria covering roughly 1/3rd of its area (average field size 1.6 ha). The fields were rasterized via the Bresenham-Algorithm to 5, 10 and 20 m resolutions creating only pixels strictly inside the vector boundaries. Their field sizes are determined from the rasterized data sets and compared to the field sizes in the original vector data set. Differences between vector and raster field sizes are, together with absolute field pixels numbers, measures for the suitability of the rasterization resolution for the analysis of in-field heterogeneities of the agricultural fields. Results of the statistical analysis of the rasterized data sets will be shown and discussed.
Preliminary results show that under real world conditions in a representative Central European landscape a considerable number of agricultural fields cannot be accessed with Sentinel-2’s current resolution. An increase of spatial resolution to 5 m would considerably improve the coverage of Bavaria’s agricultural fields by a future Sentinel satellite fleet.

Area A - Board 8

The Horizon2020 research project “Evolution of Copernicus Land Services based on Sentinel data” (ECoLaSS) aims to improve existing and develop next-generation Copernicus Land Monitoring Service (CLMS) products, with a targeted implementation schedule from 2020 onwards, focusing mainly on the High Resolution Layers (HRLs). Innovative methods and algorithms, based on dense time series processing of optical (Sentinel-2) and SAR (Sentinel-1) data, are applied to develop, test and demonstrate various prototypes of improved or novel HRLs. A special focus of the presented work is laid on the recently established prototypes for potential future HRLs on agriculture/arable land, which are based on a combination of dense optical and radar time series. These prototypes are able to both, distinguish between cropland and non-cropland, and further discriminate crop type classes meaningful for a European level.

Such possible future HRL on agricultural/arable land would be an entirely novel product consisting of a crop mask and crop type maps. Two large demonstration sites concerning seven countries (CENTRAL: Germany, Austria, Switzerland, Liechtenstein and WEST: Luxembourg, Belgium, France) are used to showcase the operational applicability and maturity of the prototypes under significantly varying agro-farming systems and Earth Observation conditions (e.g. cloud cover). The main research questions can be described as follows: (i) What is the thematic accuracy trade-off between crop type monitoring along the season and adequate time intervals for status mapping?; (ii) What is the added value of the integration of Sentinel-1 with Sentinel-2 time series?, and (iii) What are the differences in terms of performances across different European sites? The time series analysis methods and prototypes are conceptualised with a view to a potential future European-wide implementation and fitting into the operational Copernicus service environment, in case of being selected for service roll-out by the respective procuring entity(ies).

The methodological approach for deriving both the crop mask and the crop type map starts with calculating pixel-based time features (i.e. temporal statistical metrics capturing the phenological behaviour) from dense Sentinel-1 and -2 time series, such as the mean value, coefficient of variation, various percentiles, etc., from Sentinel-1 and -2 spectral bands and various derived indices (Sentinel 1: VV/VH ratio, NDVVVH; Sentinel-2: NDWI, NDVI, Brightness, IRECI). Thereafter, firstly a random forest classifier is applied to the full set of those features, which are then further narrowed down in a second step to the most meaningful ones for classification. This feature selection procedure guarantees the most cost- and time-efficient processing whilst ensuring high thematic accuracy and quality levels. The used input data comprise Sentinel-1 (of the year 2017) and Sentinel-2 (of the years 2016 and 2017) imagery. All ECoLaSS prototype implementations are designed to capitalise on the capabilities of state-of-the-art (public/private) cloud processing technologies.

For training of the crop mask (CRM) classification, samples based on the HRL 2015 and CLC 2012 products are used. These are complemented by national Land Parcel Information System (LPIS) data for training/validation of the crop type (CRT) map. The CRM products show a very satisfactory performance with an overall accuracy of >95% for the French part of the WEST site based on Sentinel-2 only, and likewise >95% for the CENTRAL site based on Sentinel-1 and Sentinel-2. Moreover, promising results are obtained also for the CRT products. The accuracy assessments show that the overall accuracy ranges from 77 % (with 13 classes) in WEST/France using only Sentinel-2, over 89 % in CENTRAL (with 16 classes) using Sentinel-1 and Sentinel-2, to 92 % in WEST/Belgium and Luxembourg (with 24 classes) combining Sentinel-2 and Sentinel-1-derived features. Looking at the thematic accuracies of specific classes, high Producer’s (PA) and User’s Accuracies (UA) are typically achieved for classes with frequent occurrence and large field sizes, such as winter crops, winter rape, summer crops and maize, whereas relatively low accuracies are often observed for classes of rare occurrence and/or small field sizes. The resulting class-based PAs range between 46 and 96% in the WEST site, with a comparably UA range (49-96%). For the CENTRAL site, the class-specific PAs range from 49 to 98%, and the UA lies between 21 and 95%. This large range of obtained class-specific accuracy values is deemed to be directly attributable to the quality of training data available from LPIS (or other sources): Typically, the mapping results for rare classes with small field sizes (and thus higher likelihood of mixed-pixel training samples) appear to be less accurate.

Finally, it is worth mentioning that these preliminary results will be further consolidated in the second project phase of ECoLaSS. Specifically CRM and CRT prototypes will be further developed over the WEST and CENTRAL demonstration sites, where it is also envisaged to develop a pan-European classification key of approx. 15-20 classes. Besides that, an additional product, distinguishing between permanent and annual crop types, is planned to be tested. Another aspect that will be analysed in the future is whether or not integrating SAR coherence for the classification of both crop mask and crop types brings further added value. Focus will also be laid on identifying appropriate time intervals for deriving specific crop types in different bio-geographical regions of Europe.

Area A - Board 9

In Portugal, cork oak (Quercus suber L.) stands cover 737 000 ha, being the most predominant species of the montado agroforestry system, contributing for the economic, social and environmental development of the country. Cork oak mortality is a known problem since the late years of the 19th century that has recently worsen. This might be confirmed by the increasing number of authorization applications to cut down dead or diseased trees that have been submitted to the national forestry service (ICNF). The causes for the occurrence of oak decline seem to be a result of slow and cumulative processes although the role of each environmental factor is not yet established. The availability of high spatial and temporal resolution global observations from Sentinel-2 mission, with Sentinel-2A been launched on 23 June 2015, enables the use of dense time series for gradual processes monitoring. These processes can be monitored using different spectral vegetation indices (VI) once their temporal dynamics are expected to be related with green biomass and photosynthetic efficiency. The Normalized Difference Vegetation Index (NDVI), the most well-known visible band index, is sensitive to structural canopy changes, however it tends to saturate at moderate-to-dense canopies. Modified VIs have been proposed to incorporate the reflectance in the red-edge spectral region, which is highly sensitive to chlorophyll content while largely unaffected by structural properties. The main goal of the present work is to assess the use of high spatial resolution Sentinel-2 time series for monitoring cork oak decline in the scope of the GEO SUBER (Cork oak monitoring) project. GEO SUBER, which is funded by the National Paying Agency (IFAP) through the European Agricultural Fund for Rural Development (EAFRD), intends to ensure cork oak vitality periodic monitoring through remote sensing and to provide information to forest farmers about dead or dying trees in an annual basis through an online platform. For this purpose, satellite data acquired from 2016 to the present, were used to generate chlorophyll indices (CIs) and NDVI time series for 2 test sites in central Portugal. In addition, information provided by the farmers on the location and vitality status of several cork oak trees is used for correlation with the CI and NDVI time series at the tree level. Preliminary analysis of the original time series seems to be promising since a slight difference between wealth and dead or dying cork oak trees was observed.

Area A - Board 10

Accurate crop type maps are vital for an agricultural monitoring and sustainable land management. When derived from earth observation data, an accurate classification of crop types requires high spatial and temporal resolutions. Furthermore, many research studies show that the fusion of optical and SAR data increases the classification accuracy compared to single sensor approaches.

With Sentinel-1 and Sentinel-2, optical and SAR data is available freely which meets the spatial, temporal and spectral resolution requirements for crop type mapping over a large area. However, large area mapping requires the derivation of spatially contiguous features from the irregular time series data due to different acquisition footprints, clouds, cloud shadows, etc. Such features can be derived by building a consistent virtual time series, temporal statistical metrics or best-available pixel composites from all clear sky observations. Several studies exist which investigated such features for classifying crop types with optical or SAR data. However, the fusion of optical and SAR data with such features has been rarely investigated. Commonly, previous studies investigated the fusion of optical and SAR data are performed on relatively small study sites and based on a stack of SAR and usually cloud-free scenes.

In this study, we evaluate the suitability of virtual time series and temporal statistical features for the classification of crop types with optical (S2), SAR (S1) and fusion of optical and SAR data. Furthermore, various feature groups, such as the spectral/polarization bands, thereof derived indices, texture features and the combination of these feature groups are evaluated. An in-depth comparative analysis is performed to identify the most important features for the overall classification as well as specific classes (e.g. winter barley) and class groups (e.g. summer cereals).

Our study site is located in Brandenburg, Germany. We use Sentinel-1 and Sentinel-2 acquired over the study site during the year 2017. For the same year, the Geospatial Aid Application (GSAA) data is used as reference data. It is generated by the farmers and comprises parcel boundaries and crop type information for the parcels where the farmers claim subsidies for. It is a nearly full coverage reference dataset and optimal for the study.

Area A - Board 11

Crop monitoring is becoming essential in terms of global food security and availability, agricultural water management and the understanding of the impact of agricultural practices on climate change. The ESA Sentinel-1 and Sentinel-2 missions allow an unprecedented amount of free data, with global coverage and repeated short frequency. The continuity of the Copernicus program offers a unique opportunity to monitor the agricultural landscape on a highly temporal base (every 5 days, even less in northern areas) on decades, providing an excellent dataset for different users and different scales: for farmers at the field level monitoring, for regional, national or European authorities and policy makers for survey and statistics, and as well for scientists and researchers to better understand and follow the crops biophysical parameters at small scale and to evaluate the role of agricultural practices at global scale in climate change adaptation and/or mitigation scenarios.

On this study we analyze the synergy between the Sentinel-1 (C-Band radar) and Sentinel-2 (13 spectral bands) on two different tests sites in France, for a variety of winter and summer crops commonly cultivated in temperate areas (wheat, rapeseed, maize, soybean, sunflower and sorghum). The first test site is in the southwest France and extensive campaigns for field measurements have taken place in the frame of the Sentinels Synergy for Agriculture (SensAgri) project and the JECAM-SAR experiment. The SensAgri project is an H2020 program financed by the European Commission and its goal is to propose prototypes for new Copernicus land services. The second test site is the Aube department (north of France) where field data was also available and a lot of expectation was held by the use of Sentinel-1 (due to the frequent presence of clouds in the area).

For the first test site, the temporal behaviors of the Sentinel-1 GRD images are analyzed on approximatively 2000 fields in a 250km x 250 km area and compared to the NDVI and NDWI indexes from the Sentinel-2 L2A images for 2016 and 2017. Concerning the radar data, there is a high variability of signal backscatter between the different crops but inter-crop as well. The analysis of rain events, satellite incident angle and orbit direction (ascending and descending) is performed to better understand this variability and its impact on the later use of these images on the crop type mapping prototype of the SensAgri project.

For the second test site, the same analysis is performed in an equivalent area, using again the S1 GRD images but as well the coherence of the Sentinel-1 complex images (SLC). The coherence of the S1 SLC images does not improve the crop type mapping classification overall (although it improves some specific crop classes) but it can be used to follow the biophysical development and practices (detect changes such as field plough, crop emergence and harvest) at field level.

Area A - Board 12

Being an effective benchmarking platform, the JECAM network provides a valuable opportunity to study and compare the available datasets for training of supervised classifiers for cropland and crop type mapping. In situ sampling is the most direct way to get reliable information to generate cropland\land use maps, but it is also costly and time consuming to get a spatially representative dataset which still does not allow to retrieve historical information. Visual interpretation of remote sensing imagery and information extracted from reference landcover maps may be regarded as other sources of training data. Spatial representativeness of visual interpretation dataset depends on the number of people involved, which is inversely related to subjectivity issues. Reference landcover maps are not always adequate to the minimum unit size of the analysis or can be out of date. The objective of this experiment is to compare different training data sources for cropland classification at different spatial scales, including: 1) in situ sampling, 2) visual interpretation of remote sensing imagery with multiple participants (crowd source, CS) and 3) available landcover reference maps (LC). In situ training data was obtained by on-road and field survey sampling for 6 contrasted JECAM sites: Argentina, Belgium, Canada, China, Russia and Ukraine for the year 2016, excepting Argentina where the study period was June 2015-May 2016. Visual interpretation of the satellite data was performed by experts related to all JECAM sites involved. Experts were requested to analyze time series of remote sensing imagery (Sentinel-2, Landsat-7,8) in order to draw and label the set of objects arranged with random stratified sampling using a crowdsourcing tool in the VEGA-GEOGLAM interface. It was a two-tier activity, when 1) the accuracy of each expert was estimated using in-situ data, and 2) the crowd source dataset was generated. Thus, the visual interpretation was performed both at small (JECAM area provided with in situ data) and large (4-6 Landsat scenes) scale. The experts were guided by the JECAM definition of cropland when labeling the objects.
Site reference maps were obtained from the Globeland30 product. The comparison of three datasets and their derived cropland maps is done by means of the overall accuracy, F-score and kappa metrics. Comparing results of visual interpretation with in situ data, the average labeling accuracy was 0.89 and the average drawing accuracy was 0.78.
In comparison with GlobeLand30 map, the experts were mostly agreed when labeling cropland class for each site, but the high discrepancy of agreement between the sites is observed in non cropland class. This may be explained by using a JECAM definition of annual cropland adopted by experts, while it is not completely applicable to GlobeLand30 map. This may lead to a phenomenon, when the clean fallow state of the arable land patch may be regarded as cropland in GlobeLand30 map (Agriculture class), while JECAM definition renders this case as non cropland.
The preliminary results of cropland mapping using CS dataset over large area together with Landsat-derived temporal metrics and locally-adaptive Random Forest classifier indicate that there is a significant variation of accuracy metrics among the sites – F-score ranges from 72% for Belgium to 95% for China. Further analysis will include accuracy estimation of the CS and LC datasets, as well as the assessment of the impact of the classifier on the accuracy considering general and locally-adaptive approach for cropland mapping over a large area. We also propose a convenient way for CS participant benchmarking on the run with no ground truth data involved.

Area A - Board 13

Spatial explicit crop type information is important for a variety of environmental applications, such as logistics, soil erosion modelling, biomass estimation and also to inform official statistics. Broad types of agricultural commodities grown each summer and winter, from 1987 to 2018 were mapped for the State of Queensland, Australia. For summer, the two types of crops are ‘Coarse-grain & Pulse’ and ‘Cotton’. For winter, the two types of crops are ‘Cereal’ and ‘Pulse’. The class Non-crop, present in both summer and winter, are ‘Bare soil’ and ‘Other’ (comprising pastures and woody vegetation). The crop-type maps are based on a time-series modelling step where the concepts of geostatistics are applied in the temporal domain to model the variation in land-surface phenology within a growing season.
Satellite imagery was combined from the Landsat sensors, as well as, when available, Sentinel-2 and MODIS (as a backup). The 30-m pixel-grid of Landsat served as the spatial reference. We applied the time-series model pixel-wise across the study region, to three variables derived from satellite imagery gathered for an individual summer or winter: enhanced vegetation index, and the sub-pixel proportions of bare-ground and non-photosynthetic vegetation. Weekly predicted phenological metrics then served as explanatory variables in a tiered, tree-based classification model. The classification model comprised two expert rules and two random forests. Prior to fitting the classification model, geospatial object-based image analysis was used to change the scale of analysis from individual pixels to (approximately) field-based segments. From the perspective of a map-user, in any given growing season we predicted ‘Coarse-grain & Pulse’ correctly in 79% of cases; the values for ‘Cotton’, ‘Cereal’, and ‘Pulse’ were 90%, 84%, and 73%, respectively; ‘Bare soil’ was 72% in summer, and 88% in winter. ‘Other’ was the most accurately mapped group (98% correct in summer, and 99% correct in winter). The classifications matched area estimates from official statistics well.

Area A - Board 14

Regular monitoring of agricultural production has increasingly gained significance in the last few decades, especially against the background of changing climatic and environmental conditions and food security issues. Since agricultural service providers and stakeholders rely on the timely provision of information about farming practices, such as harvest and mowing events, it is crucial to make crop mapping a more efficient process.

This study aims to determine crop harvest dates as accurately as possible using Sentinel-1 radar data, while complementing established optical remote sensing analyses (e.g. with Sentinel-2) of agricultural practices and vegetation parameters. The sensitivity of SAR towards plant physical structure and dielectric properties is exploited, but even more so its cloud-independent capabilities, where optical sensors suffer from limitations in determining exact harvest dates.

The 6-day repeat cycle of Sentinel-1 already provides an advantage; however, by utilising a multi-temporal data set with different ascending and descending orbits that cover the same research area, a near-daily temporal resolution is achieved. Yet, the different acquisition geometries, together with the effects of dynamically changing environmental conditions, i.e. precipitation events, plant and soil water content, varying phenological stages and surface roughness, pose some difficulty in the interpretation and determination of a clear and unbiased harvest date with radar data.

A test dataset including single agricultural fields in Germany for the harvest periods of 2017 and 2018 is analysed. Different crop types are investigated, such as maize and other cereals as well as grasslands. Furthermore, the potential and feasibility of an operational monitoring scenario, on a larger, up to national spatial scale, is evaluated. This requires big data and process automation issues to be addressed as well, given the high temporal resolution and product size of Single Look Complex images.

Different analytical approaches are compared in this study and their results integrated with harvest information gained from Sentinel-2 and in situ data. Taking advantage of the full complex SAR information, both the change in radar backscatter intensities in VH and VV polarisation, as well as the change in the phase component are examined. The former relies on changing scattering conditions due to a harvest event, e.g. a decrease in volume scattering. The latter is used for deriving harvest dates from interferometric coherence between consecutive acquisition dates, which significantly increases over bare soil or plant stubble after harvest. A first analysis shows very promising results in pinpointing harvest dates, especially when looking at the coherence, which is attributed to the short orbital and temporal baselines of Sentinel-1.

Area A - Board 15

Crop type inventory and within season estimates at moderate (<30 m) resolution have been elusive in many regions due to the lack of temporal frequency, clouds, and restrictive data policies. New opportunities exist from the operational fusion of Landsat-8 Operational Land Imager (OLI), Sentinel-2 (A & B), and Sentinel-1 (A & B) which provide more frequent open access observations now that these satellites are fully operating. The overarching goal of this research application was to compare Harmonized Landsat-8 Sentinel-2 (HLS), Sentinel-1 (S1), and combined radar and optical data in an operational, near-real-time (within 24 h) context. We evaluated the ability of these Earth observations (EO) across major crops in four case study regions in United States (US) production hot spots. Hindcast time series combinations of these EO were fed into random forest classifiers trained with crop cover type information from the Cropland Data Layer (CDL) and ancillary ground truth. The outcomes show HLS achieved high (>85%) accuracies and the ability to provide insight on crop location and extent within the crop season. HLS fused with S1 had, at times, a higher accuracy (5–10% relative overall accuracy and kappa increases) within season although the combination of fused data was minimal at times, crop dependent, and the accuracies tended to converge by harvest. In cloud prone regions and certain temporal periods, S1 performed well overall. The growth in the availability of time dense moderate resolution data streams and different sensitivities of optical and radar data provide a mechanism for within season crop mapping and area estimates that can help improve food security.

Area A - Board 16

Key words: land fragmentation, land consolidation, CAP, Romania, Earth Observation, supervised classification, unsupervised classification, segmentation.

Agricultural land fragmentation is a common phenomenon in Eastern European countries in general and for the Romanian agriculture in particular. In Romania, during communist rule, the arable land was divided into large plots for state-owned large-scale farming. Following the downfall of communism in 1989 and the subsequent privatization of land, these plots were highly fragmented. During the past decades, important land reforms (changing of laws and regulations regarding land ownership), were carried out, leading to significant land use / land cover changes.
Earth Observation data provided by Landsat and Sentinel-2 satellites is particularly suited for agricultural purposes, offering a wide range of opportunities in order to simplify the Common Agricultural Policy mechanisms and facilitating the monitoring of the full agricultural dynamics from national scale down to the single fields. Satellite imagery are being increasingly used for landscape condition assessment and land use / land cover change characterization, delivering a full overview of the complexity of the land fragmentation / consolidation phenomenon.
The aim of this study is to assess land use dynamics and to analyze the degree of fragmentation of Romania’s arable land, which is considered a major obstacle in the development of modern agriculture and land productivity. The change detection analysis was performed for an area of interest situated in Eastern Romania, with a surface of more than 500 000 hectares, and was based on multi-resolution remotely sensed data and integrating a series of additional data sources, related to geomorphological, climatic and socio-economic indicators, including land ownership at local-scale level. Also, derived additional parameters such as road density and distance from the closest urban center were taken into consideration, with regard to their impact on land cover changes and land fragmentation trends. The processing and analysis stage involved a series of supervised and unsupervised classification and segmentation techniques (pixel-based and object-based methods), mainly applied to Landsat 4, 5, 7 and 8 and Sentinel 2A and 2B satellite imagery, acquired from 1990 untill present and selected at regular 5-years intervals. Moreover, an additional moment was taken into account in order to assess the original appearance of the agrarian landscape at the beginning of the communist era, based on georeferenced CORONA KH-4B imagery from DECLASS-1 archive datasets.
Several parameters and indices were computed, in order to quantify the dynamics of land fragmentation / land consolidation, as well as the effects of different land reform measures: land use / land cover changes for each individual plot, LULC change rates for each individual parcel, calculated both in hectares and as a percentage, as well as specific fragmentation indices, regarding both the area and density of the arable land plots (number of plots, average number of plots per inhabitant, average plots’ area, RI – Reduction Index, CC – Consolidation Coefficient, SI - Simmons’s land fragmentation index, JI - Januszewski's land fragmentation index), as well as the edge and shape of the arable land plots (TE – Total Edge, ED – Edge Density, SI – Shape Index, MSI – Mean Shape Index, AWMSI - Area-Weighted Mean Shape Index, FD – Fractal Dimensions etc.). The outcomes were later compared to the key-events of the agrarian reform, and especially to the moment of granting the property titles.
The results were validated against independent sources, such as high-resolution satellite imagery or ortophotos, as well as statistical data originating from official data sources, such as the National Institute of Statistics, County Statistical and Agrarian Departments or statistical records provided by the different mayoralties.
The originality of the proposed approach consists in the fact that it integrates multiple Earth Observation data processing techniques, as well as multiple categories and sources of data, available at different resolutions, providing valuable insights on the regional differences and patterns of the land fragmentation and lately, land consolidation process. The methodology and results of the study can be used as a solid support for the refinement of the current policies driven by the implementation of the Common Agricultural Policies and also a feasible tool to investigate further changes in agricultural landscapes, which can be replicated at local, regional or national scale level and implemented for any other area of interest.

Area A - Board 17

The Copernicus Sentinel-2 mission is generating unprecedented volumes of data at high spatial, spectral and temporal resolutions. One agricultural application that potentially benefits from dense time series is the identification and mapping of crop types and crop rotations. Both, crop type and crop rotation are of key importance to determine crop-specific management in a spatial-temporal context.
Detailed spectral information provides opportunities for improving accuracy when discriminating between different crop types. Paramount, however, is the availability of multi-temporal data to capture crop specific plant development. The objective of this communication is to assess the value of multi-temporal information for crop type classification using Sentinel-2 data.
A number of recent studies used multi-temporal optical data for land cover classification. Inglada et al., for example, focused on the feasibility of large-scale crop type mapping using different algorithms and simulated S2 time series. Sheeren et al. demonstrated the suitability of dense time series for mapping dominant tree species using Formosat-2 data. Belgiu and Csillik compared a time-weighted dynamic time warping classification method and Random Forest using actual S2 time series data. Agrawal et al. demonstrated the usefulness of temporal information in crop type mapping. Some older studies such as Guerif et al. found similar results, but data sets were at this time mainly experimental as suitable sensor data was not yet available.
From an operational perspective, multi-temporal classification approaches are being implemented within the ESA Sentinel-2 for Agriculture project (Sen2Agri). Sen2Agri provides a classification plug-in to derive crop masks and crop type maps based on S2 time series data. In this context, S2 greatly improves the detection of crop specific temporal signatures due to its high temporal revisit time. At the same time, the S2 constellation is capable of mitigating prolonged adverse meteorological events (i.e. clouds, rain or monsoon season) or local unfavourable geographic conditions (i.e. mountainous or coastal areas). Additional synergies can be found when combining optical S2 imagery, with cloud-penetrating microwave S1 data.
The inherent changing nature of crops and their spectral and temporal signatures, which correlate with phenology related changes in leaf composition and canopy structure, as well as the accumulation of biomass during the growing season, make single-date classification approaches limited as such approaches capture only one point in time (i.e. snapshot). The situation is further complicated by several factors, such as: a) crop development depends on crop type, b) crop calendar differ across space and crop type, c) farmer customize sowing and harvesting dates, and d) local weather conditions affect growth rates (inter- and intraspecific), further modulated by site-specific soil conditions. This leads to immense variations in spectral-temporal profiles, which fluctuate on a yearly basis, and introduce uncertainties when selecting a single image for classification.
A first experience with crop type classification using one of the first acquired S2 images was presented in 2016 in Immitzer et al., in which we performed two classification exercises assessing the capabilities of pre-operational (August 2015) S2 data. Using spectral information only, the cross-validated overall accuracy resulted in 76% considering six summer crop species. Capitalizing on this experience, the main objective of this communication is to assess the impact of multi-temporal S2 data on crop type classification accuracy. We specifically compare classifications obtained with single-date acquisitions to increasingly denser time series over the cropping season of 2016, 2017 and 2018 and we show how much the accuracy is affected by the amount of multi-temporal information fed into the classifier. For this purpose, reference data of nine crop types were in situ collected for the three years. We perform the crop type classifications using a supervised Random Forest (RF) classifier and assess our results using the out-of-bag (OOB) accuracy from the RF model. In addition, an independent validation of the best models was performed using the parcel location and associated crop types declared by farmers to access the “direct payments” of the Common Agricultural Policy (CAP).
At the very beginning of the crop growing season (March-April), the classification achieves relatively low accuracies (OA: ~0.50). Significant increases in OA can be obtained between May and June, until the OA reaches its highest value in July. The final RF model was able to predict with very high confidence nine crop types for the three years (OA: 0.95-0.96 and F1 score: 0.83-1.00). The independent validation dataset with more than 5000 reference plots showed comparable results (OA: 91-95% and F1 score: 0.74-0.99). We conclude that the multi-temporal crop type classification efficiently mitigates negative effects observed when using single-date acquisition within sub-optimal temporal windows.

Area A - Board 18

This study proposes a simple method to extract crop rows from the high-resolution optical (UAV) images of a recently ploughed maize field. Crop rows are essential inputs for field navigation of robotic devices for activities such as cultivation and application of weed treatments. They are especially useful for incorporating a spatial dimension in the analysis of remotely sensed images of agricultural fields. For instance, crop analytics such as crop count can be extracted by localizing the area of interest. Remote sensing platforms such as airborne and satellite are not suitable for crop row extraction in bare fields due to low spatial and temporal resolution of the data. High-resolution UAV data, on the other hand, provides temporal flexibility and is suitable for frequent crop monitoring. In this study, crop rows are extracted for ploughed bare soil i.e. at the initial sowing stage with the intention of identifying the centre of the ploughed rows. The rows can then be used as contextual inputs for weed mapping, plant counting etc. The proposed procedure primarily involves template matching, image thresholding, binarization, and morphology. Template matching is used for locating features in a larger image wherein a given template is checked for the degree of match across the image to generate a correlation image. Thresholding and morphological operations are used to obtain the crop rows which are in turn converted to shapefiles. It was observed that the plant rows were extracted with >95% accuracy using this method. For the given field, the rows were typically oriented in one prominent direction with minor transverse rows in some cases. The threshold used for the template matching can be altered suitably in order to pick-up any minor variations in the directions of the crop rows. Noise can be filtered out based on length of features in the shape file. Identifying plant centres will require sub-centimetre resolution of data which will certainly increase the volume of the data and processing power requirements. However, in the proposed method, macro elements such as a row is considered thereby simplifying the entire process.

Keywords: UAV, crop row detection, template matching, ploughed fields

Area A - Board 19

Agricultural fields are naturally variable due to fluctuations in landscape, climate, soil properties, and management practices. Understanding the variability of biophysical features within agricultural fields helps to identify anomalies that occur at distinct points during the growing season. Early detection of these anomalies is fundamental to apply corrective measures before permanent plant damage occurs, preventing losses later in the season and therefore maximizing productivity and profitability. Multi-temporal remote sensing data has demonstrated great potential to describe the spatiotemporal variability of crop biophysical variables, and therefore may offer early detection of anomalies that occur within a field avoiding frequent ground inspection.
Despite the high potential of satellite multi-temporal remote sensing data to describe the spatiotemporal variability of crop biophysical variables, it has not yet been fully exploited for application in anomaly crop detection. Due to the spatial resolution of many satellite EO products, most of crop monitoring approaches are more focused on LULC (Land Use Land Cover) systems rather than on agricultural systems. This has led to the majority of analyses being focused on the pixel as the spatial unit of analysis, despite crop management decisions are generally made at per-field basis. In addition, the selection of the best threshold value to discriminate anomalous from non-anomalous areas is not a trivial task. The recent availability of Sentinel-2 imagery and object-based analysis approaches provide a promising direction for monitoring crop changes over agricultural fields, that are more meaningful for crop managers.
The image-object approach is especially important in crop monitoring, given the fact that agricultural management decisions are generally made at per-field basis. However, just a few authors have studied the use of Sentinel-2 imagery in agriculture from an object-based perspective. Although image-objects can be better characterized by integrating histogram analysis, the use of histograms to characterize image-objects is still limited. Histograms provide additional features to traditional statistic metrics that describe more effectively the heterogeneity of the pixels associated to a field plot and thus improve the study of changes in croplands. In addition, by transforming histograms into density curves the analyser can minimise the effects minor irregularities as well as outliers that can be caused by atmospheric or sensor disturbances. Traditionally, histograms have been used to establish thresholds that allow discriminate between changed and non-changed areas in Land Cover Change studies, but they have not been further explored to characterize within-field variability of croplands.
The main aim of this work is to develop a method to characterize potential in-field anomalies of asparagus crop plots over space and time, using histogram analysis. To do that, this approach exploits the increased spatial and temporal resolution of Sentinel-2 imagery and uses histogram analysis to establish thresholds that differentiate the potentially anomalous pixels in each asparagus-field.
The analysis was performed using Normalised Difference Vegetation Index (NDVI) products, derived from Sentinel-2 images, captured between April 2016 and September 2018 for the region of La Libertad-Peru. The spatial extents of the asparagus plots are known, and it is therefore proposed that the anomalous pixels can be differentiated from the non-anomalous ones using the distribution of the pixel NDVI values within the boundaries of each field. Since the same management activities are applied over the whole field, the distribution extracted from each plot is expected to be normal, and those values that form an elongated tail are more likely to be anomalous. To extract thematic information from plot data, a suitable threshold needs to be defined that optimally separates the anomalous from the non-anomalous pixels. An optimal threshold has to be different for each crop plot, as its distribution depends on biophysical variables that fluctuate over space and time, such as crop variety, landscape, climate, soil properties, and management practices. Thus, the thresholding method needs to be sensitive to the distribution of the NDVI pixel values contained within each crop field, and produce thresholds that are different at distinct points during the growing season.
The histogram analysis proposed, allows identifying dynamic thresholds, sensitive to the statistics of NDVI pixel values distribution within each plot, to differentiate the potentially anomalous pixels. A fundamental assumption is that the NDVI pixel values extracted from each crop field are expected to be normally distributed. Thus, the difference between a distribution with anomalous values and one without anomalous values is the normality of each field distribution. The removal of the atypical values in increments from the tails towards the median until the distribution is normal, separates the anomalous from the non-anomalous NDVI pixels. The distribution normality is assessed based upon measures of skewness and kurtosis for each iteration. The combination of the lowest kurtosis and skewness is used to identify the bins at which the histogram should be sliced to produce the most normal distribution.
Results show that the histogram analysis approach is able to extract thresholds sensitive to in-plot crop performance, based on multi-temporal NDVI values. The spatial distribution of anomalous pixels within each field showed to be different for each type of anomaly, and different crop stages. Most of the anomalous pixels situated under the lower thresholds were clustered near the plot edges due to the presence of paths but also potentially due to image mis-registration. It was found that when the field crops have the lowest values of NDVI (after the harvest), anomalous pixels, either above the upper threshold or below the lower threshold, formed small scattered clusters within the field crop, helping to inform in-plot management strategies.

Area A - Board 20

Managing plant pests in agriculture requires both reliable methods for early detection, to avoid or mitigate outbreaks, and methods to assess a pest’s impact in order to forecast any yield losses. Xylella fastidiosa is native to North America and considered one of the most dangerous plant pests for agriculture in many regions of the world, incurring great costs in the Californian wine-industry and in citrus cultivation in Brazil. Xylella fastidiosa now poses a danger to Mediterranean agriculture since it was found in Italy, France, and Spain and as the list of plants it can infect, which include olive trees, almonds, and citrus, grows. The pathogen had not been found in Europe before 2013, but has since caused extensive damage to many olive groves in parts of the region of Apulia, in southern Italy.

In the very south of this region the pest is now considered established as eradication is no longer possible. Pest management efforts are therefore focusing on containing the epidemic to prevent further spread, an approach that relies heavily on early detection of new infections outside the current Xylella range. This is particularly challenging as laboratory studies show that newly infected plants can go months without showing visible symptoms. Meanwhile, the exact extent of the damage in the area already infected by Xylella fastidiosa (currently ca. 600,000 km² in Italy) is not precisely known. We will present recent and new research on how remote sensing, from airborne as well as satellite platforms can help tackle these challenges, by supporting the detection of plants and orchards infected in a Xylella fastidiosa outbreak, as well as monitoring the damage left in its wake.

Recent work demonstrated, by analyzing upwards of 4000 plants in a 1200 ha area, how very high resolution hyperspectral and thermal remote sensing from airborne platforms can quantify the changes in plant traits associated with early, visible, symptoms of Xylella fastidiosa infection in olive trees. Moreover, alterations in these traits (which included chlorophyll fluorescence and thermal stress indicators retrieved from the imagery) in visually asymptomatic plants was associated with a higher likelihood of developing visible symptoms in the ensuing months.

This finding indicates that symptoms of new Xylella fastidiosa infection in trees can be detected by airborne spectroscopy and thermography before symptoms are visible to the naked eye. To test whether disease prevalence and incidence across individual orchards, rather than trees, could be operationally monitored using satellite data, Sentinel-2a data were then analyzed. It revealed a strong influence of background variation, caused by soil and understorey dynamics, both within and between orchards of different disease levels. 3-D radiative transfer modelling could largely account for this influence on atmospherically- and soil-adjusted vegetation indices from Sentinel-2a. Further tests with longer time series of vegetation indices from MODIS were then carried out to investigate whether abandonment of orchards turned unproductive by Xylella fastidiosa, caused noteable changes in landscape phenology.

When olive trees ultimately succumb to Xylella fastidiosa they show conspicuous crown desiccation and losses of leaves and branches. The number of olive trees that have been lost since the pest was first detected in Europe has thusfar not been quantified, however. While trees have in the past been counted from aerial photography, this task is made challenging in the Xylella infected region, because of its size and the associated variation in management practices, and by the facts that crowns of neighbouring trees often touch each other and that the older trees that typify the olive trees in Apulia often have segmented crowns. To address this question, we trained and deployed deep learning algorithms to identify individual olive tree crowns in historical and more recent aerial photographs and estimated olive tree losses since 2013 throughout southern Apulia.

As countries around Europe, and the world, are surveying their territories for Xylella fastidiosa, and other plant pathogens, these different uses of both legacy and new Earth Observation data illustrate how remote sensing based monitoring can support quarantine pest management.

Area A - Board 21

High value agriculture in South America is a major contributor to global food supply chains and local economies. Growers in this region typically operate at large scales, with intensive management of many hundreds of hectares. Managing intensive crops such as asparagus or grapes at these scales requires accurate, within season data on each parcel to plan activities such as irrigation scheduling, chemical applications and harvest timing. Earth observation is increasingly being used to provide growers with improved crop insight, with the Sentinel program in particular providing the high-frequency observations needed to underpin within-season monitoring.
EO4cultivar is a four year project funded by the UK Space Agency that is developing EO-based solutions that can be easily disseminated and incorporated into grower management practices. Working with commercial and government stakeholders in Peru and Colombia, the project is delivering information frequently and rapidly within the growing season via trusted channels. In this paper we will illustrate how Sentinel-1 and -2 data are being used to provide asparagus growers the following monitoring services:
1. Baseline monitoring of asparagus growth stage through the season is being delivered on a per parcel basis using Sentinel-1 radar, augmented by Sentinel-2 optical observations when weather permits. Dense time series of Sentinel observations are regularised and used to extract key phenological parameters throughout the season. These parameters are used to monitor the development of crops and highlight parcels that may be underperforming.
2. Within-parcel variation is mapped by Sentinel-2 to identify parcels with non-uniform canopy. Spatial variations in NDVI, EVI and other indices are often indicators of problems such as pests, disease or salinity. When combined with phenological stage, these observations can help identify issues as they develop and inform management activity.
We will also show how these EO data are being combined with asparagus growth models to improve the ability of growers to predict harvest yield and the optimum interval between harvests. This is of particular importance as it allows growers to better plan the harvest and provision of crop into the supply chain, which in turn reduces waste and improves competitiveness.

Area A - Board 22

Crop classification is an important application of remote sensing on agricultural areas. Knowing the crop cultivated on each agricultural field is important for administrations and stakeholders for several applications, e.g. to make agricultural statistics, plan the harvest, manage agricultural subsidies, forecast agricultural production, assess food security or estimate commodity prices. In addition, crop classification is also a necessary step prior to other remote sensing based applications (e.g., bio-geophysical parameter retrieval and crop monitoring) can be implemented at the field scale. To date, most crop classification efforts have relied on multispectral data collected by optical sensors based on the differences in the spectral behavior of crops in key moments where phenological differences are apparent. However, Synthetic Aperture Radar (SAR) observations provide complementary information to optical ones, sensitive to morphological and dielectric differences between crops. Furthermore, their ability to penetrate through clouds make them an ideal information source for areas with persistent cloud cover, where the viability of optical multispectral sensors is severely hampered.
In the last years, with the advent of the Copernicus program, the opening of the Landsat archive and the irruption of huge numbers of nano-satellites, the availability of remote sensing data has increased exponentially, enabling the development of classification methods that exploit the complementarity of different technologies and the power of time series information. The objective of this study is the implementation of a classification algorithm exploiting the information provided by Sentinel-1 (SAR) and Sentinel-2 (multispectral) time series. Our hypothesis is that the potential of the time series information can be exploited achieving accurate classification results with rather simple classification approaches. With this aim a supervised crop classification approach is implemented based on the evaluation of the temporal signature of crops using simple statistical measures of agreement. Using a training dataset, typical crop time series (i.e., temporal signatures) are obtained for the following remote sensing features: NDVI, VV backscatter, VH backscatter and VH/VV ratio. The classification approach statistically compares the time series of the field to classify with the temporal signatures of the different crops considered. Then, the field is assigned to the crop class that minimizes the overall error (RMSE) or maximizes the overall correlation (R2) with its time series.
The proposed classification concept is tested in a case study focused on the Spanish province of Navarre during the agricultural season of 2016. Navarre has an area of 10,391 km2, one third of which corresponds to agricultural areas. Due to its landscape and climate diversity a wide variety of crops are grown, being the most frequent ones rain-fed cereals, grasslands, corn, legume crops, vineyards, olives and fallow fields. The province is divided in seven Agricultural Districts, regions with rather homogenous climatic conditions, being some crop types more frequent in some districts than others. For this study the Common Agricultural Policy (CAP) declarations and field inspections files were available. The declarations file consisted of vector information for all the field boundaries (335,094 polygons) with data about the crop declared by farmers per season and field, whereas the inspections file included a sample of fields (21,950 polygons) whose actual crop type was determined after field inspections by technicians from the administration. All the available Sentinel-1 and Sentinel-2 scenes from Sept/2015 to Dec/2016 were downloaded and processed. Sentinel-2 processing consisted of an atmospheric correction using SEN2COR, a mosaic of four tiles and the calculation of the Normalized Difference Vegetation Index (NDVI). Sentinel-1 processing included orbit correction, thermal noise removal, calibration, speckle filtering, topographic normalization and orthorectification leading to backscatter coefficient values (gamma0) in VH and VV polarizations. Finally, statistics were computed per field creating a database with four features: NDVI, backscatter in VH and VV and VH/VV ratio. All the processing was done using SNAP and Python scripts.
A two level crop class legend was established with 17 and 30 classes, respectively. For each class the temporal signature of each feature was obtained by using the declarations file (training phase). At this point, a distinction was made between winter and summer crops, adjusting the duration of the temporal signature for each case. The field inspections file was used for testing and accuracy evaluation. Due to the diversity of the study region, classifications were not only performed for the whole province, but also individually for each one of the seven agricultural districts. Each field was assigned to the class providing the best temporal signature fit, for this two statistical measures were compared (RMSE and R2). The results obtained for each one of the four features were evaluated, as well as a final model that computed the best overall fit to the temporal signatures of the four features. Although still preliminary, some interesting results have been obtained. Best results were always obtained when the temporal signatures of the four features were compared and the assignation was done to the class that minimized the overall RMSE. Classification accuracies were modest (~60%) when the whole province was considered, but they significantly improved when applied individually for each agricultural district reaching values ~80% in most cases. It was observed that the proportion of fallow fields on a district negatively impacted classification results, due to the heterogeneity of this class and its frequent confusion with different crop types. Altogether, the proposed approach seems to be well suited for operational applications where crop maps are pursued.

Area A - Board 24

The central topic of this study is the derivation of yields and water use efficiency (WUE) for large areas by assimilating plant physiological data derived from COPERNICUS Sentinel-2 data into the land surface process model PROMET. The use of high-resolution Sentinel-2 time series (10 m) within the developed system enables a crop specific derivation of information on WUE, which is not possible using medium resolution data. The derived high-resolution data on WUE is then transferred into a larger scale as frequency distributions. This makes it possible to make statements on crop specific WUE on the levels of Sentinel-2 tiles or countries. Results of this method are presented for Germany (agricultural season 2018) and an intensely used agricultural region in the Romanian plain (agricultural season 2017).

Agricultural water use efficiency (WUE) is defined as kg crop yield per m³ evapotranspiration during the vegetation period. An increased WUE is generally linked to an increased yield. The global monitoring of yield and WUE is an important part to identify regional hot-spots of unsustainable water use. Implementing sustainable ways to secure food supply for a growing and wealthier global population is among the most important sustainable development goals (SDGs).

This study is embedded in the ViWA (Virtual Water Values) project of the funding program “Global Resource Water (GROW)” in the framework program FONA (Research for Sustainability) of the German Ministry for Education and Research (BMBF). ViWA aims at developing and using a global, high-spatial and -temporal resolution, remote sensing based management and monitoring system. The system is designed to retrieve information on the efficiency and sustainability of water use in agriculture and competing sectors. This enables assessing the progress in achieving water-related SDGs through more efficient use of water.

The developed assimilation method runs on representative selected crop specific pixels within the selected region, which can be on county level or bigger administrative units like full countries. A fully spatially distributed approach using Sentinel-2 data is too demanding regarding computational resources and has no qualitative advantage compared to using representative selected pixels.

Sentinel-2 satellite imagery was processed using in-house software. This comprises sophisticated methods on cloud and cloud shadow detection, atmospheric correction and derivation of plant physiological parameters with the surface reflectance model SLC (Soil-Leaf-Canopy).

Since the selection of pixels is crop specific, a crop type classification based on spectral characteristics and leaf area index (LAI) time series was carried out. Only pixels, which are classified with a high degree of accuracy as a specific crop type, are selected for further processing. Furthermore, the usage of the COPERNICUS High Resolution Layers (HRL) allows working only on agricultural used regions. The classified crop types are the main crops of the study areas (maize, rapeseed, winter/spring wheat, winter/spring barley).

A representative pixel for a specific crop type has to fulfil different requirements. It has to be agriculturally used, its position may not be at the field border to guarantee homogeneity and a continuous time series of derived plant parameters has to be available. The amount of selected pixels is linked to the areal size, or, if available, to the crop specific agricultural area of regarded districts or counties. In total, several thousand pixels were selected for each crop type. A good spatial distribution of samples was guaranteed by using Euclidean distance thresholds between them.

Through a crop specific LUT inversion of the SLC model, exact plant physiological parameters (e.g. LAI) for the complete time series of satellite imagery were derived for the selected pixels. The derived LAI time series were processed using sophisticated methods to detect and eliminate outliers and to create harmonized, daily LAI values. Afterwards the harmonized LAI time series for the selected pixels were directly assimilated into the PROMET model. For Germany, the modelled yield values were aggregated into district level and compared to measured data, showing a very good correlation. Preliminary results for the Romanian plain for maize and winter wheat, including an estimation of water use efficiency, are very promising.

Concluding, the pixel-based assimilation method works fully automatic and is designed to be applied on all globally distributed ViWA test sites. By coupling the results on WUE and yield to global model scenarios of PROMET, which will be calculated within the ViWA project with a high temporal (1 h) and spatial (1 km) resolution, the results achieved within the test sites can be transferred to a global scale.

Area A - Board 25

Land reform in South Africa is a process of trying to reverse the injustices from the past by returning farm land back to the original indigenous inhabitants of the country. This has however sparked a big debate on whether this will have an impact on food security, economic growth and production as current farmers are highly skilled and knowledgeable in agriculture. Developing countries on the African continent suffer severely from food shortages and high levels of poverty and therefore it is of vital importance for the agricultural sector to flourish to provide these extreme vulnerable citizens with food. As this is a very emotional debate there is a need for a scientific study to determine whether redistributed land that has already been given to previously disadvantaged people been functioning at its optimum potential or not. A scientific study will eliminate all emotional and racial aspects of the debate and will only present the scientific facts of the activities on the farm. This study aims at determining whether these farms has been used at its optimum potential. Satellite imagery will be used to determine the amount of crops that has been planted on the farm each year before and after it has been redistributed. The time period used will be as far back as possible to try and compensate for natural variability. Also, rainfall data from the South African Weather Service will be used to identify drought years which will possibly have an impact on crop production. Landsat, Sentinel and Airbus spot data will be used to determine whether there is a change in production or not. Spatial data from these redistributed farms will be provided by the department of Land Reform.

Area A - Board 26

In 2016, the international community defined 17 Sustainable Development Goals (SDG) to stress the importance of sustainable development. Among these, several goals focus on water. For example, SDG 6.4 is to increase water-use efficiency across all sectors substantially by 2030. Until now, there is no operational mechanism of managing or controlling this target, even though it is clear, that water resources are currently often used in a manner that is neither sustainable nor efficient. While the world population is continuously growing, water resources are substantially constant and therefore increasingly scarce. This of course has negative impact on food security and the goal to end hunger (SDG 2) as well as 92-99% of the global use of water is connected to agriculture.

Therefore, the Collaborative Project “Virtual Water Values (ViWa)” aims at developing and using a global, high-spatial and -temporal resolution, remote sensing-based management and monitoring system for the efficiency and sustainability of water use in agriculture and competing sectors (http://viwa.geographie-muenchen.de/). A system like this will enable us to transparently and neutrally assess the progress in achieving water-related SDGs through more efficient use of water. Using high resolution Sentinel-2 data as input to the system allows ViWa to consider both the “global” and the “local” scale in developing, formulating and analyzing trade-offs of practicable solutions for a more sustainable and efficient use of water resources.

Addressing the local level, the work presented here aims at reaching an increased water use efficiency on farm-level by utilizing globally available satellite data together with expert knowledge on crop modelling to calculate water demand and use this information to regulate the locally available irrigation systems.

The results of this approach will be presented for a farm in a drought-prone part of Zambia (Africa). Due to the geographical location of the farm, cultivation of wheat during dry season (May to October) is only possible with intense irrigation. The irrigation water originates from a surface reservoir whose water levels have fallen during the last few years due to a harsh drought that makes the urgency of sustainable use of water even more obvious. The farm owns innovative irrigation systems that offer the possibility to adjust the amount of water applied in a site-specific way (Variable Rate Irrigation, VRI). However, only the knowledge about the status of the plants and the corresponding site-specific crop water demand makes the use of these systems meaningful and efficient. Until now, this know-how was missing. Now, with the help of the continuous stream of Sentinel-2 data, it is possible to derive and monitor the site-specific crop water demand during the complete season.
Both the spatial and temporal scale have to be taken into account when giving irrigation recommendations. Due to spatial variations in natural properties, within a single field crop conditions can vary considerably, but also the current weather conditions make a difference in plant water demand. Defining the site-specific crop water demand as input for the calculation of the correct irrigation amount is one of the main points of this study. Leaf area [m²/m²] derived from Sentinel-2 satellite data via the radiative transfer model SLC (Soil-Leaf-Canopy) is used in combination with crop growth modelling using the highly detailed physically-based water balance and crop growth model PROMET. Based on soil and elevation data as well as hourly meteorological data from various measurement stations, the model simulates plant growth and crop water demand. By updating the model with plant status from the latest satellite data, the simulations produce site-specific maps showing an up-to-date status of the crop water demand, the field capacity or the soil moisture within a field. From these findings, ready-to-use prescription maps for the VRI control software are generated to optimize irrigation at every single point in the field. Additionally, through meteorological forecast data and simulation into the future, irrigation can be planned for the next seven to ten days. The prescription maps are then remotely sent to the African farm, which has irrigated using this system for two seasons now.

The goals of site-specific irrigation management are to minimize waste of water and prevent water stress in the crop at the same time. Daily simulations of different management scenarios showed, that, for example, with the reduction of the VRI system speed and the elongation of the revolution time from 1 day to 2 days, the crop water demand can be reduced from 8.0 mm to 6.6 mm because of lower canopy interception. This has saved up to 1750 cubic metres of water for one pivot and in one day of irrigation alone, helping to stabilise the vulnerable water reservoir. This stabilisation in the dry season of 2018 was also mapped and monitored using Sentinel-2 data.
These first experiences at our demo farm in Zambia, where the approach has already been implemented successfully, show great promise for irrigation management, especially since with the adjustment of the irrigation system based on the site-specific crop water demand, the use of irrigation water could be decreased significantly while the yield increased compared to the last years.

Area A - Board 27

Monitoring the irrigation amounts and timing is of paramount importance in arid and semi-arid regions for an efficient management of water resources. The crop water requirements have been commonly estimated from FAO-56 dual crop coefficient (FAO-2Kc) model by simulating evapotranspiration (ET) that is partitioned between the soil evaporation (E) and the plant transpiration (T). However, one of the main limitations of FAO-2Kc for operational irrigation management over large areas is the unavailability (over most irrigated areas) of irrigation data. To overcome this issue, this study proposes a new approach to estimate the irrigation volumes and dates at the crop field scale throughout the agricultural season from land surface temperature (LST) and vegetation index (VI) data integrated in a FAO-based modeling. LST is first partitioned into its soil and vegetation components to force E and T separately via thermal-derived estimates of evaporation and transpiration reduction coefficients (Ks and Kr, respectively). The thermal-derived Ks is also used to derive a first-guess (LST-derived) root-zone soil moisture (RZSM) estimate, based on the FAO-2Kc relationship between total available water and Ks. The dynamics of first-guess RZSM is then analyzed to retrieve the irrigation amounts and dates by inverting a water balance model using precipitation, LST-derived ET and RZSM. FAO-2Kc is next forced by the previously retrieved irrigation and re-analyzed estimates of daily ET (ET_FAO+LST) are finally provided for the full season. This approach is tested over an irrigation perimeter area of 12x12 km2 in central Morocco covered mainly with winter wheat by using 13 Landsat-7 and -8 images along the agricultural season. The total irrigation depth (274 mm) is correctly estimated and is very close to the actual effective irrigation (288 mm) applied by the farmer. Daily ET_FAO+LST is estimated with a R2 value of 0.64 and a RMSE value of 0.77 mm.d-1 (28%) by forcing FAO-2Kc using the retrieved irrigation (from LST/VI data and precipitation only). This research demonstrates the utility of the high spatial resolution but temporally-sparse thermal data of Landsat for the application of FAO-2Kc model at regional scale in order to estimate the crop water requirements and ET on a daily basis. However, improving the temporal frequency would be an essential asset to better constraint the irrigation timing.

Area A - Board 28

The quantification of water stored in agricultural plants plays an essential role in understanding the impact of cultivated areas on the earth’s water cycle. Due to its close association to biochemical factors, such as vegetation transpiration and net primary production , the knowledge of quantities of water contained within agricultural crops is crucial, particularly for the development of environmental process models. Moreover, quantifying canopy water content (CWC) is important in view of water use efficiency of plants, evaluation of plant physiological status and health, and crop ripening monitoring, as it can be monitored by the future hyperspectral missions, such as EnMAP. Crop water content is commonly expressed as equivalent water thickness on canopy level (EWTcanopy) and its retrieval from remote sensing has been targeted by many studies. However, the effect of radiative penetration into the canopy and thus, detectability of water stored in plants, has not yet been fully understood. Therefore, in our study the Beer-Lambert law is applied to inversely determine water content information in the 930 to 1060 nm range of canopy reflectance from in-situ measured winter wheat and corn spectra collected in 2015, 2017, and 2018. The spectra were converted into the EnMAP spectral bands using the actual spectral response functions. Thereby, the developed algorithm inversely derives the thickness of the optically active water layer responsible for the 970 nm absorption in a measured reflectance spectrum. With regard to potential dependencies of the retrieval results on the across track platform pointing capability of ±30° (EnMAP observer angle), the spectroscopic measurements were carried out in three different viewing angles (nadir, +30° observer zenith angle (OZA) and -30° OZA). The spectral model was calibrated using a look-up-table (LUT) of the radiative transfer model (RTM) PROSPECT containing 50.000 spectra. Internal model validation was performed using two leaf optical properties datasets (LOPEX93 and ANGERS). Destructive in-situ EWTcanopy measurements were collected separately for leaves, stalks, and fruits. At nadir, correlation between measured and modelled water content were most promising for leaves and ears in case of wheat, reaching coefficients of determination (R²) up to 0.72 and relative RMSE (rRMSE) of 26% and in case of corn for the leaf fraction only (R² = 0.86, rRMSE = 23%). These findings indicate that, depending on the crop type and its structure different parts of canopy water are observed by optical sensors. Regarding the off-nadir spectral measurements we expect that different shares of plant compartment specific water content will be detectable. By assessing the estimation uncertainties from off-nadir data, the monitoring capabilities of the future EnMAP sensor will be very advantageous for agricultural monitoring applications, with data availability of up to 4 days. Moreover, the plant water retrieval (PWR) model will be provided as a tool within the EnMAP-Box 3 software, freely available as an approved QGIS plugin.

Area A - Board 29

Critical Zone Observatories (CZO) provide an integrated research system to determine how geological, chemical and atmospheric interactions impact the ecosystem over different spatial and temporal scales. The Ganga Alluvium Plain (GAP) CZO, designed by the Indian Institute of Technology in Kanpur (IITK) through the Ministry of Earth Sciences (MOES), encompasses an agriculturally-dense landscape (major crops being wheat and rice) that is particularly vulnerable due to poor agricultural practises, excessive water extraction and fertiliser-use. Additional pressures from increasing average temperatures, drought and extreme rainfall events, combined with the expected population rise over the next decade will likely create an additional burden on the landscape threatening future food production and the livelihoods of rural communities.

Challenges in monitoring the Ganges basin include a lack of high temporal and spatial resolution data at sub-regional, district -level scales as well as large uncertainties in land surface measurements over the heterogeneous landscape. To overcome such issues, the GAP CZO comprises a high-resolution data repository to monitor hydro-meteorological parameters, soil physical, chemical characteristics and geochemical parameters. Supported by sensors on-board space- and unmanned aerial vehicle (UAV) platforms, detailed landscape mapping will provide independent sources of information about the ecosystem and enable an evaluation of current agriculture practises.

We will show how land measurements from the Sea and Land Surface Temperature Radiometer (SLSTR) on the Sentinel-3 missions have been used to characterise changes in agriculture, crop stress and water availability. Furthermore, complementary measurements from thermal and multi-spectral sensors on UAVs and ground-based thermal infrared (TIR) radiometers will provide an in-depth mapping of the CZO from which land-use processes contributing to groundwater depletion and crop stress will be identified. By mapping these changes over varying crop cycles and spatial scales, the datasets will be used to identify agricultural practises in the region and support in-country Non-Government Organisations (NGOs) focussing on agricultural management.

Area A - Board 30

We are using the latest innovations in satellite remote sensing to estimate crop water use and provide timely and consistent feedback to farmers. This information is crucial for efficient irrigation management and can inform practices that increase agricultural productivity and sustainability from small to large scales.
Background
Monitoring the changes of soil and crops status in agricultural fields throughout the growing season is key in increasing the production efficiency. As the crop develops from seedling to a fully mature plant, its transpiration changes accordingly, and so do the crop water requirements. In the early stages of the growing season, the majority of the evapotranspiration, which represents the combined water loss due to evaporation from the soil surface and transpiration from the crop, is attributed to soil evaporation. Yet, as the crop develops, the relative contribution of transpiration to the evapotranspiration increases as the vegetation cover increases, and eventually declines with maturity and senescence. Therefore, information about the daily crop evapotranspiration, can facilitate better irrigation planning, and ultimately, increase the water use efficiency.
The crop coefficient, which represents the crop water demand, is a useful and widely used approach for irrigation management. The United Nations’ Food and Agricultural Organization (FAO) has provided details on the development and use of the crop coefficient, for different crops in different parts of the world. However, these crop coefficients have been shown to vary between sites and between seasons. Additionally, in cases of atypical crop development and water-use patterns caused by weather anomalies, adopting the FAO-recommended crop coefficient values often results with imprecise crop water use estimations. As a result, local adaptations to the FAO-recommended crop coefficient are implemented to form local coefficient tables, but even these sometimes fail to capture deviations from standard conditions due to specific fertilization, variations in crop planting density, and stress factors such as pests. In addition, the spatial variation in crop water use due to spatial heterogeneity in soil characteristics such as water holding capacity and nutrients availability is not reflected in standard coefficient tables. Accordingly, in the absence of reliable, real-time information about crop water use, there is a need for better crop coefficient estimates.
One approach to address this need is by using satellite remote sensing imagery. This technology is attractive for modeling crop water use since it provides a synoptic coverage at fixed time intervals, and can therefore monitor changes over time. Moreover, spectral indicators derived from remote sensing imagery are highly correlated with crop characteristics including biomass, Leaf Area Index (LAI), plant height, and yield. Similarly, these spectral indicators can serve as near-real-time surrogates for crop water use since they depict a similar temporal pattern. The reason for that is that both plant transpiration and light absorption increase roughly at the same rate throughout the growing season. In order to model the crop water consumption with satellite observations, we are performing evapotranspiration measurements in the field on the same days as the satellite overpass occurs.
The basic limitation of satellite remote sensing application for irrigation management is the compromise between the sensor’s revisit time and spatial resolution. Sensors with a short revisit time such as the moderate-resolution imaging spectroradiometer (MODIS) that provides daily coverage are characterized by a coarse spatial resolution (>250 m), while sensors with medium spatial resolution such as the Landsat series are characterized by longer revisit times (16 days). Cloudy conditions further reduce the temporal resolution for all optical sensors, thus posing another limitation on operational applications. Irrigation management decisions for field crops should ideally be based on a dense time series of imagery that are fine grained enough to distinguish between field plots. Commercial high spatial resolution satellite sensors are usually not employed for crop monitoring because their imagery are not public domain and come at a significant cost, rendering them too expensive for most operational agricultural applications. Therefore, in spite of remote sensing models for crop water consumption, the limited availability of imagery with suitable temporal and spatial resolutions at no or low cost hindered the development of worldwide remote sensing application for near-real-time irrigation decisions.
The successful recent deployment of the two Sentinel-2 satellites creates a unique opportunity for operational crop water consumption estimates. Sentinel-2 multispectral spaceborne imagery with a 5-day revisit time (obtained by the combination of Sentinel-2A and Sentinel-2B data) can potentially create a dense observation time series at 10-20 m spatial resolution, which would allow the application of this technique even for small fields. Sentinel-2 imagery offers an acceptable compromise between the revisit time and spatial resolution, with increased spectral abilities for vegetation monitoring compared to previous public domain spaceborne imagery. Hence, the overarching aim of this research was to develop methodology to estimate cotton water consumption based on Sentinel-2 imagery. The key objectives of this study were to (1) estimate daily crop water consumption experimentally in the field, and (2) develop empirical models that link crop water consumption with remotely sensed spectral indicators from Sentinel-2.
Methods
The measurements took place during the summer of 2016 in a cotton field near Gedera, in the Shfela region in Israel. Direct measurements of actual crop evapotranspiration were done by an eddy covariance system that measures at high frequency the water vapour concentration together with the wind speed and direction (Fig. 1). A total of seven relatively cloud free Sentinel-2A images acquired during the cotton-growing season were analysed in this study, including five images that coincide with our agro-meteorological field measurements. We calculated 22 vegetation indicators based on different combinations of the unique Sentinel-2 spectral bands. We then checked to see which of these spectral indicators is most related to the crop coefficient based on our field measurements of crop water consumption. In order to compare our findings with previous studies we estimated the crop coefficient according to relationships found in previous studies between the crop coefficient and a spectral indicator, and to the standard crop coefficient recommendations by the Israeli Extension Service.
Results
We found that most of the spectral indicators can be used to estimate the crop coefficient, but that the three best ones rely on the red-edge spectral bands, which are unique to Sentinel-2. For the particular field in our study, the relationships that were found in other studies between spectral indicators and the crop coefficient grossly overestimated the crop water consumption that we measured in the field. The Israeli Extension Service recommendation was also higher than the crop coefficient derived from the measured crop water consumption (Fig. 2).
Conclusion
Sentinel-2 is superior to older generations of public domain satellite data in terms of spatial, temporal and spectral resolutions. This allows, for the first time, to estimate the crop water consumption, an important parameter for irrigation management, at a high frequency that can support irrigation decisions, at a fine spatial resolution of 10 m that well captures within field variability, and at higher accuracy than before, owing to the sensor’s unique spectral bands that cover the red-edge region.

References
Kamble, B., Kilic, A., & Hubbard, K. (2013). Estimating crop coefficients using remote sensing-based vegetation index. Remote Sensing, 5, 1588-1602.
Montgomery, J., Hornbuckle, J., Hume, I., & Vleeshouwer, J. (2015). IrriSAT—Weather based scheduling and benchmarking technology. Proceedings of the 17th ASA Conference, 20-24.
Rozenstein, O., Haymann, N., Kaplan, G., & Tanny, J. (2018). Estimating cotton water consumption using a time series of Sentinel-2 imagery. Agricultural Water Management, 207, 44-52.

Fig. 1: Field measured water consumption for cotton in this study, and the standard recommendation by the Israeli Extension Service (IES) for cotton growers in this region compared to estimates according to three models: (a) The model developed in this study; (b) Kamble et al. (2013); (c) Montgomery et al. (2015).

Area A - Board 31

The FAO portal to monitor Water Productivity through Open access of Remotely sensed derived data (WaPOR) monitors and reports on agriculture water productivity over Africa and the Near East. The database is built on three levels: i) continental level, at 250 m resolution, covering the whole of Africa and the Near East, ii) national and river basin level at 100 m resolution covering 18 countries and four river basins and iii) irrigation scheme and watershed level, at 30 m, covering 5 areas of 100,000 ha each. The temporal span of the database currently ranges from 2009 to date, with possibilities of extend it over space (and time) through additional donors’ engagements.
WaPOR is an open access platform:
The methodology and algorithms used to calculate variables are free and openly accessible. A dedicated FAO web portal provides free direct access to the database including: decadal actual evapotranspiration, decadal above-ground biomass production, decadal net primary production, daily reference evapotranspiration, daily precipitation, annual and seasonal land cover, crop calendars.
Water and land productivity, both under rain-fed and irrigated conditions, are computed and made available on-demand. The database provides the information base for water balance analyses at basin scale, and thus allows for developing water accounts using tools such as Water Accounting Plus.
The database already comprises more than 7 000 raster files which have been organized in a convenient catalogue with supporting metadata to facilitate data navigation and comprehension.
Through the portal, the data can be quickly and easily; viewed, interrogated, analyzed and downloaded.
Available analyses include time series and area statistics. More complex user-defined calculations are made available through dedicated python scripting.
In order to deliver these capabilities, we take advantage of Google Earth Engine’s data and compute capabilities, and employ modern scalable cloud native computing technologies.
The documented Application Programming Interfaces (ReST APIs) enables the data to be directly incorporated into other websites and consumed in user data pipelines.
The poster highlights key technologies used to develop the portal and database, with the aim of facilitating wider uptake and application development by other partners.

Area A - Board 32

FruitLook is a unique program to the world. Via funding of the Western Cape Department of Agriculture innovative satellite data products are openly available to all farmers in the Western Cape province of South Africa. Launched in 2010, it is currently used on 50 000 hectares for crop monitoring and efficient farm resource management.

Via FruitLook, every week quantitative data products describing crop growth, crop evapotranspiration and crop mineral content are made available to the farming community. This data is created through energy balance algorithms using raw satellite imagery from Sentinel-2 and Landsat 8 in combination with local weather information. Fruitlook offers nine different information data sets, from biomass production showing crop productivity to an evapotranspiration deficit parameter indicating whether crops are experiencing stress.

FruitLook nowadays services an agricultural area of 1.5 million hectares; this is the extent of data available via the web-portal to the local farming community. This includes close to 200 000 ha of (irrigated) fruit crops. The data is available via www.fruitlook.co.za and can be accessed on a field-by-field basis. All major agricultural areas in the Western Cape are covered meaning every single farmer in the province can make use of the service. Local partners support farmers in the successful uptake of the service.

During the 2017-18 summer production season FruitLook counted 776 users on 16 507 blocks and 53 049 hectares. Users indicate that FruitLook is an extremely valuable tool: in a questionnaire from early 2018, 71 of the 100 respondents indicated that using FruitLook improved their water management. Additionally, users became better aware of crop development throughout the season (57%). More than 30% of the respondents to the survey indicate they gained more insight in how blocks compare to each other in terms of growth and water use, improved their uniformity of blocks, detected/solved growth issues in the midst of the season and increased their yield.

The efficient use of water is a vital component of sustainable and responsible (irrigated) farming in the Western Cape. It means FruitLook, along with various other tools, will remain a powerful instrument to help farmers optimize water use and general farming efficiency. During this presentation the FruitLook story and lessons learned are presented: it will show how satellite data products are efficiently used on ground level and what it took to get there.

Area A - Board 33

Irrigation is a crucial practice for agriculture, especially in arid and semi-arid regions, because it considerably influences the crop production and quality. Remotely sensed data can be efficiently used to outline irrigated croplands, offering a synoptic overview of wide areas and of their extent. The most of the methods so far proposed to produce irrigated /not irrigated areas maps by remotely sensed data are based on optical images [1].
More recently, the availability of operational SSM products derived from spaceborne radiometers and or scatterometers (e.g. SMOS, SMAP, ASCAT) has stimulated the investigation of Surface Soil Moisture(SSM) potential in the detection of irrigated areas [2]. One limitation of these studies is related to the coarse spatial resolution of the aforementioned operational products (typically on the order of 25 km).
Here, a methodology using high resolution SSM data from Sentinel 1 images is presented for the identification of binary masks of irrigated/not-irrigated fields. It is based on the exploitation of local statistics computed, at different scales. The basic idea is that the SSM contrast between irrigated and not-irrigated areas reduces as the timespan between the irrigation event and the SAR acquisition increases. Nevertheless, a significant SSM local contrast after the water supply will last for a timespan that mainly depends on the soil type, on the amount of water received by the field and by the evapotranspiration rate. Clearly, the shorter the SAR system revisit, the higher is the potential to accurately discriminate irrigated vs not-irrigated fields. For this reason, the S1 constellation that has an exact revisit of 6 days represents an important improvement with respect to past SAR systems.
In particular, local statistics are computed on the SSM maps by considering respectively n₁ X n₁ and n₂ X n₂ moving windows, surrounding each pixel, with n₁ < n₂. Then, a pixel X is selected as corresponding to an irrigated area, if and only if the following condition holds

p_(n₁ ) (x≤X)≥p_(n₂ ) (x>X)+TH

where p_(n₁ ) (x≤X) is the probability to find, in a patch surrounding X, pixels exhibiting a SSM value minor than the X one (i.e. it is the local cumulative distribution function computed on a patch of n₁ dimension), p_(n₂) (x>X)=1- p_(n₂ ) (x≤X), where p_(n₂ ) (x≤X) is the local cumulative distribution function, computed on the patch of dimension n₂, TH is a threshold value opportunely set by the user.
A detailed analysis of the proposed algorithm has been carried out on two SSM maps, corresponding to the S1 images acquired respectively on 9th April and 25th August 2017, on the Riaza irrigation district (Spain).
Several observations can be desumed from the experimental test. First of all, the algorithm performances are several biased by the contrast in the SSM maps between irrigated and not irrigated areas. Consequently, in very dry period, characterized by high rate of evapotranspiration, the algorithm performances are low (Overall Accuracy=64%), as can be seen on August 25th. On the contrary, in April, when the evapotranspiration rate is lower, better results have been obtained (Overall Accuracy=79%).
It is also clear that there is a strong correlation between algorithm performances and the gap between irrigation time and S1 image acquisition. Also in this case, the evapotranspiration plays a key role. Other factors, such as the crop, seem to are unrelevant because different crops exhibited the same behaviour of the True Positive Rate respect to the gap between S1 acquisition and irrigation practice time.


[1] Ozdogan, M. et al. Remote Sensing of Irrigated Agricolture: Opportunities and Challenges. Remote Sensing 2010, 2, 2388–2412.
[2] Qiua et al. Comparison of temporal trends from multiple soil moisture data sets and precipitation: The implication of irrigation on regional soil moisture trend, International Journal of Applied Earth Observation and Geoinformation 2016, 48, 17–27.2016.

Area A - Board 34

The microwave-derived near-surface soil moisture and thermal-derived land surface temperature are integrated simultaneously within a new energy balance model. In practice, both information types are used to retrieve the main parameters of soil evaporation (soil resistance, rss) and plant transpiration (Priestly Taylor coefficient, αPT). For that purpose a two-source energy balance model named TSEB-SM is first derived from the TSEB formalism by explicitly representing soil evaporation using the empirical parameters (a_rss, b_rss) of the relationship between r_ss and soil surface moisture. For the fraction cover fc0.5, the model calibration consists of inverting r_ss at Terra and Aqua overpass times. When fc>0.5, the calibration consists in estimating αPT at the daily scale. The procedure is applied over a rainfed wheat field in the Tensift basin, central Morocco. The field was seeded during 3 wheat seasons (September 2014-Juin 2015 (S1), September 2016-Juin 2017 (S2), and October 2017-May 2018 (S3)), while it had remained under bare soil conditions during the 2015-2016 (B1) agricultural season. The mean retrieved values of soil resistance calculated for the entire study period using satellite data are (7.32, 4.58), which are relatively close to those estimated in Sellers et al. (1992) (8.2, 4.3). The calibrated daily αPT for S1 and S2 ranges between 0 and 1.38, and is mainly dependent on the rainfall distribution along the agricultural season. Moreover, the retrieved αPT shows similar dynamics when using in-situ and satellite data, with a relative error of about 11 and 19 % for S1 and S2 respectively. The calibrated daily αPT for S3 remains at a mostly constant value (~0.7) throughout the study period, and a significant relative difference of about 34% is obtained when comparing results using satellite and in-situ data. For all four seasons TSEB tends to significantly overestimate latent heat fluxes. The overall mean bias is 119, 181, 94 and 128 W/m2 for S1, B1, S2 and S3 respectively, when using satellite data. The errors are much reduced when using TSEB-SM witha mean bias of 39, 62, 4 and 7 W/m2 for S1, B1, S2 and S3 respectively. The analysis of the retrieved (arss, brss) and αPT variabilities using satellite data indicate the robustness of the approach, which combines microwave and optical/thermal data to retrieve a water stress indicator at the daily time scale.

Area A - Board 35

The construction of houses and roads on agricultural areas results in permanent reduction of land area available for farming. For countries like Norway with limited agricultural area, protection of agricultural lands form such permanent changes is a national policy priority. Therefore, agricultural areas have to be monitored frequently and any activities involving permanent changes need to be detected as early as possible. The relatively high temporal and spatial resolution of the Sentinel-2 images provides opportunities for monitoring land surfaces. Here, an automated machine learning workflow aided by high quality national geospatial datasets is tested to detect construction of roads and buildings on agricultural areas from sentinel-2 optical satellite images. The training data for the machine learning algorithm is collected automatically using different data sources where the data sources complement each other in filtering and improving the quality of the training data. First, a number of normalized indices such as the normalized difference vegetation index (NDVI), the normalized difference water index (NDWI) and the normalized difference building index (NDBI) are used to define areas that can potentially be used as training data for built-up areas. However, the use of these indices and thresholding includes bare grounds into potential built-up areas. Second, the detected potential built-up areas are filtered using cadaster dataset by excluding areas that do not overlap with the building dataset from the cadaster. The reason why cadaster data alone is not used as training data is due to the potential geometric misalignment with the images. Third, the training data for the other land cover types such as forest, grassland, farmland and water are collected carefully with a conservative procedure from the national detailed land cover map. The Support Vector Machine (SVM) classifier is then trained on the training dataset by fine tuning the different parameters. After the classifier is satisfactorily trained, it is applied to the scene. Built-up areas detected in the regions classified as agricultural area in the national land cover map is further processed. They are filtered and refined using morphological image processing algorithms. The method detects buildings and roads on agricultural land with limited commission of bare grounds. The method has shown great potential for monitoring of agricultural areas. Besides, its fully automated procedure is expected to make it applicable at national level. Further work needs to be done regarding computational efficiency and general improvement of its accuracy. Other machine learning algorithms ought to be explored to optimize both the accuracy and computational efficiency of the method.

Area A - Board 36

Sentinel2A/B (S2) image collections allow for the creation of multiple-year time series with a nominal time sampling of about 5 days in Europe, a bit longer over the rest of the world. From the theoretical point of view, this gives the chance to monitor the phenological evolution of crops through the different seasons of the year. On the one hand, different growing stages of crops over the same area of interest are now observable, both in term of timing and abundance. On the other hand, the usage of advanced analytical tools for crop yield real-time prediction start to become meaningful. The main requirement to envisage such an analysis is the availability of reliable S2 multispectral information during the whole growing season of crops, or at least during the key development stages between steaming and flowering. This knowledge can give farmers the opportunity to intervene if necessary in order to improve the quality as well as the quantity of final harvest. Yet, the cloud-free dependence of S2 useful captures may drastically reduce the number of time samples at disposal of a specific area of interests, with important limitations of remote sensing based recommendation to farming.

The Sentinel1A/B (S1) Copernicus mission provides free SAR data at C-band overall the world with a revisit time over Europe of 6 days, approximately. It is well known that clouds are not an issue for these active sensors, and captures from both ascendant and descendant orbits are useful for surface monitoring purposes. SAR backscattering at a specific wavelength is related to the observation geometry but also to physical properties of the surface, such as soil roughness and moisture and vegetation cover [1]. The presence of vegetation generates the so-called volumetric scattering that characterizes the redistribution of energy among the cross-polar bands with respect to the copolar ones, demonstrating the sensitivity of SAR response on the amount of vegetation onto the illuminated surface. This property has led to the definition of specific Radar Vegetation Index (RVI), when dealing with fully polarimetric data [2], and a simplified version of it for dual-pol data [3].

In this work, we tackle the problem of optics time series gaps over crop areas caused by clouds by establishing a synergy between S1 and S2 data collections. Taking advantage from the similar spatial resolution, we use the time evolution of the dual-pol Radar Vegetation Index to fill the missing vegetation information (NDVI and LAI) from cloudy S2 captures. For this purpose, we employ the so-called Multi-Output Gaussian Process (MOGP) approach proposed in [4] an innovative kernel-based machine learning technique able to stress the dependencies among multiple distinct but related time series. To show the potentials of MOGP, we select a crop area within the Castilla Y León region, North-West Spain. We obtain the S2 collection over the area of interest from the Theia distribution of atmospherically corrected multispectral reflectance data based on MAJA processor [5]. For S1, we download the GRD images from ESA Scientific Data Hub, preprocess them to obtain speckle-filtered dual-pol σ0 data. Once referred to the same projection reference, we apply MOGP at pixel level to reconstruct gap-free profile of vegetation descriptors. We present results showing the effectiveness of the proposed methodology over multiple crop types both at pixel and polygon-averaged level (1D profile) and over the whole area of interest (2D gap-filled images) for both NDVI and LAI estimation. We will also provide first quantitative performance assessment using leave-an-image-out validation technique.



[1] Bousbih, S.; Zribi, M.; Lili-Chabaane, Z.; Baghdadi, N.; El Hajj, M.; Gao , Q.; Mougenot, B. Potential of Sentinel-1 Radar Data for the Assessment of Soil and Cereal Cover Parameters. Sensors 2017, 17, 2617.
[2] Y.J.Kim and J. VanZyl, “A time-series approach to estimate soil moisture using polarimetric data,” IEEE Trans. Geosci,. Remtote Sens., vol. 47, no. 8., pp.- 2519-2527, Aug. 2009.
[3] Y. Kim and J.J. van Zyl, “Comparison of forest estimation techniques using SAR data,” IEEE International Geoscience and Remote Sensing Symposium, pp. 1395–1397, 2001.
[4] Alvarez, M.A., Rosasco, L. and Lawrence, N.D. orcid.org/0000-0001-9258-1030 (2012) Kernels for Vector-Valued Functions: a Review. Foundations and Trends® in Machine Learning, 4 (3). pp. 195-266. ISSN 1935-8237
[5] A multi-temporal method for cloud detection, applied to FORMOSAT-2, VENµS, LANDSAT and SENTINEL-2 images, O Hagolle, M Huc, D. Villa Pascual, G Dedieu, Remote Sensing of Environment 114 (8), 1747-1755

Area A - Board 37

Accurate monitoring of Leaf Area Index (LAI) and Chlorophyll are vital for assessing the vigour of vegetation and reducing the environmental impact of Nitrogen (N) fertilisation in agrosystems. Current imaging sensors offer the necessary spatial, spectral and temporal information to tackle a wide range of problems through Earth Observation, such as crop monitoring. Here, we present an operational chain of retrieval of LAI and leaf chlorophyll content (LCC) from multi-temporal and high spatial resolution data for irrigated rice. The paper aims at testing the best machine learning regression algorithms (MLRA) and the inversion of PROSAIL radiative transfer model in MRLA for the retrieval of LAI and LCC in irrigated rice. Sentinel 2 satellite and a multispectral camera mounted on Unmanned Aerial Vehicle (UAV) were used for multi-temporal retrieval of LAI and LCC at high and ultra-high resolutions respectively. The measured data was collected from January 30th to April 10th, 2018 in Nigeria where a 3X3 factorial experimental plot of 30 ×25m was designed with different water and fertilization levels. The validation of LAI and LCC estimates were collected during the dry season where the collection of LAI was achieved with the LAI-2200 meter while chlorophyll was collected with the atLEAF meter. Data was collected to coincide to the acquisition of Sentinel 2 over the experimental plots. Performances were evaluated over both actual experimental observations and model simulations. To assess the performance of the models, the normalized RMSE and R2 between independent in situ measurements and estimated parameters were used. The paper will demonstrate the importance of irrigated rice in order to address in background condition of water management and nitrogen (N) fertilisation during the various growth stages of rice. From the analysis, the added value derived from high spatial and temporal data from LAI and LCC maps from Sentinel 2 and UAV will be quantified for monitoring rice.

Area A - Board 38

Current operational retrieval services of leaf area index (LAI) are mainly derived from observations of Landsat, SPOT and MODIS sensors. However, the demand for improved services and relevant biophysical parameters requires an enhanced land surface monitoring system in terms of spectral, temporal and spatial coverage. The Sentinel-2 (S2) mission aims to provide such continuity to monitoring services but with advanced features compared to the previous sensors. Thanks to S2 spectral bands configuration, an explicit distinction can be achieved for when crops are still in vegetative, photosynthesizing stage (green vegetation) or when crops are senescing (brown vegetation). While the LAI mapping of green vegetation is well established, that is less the case for senescent vegetation. Because senescent vegetation is spectrally similar as bare soil, current LAI algorithms underestimate LAI when vegetation starts senescing.

With the purpose of exploiting the Sentinels and to enable the retrieval and monitoring of plant biophysical parameters, the HORIZON 2020 Sentinels Synergy for Agriculture (SENSAGRI) project (Grant Agreement nº 730074) aims to exploit the capacity of Sentinel-1 and S2 images to develop an innovative portfolio of prototypes agricultural monitoring services, taking advantage of the synergy of using both types of measurements. One of these new products involves the simultaneous retrieval of both LAI green and LAI brown. SENSAGRI has been running for almost two years now, and operational LAI green and brown algorithms are still under development.

A first step was identifying a suitable operational method that is both: (1) accurate, (2) fast, and (3) provides uncertainty estimates so that the portability of the algorithm can be evaluated. Based on Verrelst et al., (2015) the Bayesian Gaussian processes regression (GPR) has been evaluated as most robust due to their superior accuracies and, because the delivery of associated uncertainty estimates and insight in relevant bands. It is demonstrated that GPR models based on S2 data can also achieve an explicit distinction between LAI green and brown largely thanks to the new S2 bands in the red edge (B5: 705 nm & B6: 740 nm) and in the shortwave infrared (SWIR) (B11: 1610 nm & B12: 2190 nm). Henceforth, operational algorithms for explicit retrieval of LAI green and brown from S2 data can be developed using GPR.

In the v.0.2 prototype algorithm, LAI green model performed more accurate (R2= 0.89, NRMSE= 7.4%), than LAI brown (R2= 0.76, NRMSE= 13.5%). The uncertainty estimates were used to map only those LAI estimated values that fall within an uncertainty threshold of a 40%, as such allowing to transfer the algorithm in space and time. For multiple European core test sites, LAI green and brown composite maps were generated from S2 images, enabling the description of a great spatial variability across the European region. Moreover, a multi-temporal analysis of LAI green and brown retrieval was conducted from a time series images from an agricultural site in Valladolid, revealing the consistency of both models to estimate and represent the transition from green to senescent vegetation.


Verrelst, J., Rivera, J.P. Veroustraete, F., Muñoz-Marí, J., Clevers, J.G.P.W., Camps-Valls, G., Moreno, J. (2015). Experimental Sentinel-2 LAI estimation using parametric, non-parametric and physical retrieval methods – A comparison. ISPRS Journal of Photogrammetry and Remote Sensing, 108, p. 260-272.

Area A - Board 39

In recent years monitoring of earth system dynamics by means of remote sensing satellites has begun to play an important role in the provisioning of spatially distributed input or validation data for earth system and or land surface models. The Global Climate Observing System (GCOS) has identified a set of essential climate variables (ECV), including LAI, FAPAR, and soil moisture. For a subset of these ECVs the European Space Agency’s Climate Change Initiative (ESA-CCI) and the European Commissions’ Copernicus programme are developing and operating retrieval schemes for long term Climate Data Records (CDRs). Often, the retrieval scheme for a given ECV relies on either optical or microwave data streams. Consequently the CDRs tend to suffer from discontinuities or data gaps in the optical domain due to clouds or aerosol contamination. The synergistic use of the combined information content of all available frequency domains (e.g. VIS, NIR, SWIR, TIR, microwave) in a joint retrieval scheme is algorithmically much more challenging but also very promising. It is expected to improve the temporal resolution and accuracy of the CDRs and also ensures physical consistency of the ECVs that are jointly retrieved.
We present results of a joint retrieval of land surface variables over agricultural landscapes such as LAI, FAPAR, vegetation height, phenology and soil moisture by means of a weak constraint variational data assimilation approach. Here we jointly invert two physically consistent radiative transfer models, one for the microwave domain and one for the optical domain, using Sentinel-1 and Sentinel-2 data respectively and employing prior information about the land surface provided by the Joint UK Land Environment Simulator (JULES). In addition we include a regularisation term, which imposes, as a weak constraint, our expectation of the temporal evolution of the land surface state. Outputs are generated in terms of target variables such as LAI, FAPAR, vegetation height, phenology and soil moisture as well as their associated uncertainties.
Results are validated against field measurements over the Munich-North-Isar site, Germany. It will be shown that the retrieval of bio- and geophysical parameters from remote sensing data greatly benefit from a joint retrieval from optical and microwave data. It will be highlighted, that in a joint retrieval scheme, the retrieval of soil moisture greatly benefits from LAI retrievals at the same time. On the other hand, the high frequency of Sentinel 1 acquisitions improves the retrieval of an LAI time-series. In addition it will be shown, that a synergistic retrieval of Sentinel-1 and Sentinel-2 data leads to an uncertainty reduction and allows to provide a consistent retrieval of the spatio-temporal dynamics of land surface variables.

Area A - Board 40

Mapping the crop state along the growth cycle is one of the major concerns for an efficient management of agriculture. SENTINEL-2 observations are particularly well suited for this kind of applications due to its short revisiting time and its decametric spatial resolution. However, crop monitoring requires getting the crop state at specific phenological stages during the growth cycle. Due to the presence of clouds in many areas, large periods without images can be observed in the S2 time series despite the high revisit frequency. The objective of this study is to demonstrate the interest of combining in situ continuous ground measurements and S2 data acquisitions to provide near real time daily Green Area Index (GAI) images.
We present the FieldSensor systems that we specifically designed for agricultural applications. They acquire in situ RGB images continuously along the day and the season, PAR absorption and canopy reflectance in six wavebands. We first study the consistency between S2 pixels values and FieldSensor measurements for both the reflectance and GAI estimates. Assuming that the diversity of FieldSensor time series represents well the diversity of S2 time series, we compare two schemes to combine FieldSensor and S2 GAI time series. The first one applies a correction to the FieldSensor time series based on iterative optimization, the correction being updated at each new S2 acquisition. The second one exploits a Kalman Filter where S2 daily time series are computed as a dynamic linear combination of the FieldSensor measurement and S2 images considering the uncertainties associated to each observation source. Both schemes are then run in backward mode to smooth the temporal profiles.
Results show that both methods provide similar pattern of GAI spatial distribution. However, the Kalman filtering based method provides a more continuous spatial distribution while the other method presents some discontinuities between contiguous pixels. The same is observed when considering the temporal evolution: the Kalman filter shows a smooth temporal profile over the whole season while the first scheme shows discontinuities at each S2 acquisition. Finally, the Kalman filter performs better between two S2 observations as compared to the projection mode, i.e. after the last S2 image. Both methods are sensitive to the quality of FieldSensor acquisitions and GAI estimates.

Area A - Board 41

The Land Parcel Identification System (LPIS) of the European Commission is used for registration of agricultural reference parcels considered eligible for annual payments of European Common Agricultural Policy (CAP) subsidies to farmers. LPIS and the subsidies presume a system of controlling the quality of the management for each parcel. The control has been done through field visits and partially through aerial photographs, both methods being quite expensive. The European Union is interested in a more cost-efficient way of controlling the management of crop fields, e.g., the utilization of the space-borne remote sensing data. SAR data is the only space-borne data source for rapid and almost on-line monitoring of the crop fields during the growing season in the European countries.
Methods to use space-borne remote sensing data to monitor agricultural regimes have been developed with C-band and X-band SAR and C-band InSAR data, e.g. for detecting ploughing, moving or sowing activities. Studies in crop classification with fully polarimetric C-band SAR and optical area data have also been reported. Many studies concentrated on tree orchards, grapes or sugarcanes. A few studies exist with detailed corn species recognition.

The objective of the study is to develop methods based on space-borne SAR data to monitor crops and fields from the point of view of the agriculture subsidies. The objectives include recognition of individual corn species or at least spacies groups and also the management regimes.

Three representative sites were originally selected from Finland located in South (Eura), East (Liperi) and Northern Finland (Oulu). In this study, the first results from the Eura site are presented and discussed.
Eighteen Sentinel-1 SAR scenes (VV and VH polarizations) from 1 June - 31 August 2107 were selected based on the weather conditions, downloaded and preprocessed. The averages and standard deviations of the intensities of the both polarizations were calculated for each agriculture parcel, as well as ratios The Eura site included 10 500 agriculture parcels the total area amounting to 25 000 hectares. The number of different crop species in 2017 was 119.

Two methods, the improved ik-NN estimation with genetic algorithm feature optimization and generalized regression models for categorical variables were tested with optional Sentinel-1 data sets, species groupings and thresholds for the minimum parcel area. The number of species groups varied from 7 to as large as 41, the latter one requested by the agriculture authorities for the Eura site. Leave-one-out cross validation and splitting the data to training data (2/3) and evaluation data (1/3) were used in result validation.

The overall accuracies (OA) somewhat varied depending on the number of species included in the classification, and whether or not all blocks were included exhibiting also dependence on the size of uncluded blocks. OA with nine species groups was around 70% when all blocks were included, around 80% when the block area threshold (for incorporating blocks into classification) was 0.5 hectares and around 90% when the block area threshold was 4 hectares. OA was at the level 82% with 41 species and area threshold of 0.5 hectares. However, the results with 41 species is dominated by a few species with large areas explaining a surprisingly large OA. The OA increased with the number of Sentinel-1 scenes until the early July when the first scenes were from the early June. After that, only minor improvement in the crop recognition accuracy was noted. These results were well-received by the authorities and encourage to continue the study towards an operational method in which the space-borne SAR data is a part of the information chain.

Area A - Board 42