Cereals are the base of a large part of the world's population alimentation, being wheat the most widespread cereal in the world, and the second one in terms of production after corn. Some recent studies claim that wheat provided 20% of the daily calories to 4,500 million people, especially in developing countries. As a result, quantify wheat’s production is essential to support food security management. It can be achieved with agronomical models, which can be developed with the information provided by remotely sensed vegetation indices (VI).The aim of this work was to evaluate the performance of different time series of vegetation indices for the predictive modelling of wheat production and yield in Spain.
The wheat production and yield was modelled using time series of different vegetation indices: MODIS NDVI (Normalized Difference Vegetation Index), MODIS EVI2 (Enhanced Vegetation Index) and MERIS MTCI (Meris Terrestrial Chlorophyll index). The MODIS NDVI and EVI2 time series covered the period from 2006 to 2016 and were computed from weekly surface reflectance products (MOD09Q1 v6) at a spatial resolution of 250 metres. The MTCI time series cover a shorter period (2006-2011) and was computed from surface reflectance at a spatial resolution of 300 m. The gap-filled VI values were smoothed using a Savitzky-Golay filter which included a fitting bias towards maximum values. The sum of IV values of one month after the maximum was used in the modelling as this period correspond with the reproductive and grain filling stages of wheat and it is related with yield and production. The wheat croplands were identified by combining a national database with the boundaries of crop parcels with the information about crop types obtained from the CAP (Common Agricultural Policy). Both information sources were joined to create yearly masks of wheat pixels that were applied to the VI time series. The information about the production (t) and yield(t ha-1) was obtained from the Spanish Ministry for Agriculture at a province level (NUTS3). Therefore, the VI values were aggregated spatially at the province level. Subsequently, a regression analysis was carried out for each NUTS2 region and index. Different models were obtained for the main wheat producer regions of Spain: Castilla y Leon, Andalucía, Aragón, Cataluña and Extremadura.

The production models outperformed the yield models. The percentage of variance explained by production models (R²) ranged from 0.96 (Castilla y León – MTCI) to 0.52 (Andalucía – NDVI), with normalized relative error (NRE) 14.72% and 54.53%, respectively. In relation to yield, models had a R² ranged from 0.85 (Aragón - EVI2), to 0.01 (Extremadura – NDVI), with NRE of 10.57% and 24.54%, respectively. In general, MTCI seemed to produce more accurate models for wheat production and yield, and NDVI had the weakest relation to wheat production. The findings of this work suggest a high potential for the use of MTCI, which could be obtained from Sentinel-2, for the predictive modelling of crop production, and in a broader sense for food security management.

Area A - Board 23

An effective establishment of a winter cover crop is important for reducing nitrogen leaching to groundwater in the maize-based cropping system of the Netherlands. Cover crop establishment after maize cultivation is obliged by law for sandy soils and consequently implemented in nearly all maize fields, but the vegetative ground cover in winter varies significantly between fields. The objective of this study is to evaluate to what extent differences in winter vegetative cover can be explained by the timing of cover crop establishment and weather conditions in two growing seasons (2017-18). We used Sentinel-2 satellite imagery, in combination with the Dutch Basic Registration Parcels (BRP) data, to construct NDVI time-series profiles for parcels known to be cultivated with maize within Province of Overijssel in 2017. We developed a piecewise NDVI time-series smoothing method to fit the time profile of maize-cover crop rotation, from which parameters were then retrieved, including the cover crop sowing date and the fitted NDVI value for 1 December 2017 (NDVIdec). The fitted NDVIdec was used to denote the quality of cover crop establishment in the winter season. A comparison against ground reference data for eight parcels showed that the Sentinel-2 retrieved sowing date was an efficient proxy of on-farm actual sowing date, with low root-mean-squared error (RMSE=6.6 days). To eliminate spatial autocorrelation effects, sowing dates for 500 out of 13980 randomly selected maize parcels were then retrieved, which could explain 38% of NDVIdec variability. Finally, we mapped the spatial patterns of sowing time and its impact on winter cover crop growth for all maize parcels in Overijssel at fine resolution (Figure 1). Besides sowing date, we also explored various other spatially-variable predictors for difference in NDVIdec, such as soil type and parcel planting history. Weather conditions in 2018 were substantially different from 2017 because of low levels of rainfall and high temperatures. As a result, maize harvesting times and cover crop sowing times shifted forward significantly. Although at the time of writing we cannot yet perform a full assessment of the 2018 winter season, our presentation will include also an assessment of the effect of this drought on estimated cover crop sowing dates and the resulting quality of the cover crop establishment. Our results demonstrate the potential of Sentinel-2 image time series for retrieving seasonality parameters for small agricultural parcels, allowing for effective analysis of the effect of farmer practices (such as timing of maize harvest and sowing of cover crop) on cover crop establishment. As such, these provide actionable information both for farmers as well as regulatory bodies aiming to reduce nitrogen leaching.

Area A - Board 58

Motivation and Background: The European Common Agriculture Policy (CAP) was set up five decades ago with the goal of ensuring food security, and to support the farmers and agricultural activities of Member States (MS) by implementing a system of agricultural subsidies. The CAP has responded to the economic, environmental and other challenges that Europe has faced at different times by undergoing several reforms. The latest CAP reform, planned for the period after 2020, will see, among other commitments, a reshaping of the updates in the monitoring procedures of CAP measures, including the procedure for control and monitoring system support (Integrated Administration and Control System, IACS) and the procedure for Control with Remote Sensing (CwRS). The CAP already encourages MS to use Copernicus services (Sentinel data, Copernicus Data and Information Access Service - DIAS) for monitoring the implementation of its measures. However, after 2020 the MS will need to implement some of those updates with reference to the use of common data (Sentinel) to enable better comparability of CAP results achieved by different MS, and to shift from sample inspections to large-scale monitoring.
An important step towards modernization and simplification of the CAP in the post-2020 timeframe is being made with the implementation of the SEN4CAP project by the European Space Agency (ESA) in direct collaboration with DG-AGRI, DG-Grow and DG-JRC. The SEN4CAP project aims at providing the European and national CAP stakeholders with validated algorithms, products and best practices for agriculture monitoring relevant for the management of various CAP measures. The main premise of the suggested monitoring of CAP measures is that the declared agricultural parcel state can be tracked using a proposed color-code expressed as 'traffic lights' for each payment scheme and the related eligibility conditions. The process of allocating traffic lights consists of assigning an eligibility status to an agricultural parcel. The parcel is monitored until a characteristic manifestation through the applied markers is observed that allows a decision to be made for the application year. The markers are customized observations of signs of plant status in relation to the agricultural activities performed (e.g. signs of ploughing, harvesting, grassland mowing, overgrowth). Our study was thus designed in the above respect with aim to help support timely preparation of national authorities (National payment agency responsible) to anticipated changes in CAP reporting and monitoring processes.
The focus of our study was to evaluate the time series approach using Sentinel-2 data for the identification of irregularities in the assessment of the agricultural parcel-based management practice, in a highly fragmented agricultural landscape. The objective is to identify certain signs of ineligible (inconsistent) areas on permanent meadows in one growing season. To better assess the workflow operation we investigate also certain crop fields’ activities.
Appropriate measures of agricultural policy require only mowing to be carried out on permanent meadows (mowing is also time defined and limited), and that crop fields should be ploughed annually or harvested depending on the type of crop planted. Permanent changes in land cover may be the consequence of biotic or abiotic factors, and should be treated in a meaningful manner in the monitoring process. Therefore, when analyzing the detection of irregularities of land use, we are paying particular attention to signs of ploughing or other permanent changes in the land cover in the case of permanent meadows, whereas for crop fields we focus on overgrowing, abandonment or absence of ploughing. In both cases, the observation of the phenomenon is associated with the observation of the continuity in greenness (the unbroken and consistent existence of vegetation).
Data and Methodology: In the study we investigate the analysis of a time series approach using Sentinel-2 images and the suitability of the BFAST (Breaks For Additive Season and Trend) Monitor method to detect changes that correspond to land use anomaly observations in the assessment of agricultural parcel management activities. We focus on identifying certain signs of ineligible (inconsistent) use in permanent meadows and crop fields in one growing season, and in particular those that can be associated with time-defined greenness (vegetation vigor).
Observation of the vegetation cycle and knowledge of periodic processes and causes for their occurrence lead to the understanding and detection of anomalies in the growth cycle. For the purpose of our study we examined the reference state and characteristic development of the growth cycle of a permanent meadow and crop field in one growing season, in addition to the visual interpretation of the NDVI vegetation index time series.
Two important considerations influenced the design of the approach: the selection of mapping unit or unit of phenomena observation (i.e. parcel), and the short timeframe of available data to conduct an analysis of time series between the years.
The main prerequisite for most time series analysis algorithms is a sufficient length of consecutive observations, providing a reference over several years, also referred to as stable history. In this respect, Sentinel-2 data with the available observation period (data delivery) since 2015 enables only a short time series dataset. In the timeframe of this study the observation period is only over three growing seasons. This is identified as a weak starting point in providing sufficient and stable reference/history information (two growing seasons with normal development expected) to find the true trends in an observed year (and thus relevant anomaly detection) for the BFAST Monitor algorithm. Due to this poor starting point for anomaly detection performed on short time series analysis, we introduce two additional procedures to control and verify the BFAST Monitor algorithm results.
Thus in our case, BFAST Monitor was used in combination with customized procedures to control and approve the BFAST Monitor anomaly detection results. Customized procedures are using the analysis of NDVI object-based temporal profiles and computation of pixel-based time-series standard deviation as an indication of NDVI variability in one growing season, where objects of interest are preselected parcels of particular land use: permanent grasslands or crop fields.
Firstly, NDVI layers calculated from pre-processed Sentiel-2 data satellite data enter a BFAST Monitor time series analysis. The advantages of analyzing permanent meadows and crop fields with time series are that they can detect dynamic (short and long term) processes, as well as the magnitude of these changes. Each change of values on a particular image indicates possible changes, from which we can get information on both the locations and times of such changes. Secondly, the objects of potential change, which are the results of the automatic BFAST Monitor time series algorithm, present input data for time series graph analysis, consisting of descriptive statistics on the temporal development of objects. Finally, these results are compared with the time series standard deviation output raster, which is calculated using the NDVI time series stack. The final results of land use type anomalies are selected change objects detected with the BFAST Monitor time series method, which are refined and additionally examined using time series graphs and a standard deviation computation approach. The procedure is semi-automatic and runs separately for permanent meadows and crop fields.
Results: The results indicate that a straightforward time series analysis approach can provide an efficient tool for ineligible land use monitoring of permanent meadows and other agricultural land when the anomaly context is clearly defined (e.g. expected permanent greenness), even with short time series. The complete semi-automatic methodology was developed to obtain results in a relatively fast time and designed in a way that the outputs can serve as a risk parcel warning layer of unjustified use (to support further control procedures of the National paying agency).
Validation: The validation of permanent grasslands land use candidate anomalies in view of land use ineligibilities was performed with the information on declared land annual use and field controls, as obtained in the framework of subsidy granting in Slovenia. The results confirm that the proposed combined approach proves efficient to deal with short time series and yields high accuracy rates in monitoring agricultural parcel greenness conditions. As such it can already be introduced to help the process of agricultural land use control within certain CAP activities.
Conclusion: The advances and limitations of the workflow development, results and workflow assessment focusing on permanent grasslands and crop fields abandoning monitoring will be presented. Also, a further contribution of this study, the assessment of the suitability of Sentinel-2 data for monitoring fragmented and small-size land or habitat areas considering the spatial, spectral and time resolution of Sentinel-2 observations will be discussed.

Area A - Board 59

Remote sensed data in agriculture have been used for several years, with interesting studies and operational applications, but with several limitations. One interesting case is that of the controls to be performed for the Common Agricultural Policy (CAP). The CAP started in Europe around fifty years ago, in order to set of legislation, procedures and tools trying to ensure self-sufficiency of food supply to Europeans as well as sustainability and development of rural areas. According to this mandate, the CAP policy distributes subsidies to each claiming EU farmer, after specific analysis and controls aimed to safeguard the Community Funds from frauds and improper payments. From the CAP Reform of 1992, the management system for national controls foresees, in its workflows, the use of remote sensed data establishing for example, a 5% On-The-Spot Control (OTSC) of aid applications per payment scheme, based on very high resolution satellite data acquired on more than 700 control zones distributed on EU selected areas. This approach, based on spatially sparse sampling in a limited period of the year, was driven by the unavailability in the past years of cost-effective technology and remote sensed. However, with this approach only minimal and localized checks can be made. In addition, it is impossible to verify specific practices (e.g. permanent grassland vs annual, crop diversification analyses, etc.), that could be performed through a phenological analysis of the crop phases, along the entire agronomic year.
The recent launch of the Copernicus Sentinel-1A/1B (S1) and Sentinel-2A/2B (S2)satellite missions, carrying on-board synthetic aperture radar (SAR) and multispectral optical sensors, respectively, has opened new possibilities for agriculture monitoring, providing good spatial resolution together with unprecedented temporal acquisition frequency (including all-weather imaging capability for the SAR) at global scale. In particular, these capabilities will help a transition from a control to a monitoring approach in agriculture, as foreseen in the next 2020+ CAP Reform.
Among the CAP greening measures to be monitored, the maintenance of permanent grassland is important because it helps to reduce soil erosion, ì store more CO2 on ground, ì reduce emissions, etc. Different countries implement different rules to define permanent grasslands, e.g. through the set-up of a minimum time range of persistence (e.g. 5 years) or through a reference date or period for the grassland mowing.
In the framework of the ESA SEN4CAP Project, we have developed and validated a new algorithm (and a tool) for grassland mowing detection based on S1 and S2 image time series, building on the following proofs of concept from recent technical and scientific literature:
- Usage of a wide range of electro-optical and radar derived markers relevant to grassland mowing detection: Vegetation Indexes (VI) from multispectral optical data as Normalized Difference Vegetation Index (NDVI), Leaf Area Index (LAI), fraction of Absorbed Photosynthetically Active Radiation (FAPAR), backscatter and coherence values from SAR data, as generated by the SEN4CAP modules
- Identification of the main disturbances with negative impact on the product accuracy, as cloud coverage for optical data, and rain, morning dew, and soil moisture for SAR data
In the first part of the development activities, we tested different types of approaches and algorithms, in order to choose the most convenient for automatically analysis of S1 and S2 temporal series, taking into account expected algorithm reliability, robustness with respect to the disturbances identified, easy availability of any need ancillary data, robustness with respect to regional variability etc. Change Detection approach was preferred to Machine Learning, mainly for the lack of reliable and geographically well distributed truth dataset (which would be necessary for training a Machine Learning method).
According to these assumptions, two separate mowing detection algorithms, based on S1-derived interferometric coherences and S2-derived Vegetation Indices, have been developed. Then, the two outputs are fused according to specific rules to obtain the final detection.
The S1 algorithm is based on short-term (6 days) coherences time series analysis. For each grassland parcel, the time series is processed to detect a mowing event at the occurrence of a strong increase of the last measured coherence with respect to the previous values.
The S2 algorithm analyzes, for each grassland parcel, the VI time series to detect a mowing event in correspondence of a decrease of the value with respect to an adaptive model, characterizing the VI trend of unmown grassland.
The mowing detection results from S1 and S2 data are fused together according to a set of criteria taking into account the different reliability of the S1- and S2 detections and estimated mowing temporal ranges.
In the prototyping phase of the SEN4CAP project, the algorithm was preliminarily tested and validated over pilot Areas of Interest (AOIs) in Netherlands and Czech Republic. The comparison with the 2017 ground truth from farmers’ mowing dates declarations gave a detection accuracy of about 90%. The project has now entered a 16-month phase of demonstration at national scale, with 2018-2019 grassland mowing products to be continuously generated over six pilot countries (Czech Republic, Italy, Lithuania, Netherlands, Spain and Romania).

Area A - Board 60

Current capabilities of high resolution Earth Observation systems provide considerable amount of data which can be used for various vegetation analysis in both quantitative and qualitative meaning. Therefore, there is a strong interest in linking qualitative products (e.g. crop maps) with quantitative indicators describing the current status of agricultural crops. The main scope of our presentation is development of software system for estimating key biophysical and productivity variables. Such information has huge potential to support the ongoing transformation of agricultural sector towards smart farming and precision agriculture practices.
PROSAIL radiative transfer model was used for simulating databases of spectral signatures which were further transformed to be corresponding to the spectral properties of the Sentinel-2 MSI sensor. The simulated spectral signature were then used for retrieval three key biophysical variables: chlorophyll (Cab) and water (Cw) content together with leaf area index (LAI) using Artificial Neural Network (ANN) approach. Crop specific optimization of the PROSAIL model was performed to get the best possible fit between simulated and real spectra. Six different crops were considered for the parameterization process: winter cereals (winter wheat), spring cereals (spring barley), winter rapeseed, fodder crops (alfalfa), sugar beetroot and maize, as they represents different architectures of crop canopies. Intensive collection of in-situ data was conducted in the vegetation season of 2017 and 2018 to obtain representative datasets for PROSAIL parameterization and for validation of the final products. In-situ data collection was performed repeatedly during the vegetation season to cover different growing stages of the particular crops. Qualitative crop map based on fusion of Sentinel-1 and Sentinel-2 data (together with other supportive layers) is then used as an input for initial decision tree to detect actual crop and its growing stage for each pixel of the processed Sentinel-2 scene. The final products are distributed on several levels (in addition to the original absolute values of Cab, Cw and LAI) from the field level (showing relative spatial gradients within one agricultural parcel) to the farm level (showing the normalized values across the all parcels with the given crop within one agricultural farm).
In addition to estimation of crop biophysics, temporal dynamics of spectral properties was assessed in order to estimate crop productivity parameters. Experimental growths of winter wheat and winter rapeseed were cultivated in specific conditions with different levels of nutrition. In-situ acquisitions of TOC reflectance accompanied with chlorophyll content, LAI and nitrogen uptake measurements were performed here several times during the vegetation season. On top of that, data from yield meters and yield registry were available for ca. 200 standard agricultural parcels. All these data were used to establish relationships between spectral information obtained from Sentinel-2 imagery and productivity parameters such as nitrogen uptake and yield.
The last part of our presentation is devoted to the assessment of the actual crop status relatively to the 30 years long-term normal. Reference temporal NDVI curves were generated for the six considered crops and various climatic regions using the time series of Landsat-4, 5, 7 and 8 data covering the time period of 1985 – 2015 as the crop maps were available for each of these years. Actual NDVI curve is then constructed from Landsat-8 and Sentinel-2 data using spectral harmonization of the Sentinel-2 scenes to be radiometrically corresponding to Landsat data. The comparison of the actual and reference curves is performed continuously during the whole vegetation period in 14-days interval. The output is then expressed as the percentage of the long-term normal.
The development of all products was carried out in close cooperation with the representatives of the target users (farmers, agronomists etc.) to take into account their needs and requests. The overall aim of all these activities was to set-up an operational innovative service providing support to decision making on the farm level, such us variable application and management optimisation of seeding, irrigation, fertilization & crop protection. The service based on the above products is already in place and we will present the feedback from farmers during the conference. Although the current methodological setting limits the use of the products to the region of Central Europe, transformation for the areas with different climatic characteristics is foreseen for the nearest future.

The results were achieved within the framework of the project AGRITAS supported by grant no. TH02030248 of the Technological Agency of the Czech Republic

Area A - Board 61

Diversified agroecosystems become key pillar for sustainable intensification of agri-food systems while ensuring ecological functions under changing climate, diet and demography. The high diversity within agroecosystems is always considered as a fundamental to restore the ecosystem equilibrium through spatial and temporal interactions among the various components of the system. Hence, digitization of the agroecosystems diversity become most essential entry point for any sustainable developmental entities whether it is plant genetics for breeding better varieties, crop diversification and intensification, efficient use of farm inputs, agronomic practices, economic benefits, agroecological value chain to ecosystem services. That leads to delineation of spatial-temporal typologies at all levels from field, farm, landscape and region. Recent advances in Earth Observation System (EOS), Open-Access (OA), Artificial Intelligence (AI), Machine Learning (ML), Information and Communication Technologies (ICTs), Freely available Cloud Computing Platforms (CCP) such as Google Earth Engine (GEE) along with smartphone enabled Citizen Science (CS) are making GeoAgro Big-Data analytics much smarter, interoperable and substantially useful ever before. This has opened tremendous opportunity to address information gaps at multiple levels (e.g., data, yield, ecology, economy, resilience) for implementation and monitoring of demand-driven ecological interventions across the scale (e.g., space, time and package). Ongoing efforts in big-data driven digital augmentation aim at quantifying functional production dynamics and drivers to target site specific developmental interventions and scaling the ecological intensification such as intensification of pulses in cereal based systems, adoption of conservation agriculture, bridging the yield gaps, geo-localization of the research and impact reporting which contributes to 8 of the 17 SDGs. Here we present one of the ongoing efforts in real-time monitoring for agroecological intensification programs targeting pulses intensification in cereal systems in India using ESA Copernicus Sentinel and NASA Landsat series along with Digital Globes GBDX platform for near real-time monitoring of the farming system dynamics for targeting pulse diversification in cereal systems. Approach can be used for other cropping systems, regions in assisting in decision making, policy framework and socio-ecological stability.

Identification and characterization of winter land use is a major component of water quality restoration processes and sustainable management in agricultural landscapes [1]. While optical remotely sensed data have been largely used to determine annual land-cover and land-use [2], they remain underexploited to characterize land-use during winter, due to the lack of cloud-free images during this season. Sentinel-1 and Sentinel-2 sensors, which acquire SAR and optical data respectively with high spatial and temporal resolutions (HRS and HRT), open interesting opportunities for that purpose. The main purpose of this study was to evaluate the ability of Sentinel-1 and Sentinel-2 time-series to understand the temporal behavior of winter land use in agricultural landscapes. The developed methodology was applied on The “Zone Atelier Armorique” study site, which is an agricultural area located in the southern part of the Bay of Mont-Saint-Michel, France. Sentinel-1 and Sentinel-2 time-series acquired during the winter period 2016-2017 have been processed in order to model the temporal dynamics of winter land use classes using optical and SAR parameters such as the Normalized Difference Vegetation Index (NDVI) or the Shannon Entropy (SE) computed from Sentinel-2 and Sentinel-1 time-series respectively. Then, two classifications of land-use during the winter period were performed: one using the Random Forest algorithm (RF), which is an ensemble classifier that uses a set of Classification and Regression Trees (CARTs) to make a prediction [3]; one using Time-Weighted Dynamic Time Warping technique (TWDTW) which is a time weighted version of the Dynamic Time Warping (DTW) method specifically developed for land-use and land-cover classification using remote sensing time-series [4]. Classification results obtained with RF and TWDTW were compared to evaluate the suitability of Sentinel-1 and Sentinel-2 techniques for the classification of winter land-use. The preliminary results showed the respective interest of these time-series but also the potential synergy when using both of them to characterize winter land use. Moreover, the results outline the relevance of TWDTW (classification overall accuracy higher than 85% for 12 winter land-use classes) that takes into consideration the entire temporal behavior of winter land use classes, in comparison with more commonly used classification algorithms.

References

[1] Corgne, S. (2004). Hiérarchisation des facteurs structurant les dynamiques pluriannuelles des sols nus hivernaux. Application au bassin versant du Yar (Bretagne). Norois. Environnement, Aménagement, Société, (193), 17–29.

[2] Fieuzal, R., Baup, F., & Marais-Sicre, C. (2013). Monitoring Wheat and Rapeseed by Using Synchronous Optical and Radar Satellite Data—From Temporal Signatures to Crop Parameters Estimation, 2013. https://doi.org/10.4236/ars.2013.22020

[3] Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324

[4] Maus, V., Câmara, G., Cartaxo, R., Sanchez, A., Ramos, F. M., & de Queiroz, G. R. (2016). A time-weighted dynamic time warping method for land-use and land-cover mapping. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 9(8), 3729–3739.

Area A - Board 63

Globally, droughts constitute a serious threat to food and water security. The complexity and multivariate nature of a drought challenges its characterization, especially on regional and local scales. We used time series from optical (Landsat, MODIS) and SAR (Synthetic Aperture Radar) data to identify the main characteristics of agricultural droughts in Ukraine. Indicators were derived based on optical (Normalized Difference Vegetation Index (NDVI), Normalized Difference Moisture Index (NDMI), Land Surface Temperature (LST), Tasselled cap indices) and Sentinel-1 (backscattering intensity and relative surface moisture) data. We used logistic regression to evaluate the drought-induced variability of remotely sensed parameters estimated for an overall growing season and over its specific period. The variables with the highest prediction rate were further used to estimate thresholds for drought/non-drought classification. The models were evaluated using the area under the receiver operating characteristic (ROC) curve. The results revealed that remotely sensed variables do not respond similarly to drought conditions and specifically growing season maximum NDMI and NDVI (70-75%) and SAR backscatter (60%) reflect the impact of agricultural drought. These metrics also depict stress affected areas with bigger spatial extent. LST was a useful indicator of crop condition as well especially for maize and sunflower with prediction rates of 86% and 71% respectively. When comparing drought characteristics, Landsat and Sentinel-2 NDMI exhibit higher drought severity compared to the other metrics. On the other hand, SAR metrics characterize relatively larger drought duration, especially for maize. Coupled with the declined precipitation during the growing season as well as reduced crop yields, this indicates that drought-induced properties of time series can be characterized based on derived temporal metrics. The findings show that NDMI, SAR backscatter, and land surface temperature could explain drought occurrence and severity. The methods demonstrated here can be applied to other areas requiring early warning of food shortage or improved agricultural monitoring to ensure greater sustainability within the agriculture sector.

Area A - Board 64

The actual extent of irrigated areas and the corresponding irrigation amounts are difficult to assess in most of the countries of the World. In EU regions, statistical data acquired by means of periodical surveys, i.e. Eurostat, are meant to give a picture at the national scale, but they are rather imprecise at regional and local scales, which are the most relevant for the management of water resources in hydrographic basins. The lack of accurate and up-to-date data is hampering the full implementation of the Water Framework Directive (WFD) and its compliance especially in Mediterranean regions. This has been highlighted in Communications from the EU Commission to the European Parliament (see for example the Com.673/2012), which have led to the imposition of additional conditionalities for the utilization of EU funds by Member States in the agricultural sector. For example, the Italian Ministry of Agriculture has adopted specific actions (Decree 31/07/2015) for monitoring irrigation areas and volumes on a regular basis to improve the compliance to the Water Directive. To this end, information provided by Sentinel-2 satellites are representing the most suitable technical basis for developing tools for final users and stakeholders.

The DIANA project (H2020 Grant agreement No 730109; http://diana-h2020.eu/en/) is focused on the development of Earth Observation tools for quantifying on a regular basis the irrigated areas and the irrigation abstractions and delivering these data to stakeholders in a simple and intuitive way by means of the most advanced Information Technologies.
The procedure developed in DIANA for discriminating irrigated and non irrigated crops is based on their different seasonal development or phenology assuming that in arid and semi-arid environment (like the Mediterranean Region), high trends of vegetation growth are only compatible with irrigation. To this end, dense time series of Sentinel-2 data have been processed for monitoring the seasonal pattern of changes in leaf area index (LAI) and Vegetation Indices such as NDVI (Normalized Differential Vegetation Index) and NDWI (Normalized Differential Water Index). This methodology benefits from improved spectral and temporal resolution of Sentinel 2A and 2B; in addition, SWIR bands provide very significant information for better identifying irrigated areas.
The case-study is the Sannio Alifano Irrigation and Land Reclamation Consortium located in Campania region (Southern Italy) during the irrigation season 2018, using dense time series of Sentinel-2A/B images and other ancillary data (ground truths, irrigation scheduling and agrometerological data).
From an operational point of view, the procedure consists of the following stages:
- Preliminary processing, consisting of atmospheric correction of S2 data, cloud masking and gap filling procedures;
- Production of multi-time series of vegetation index NDVI;
- Masking of areas out of interest, typically urban, mountain and wetlands such as rivers, lakes and water basins;
- Unsupervised classification (clustering) applied to time series of NDVI index maps: 70 clusters are generated and the corresponding temporal NDVI index pattern are generated, in fig.1 there are typical examples of irrigated (left) and non irrigated (right) pattern.
- Labelling of vegetative areas, identification of classes of vegetation categories that, given the water deficit conditions, show NDVI pattern compatible only with irrigation;
- Supervised classification: this is done using such training pixels of irrigated areas identified via the multi-time classification of NDVI index combined with on field inspection and ground-truth data; this phase is aimed at improving the accuracy of the unsupervised classification step;
- Mapping of irrigated areas along with possible integration with other GIS data (such as maps of parcel holding irrigation water rights).
Different machine learning supervised classification algorithms have been tested in order to analyse the impact in term of overall accuracy by introducing additional information, such the soil moisture index, derived on SWIR bands with the algorithm OPTRAM [1].
The quantification of irrigation abstraction has been carried out by applying the Penman-Monteith equation, with appropriate values of the canopy parameters namely the canopy resistance, crop height and surface albedo [2]. The canopy resistance is calculated from the Leaf Area Index, also derived from S2 data, and a stomatal resistance which takes into account the soil and canopy water status from soil moisture index of OPTRAM. From a comparative analysis it has been proven that the estimations of water volumes from Sentinel-2 are close to the data provided by in-situ metering devices of the Sannio Alifano consortium; furthermore, S2-based monitoring has an operating cost which is one or two orders of magnitude smaller than field installations.
The procedure is now currently used in the Campania region for updating every year the actual consistency of irrigated areas and the corresponding water volumes, in line with the conditionalities for the EU Water Directive compliance.

[1] Sadeghi, M., Babaeian, E., Tuller, M., & Jones, S. B. (2017). The optical trapezoid model: A novel approach to remote sensing of soil moisture applied to Sentinel-2 and Landsat-8 observations. Remote Sensing of Environment, 198, 52-68.
[2] Vuolo, F., D’Urso, G., De Michele, C., Bianchi, B., Cutting, M. 2015. Satellite-based Irrigation Advisory Services: A common tool for different experiences from Europe to Australia. Agricultural Water Management, 147, 82–95.

Area A - Board 65

The Normalized Difference Vegetation Index (NDVI) is an index of plant “greenness” or photosynthetic activity, and is one of the most frequently used vegetation indices. Vegetation indices are based on the reflection that different surfaces imitate different types of light differently. By taking the proportion of red and near infrared bands from a satellite image, an index of vegetation “greenness” can be defined. The Normalized Difference Vegetation Index (NDVI) is probably the most common of these proportion indices for vegetation.

Many factors affect NDVI values like plant photosynthetic activity, total plant cover, biomass, plant and soil moisture, and plant stress. Because of this, NDVI is correlated with many ecosystem traits that are of interest to researchers (e.g., net primary productivity, canopy cover, bare ground cover). Also, because it is a ratio of two bands, NDVI helps compensate for differences both in illumination within an image due to slope and aspect, and differences between images due things like time of day or season when the images were acquired. Thus, vegetation indices like NDVI make it possible to compare images over time to look for ecologically significant changes. Vegetation indices like NDVI, however, are not a remedy for rangeland valuation and monitoring.

A Primary conclusion can be derived that NDVI value is a number which mainly derive from satellite image based on the greenness of the crops. This value is just a coefficient which mainly helps us to determine the estimation of yield and pixels under same crops. We cannot predict crop name or type or crop‘s health by analysing NDVI values, we only can estimate the crop’s vegetation strength while it is predicted the crop type by analysing spatial reflectance. NDVI values also vary based on the crops’ growth stage. Seeding and harvesting stage is mostly having less NDVI and middle of cultivation is having more NDVI generally than seeding and harvesting stage but this can vary based on different parameters effect on crops.

Area A - Board 66

Agriculture is characterized by a high variability in both space and time as it highly depends on seasonal patterns, landscape, climate as well as policies and management practices. Land use conflicts, intensive freshwater use and other environmental concerns pose conflicts of interest to agricultural production. At the same time, a projected global population of 9.7 billion in 2050 and an increasing use of energy crops, will result in a drastically rising demand for agricultural products in coming decades. It is therefore crucial to ensure a timely and accurate monitoring of crop production on large scales (regional, national or continental). This has been attempted using various techniques, for example, in the Crop Growth Monitoring System of the European Joint Research Centre. To extend crop production monitoring outside of regions with high data availability, it is necessary to conceive approaches that can function with limited data input and independent of existing agricultural monitoring programs.

Optical satellite remote sensing data is particularly suited for timely cropland observations. It delivers data on a fairly regular schedule, under comparable conditions and on a global scale. Secondary products such as vegetation indices are in common use and relationships to various plant characteristics, including growth patterns, stress reaction and crop yield, were proven in many studies.

Thanks to open data policies of space agencies like ESA and NASA, we encounter an unprecedented availability of free earth observation data. This enables large-scale applications making use of spatial coverage and long time series instead of single image processing. Most existing approaches are, however, held back by computational and/or data requirements limiting their ability to scale to large datasets: (a) they require a large number of different input variables and extensive knowledge (or good approximations) of location, crops and growing conditions within a field, or (b) they rely on extensive statistical datasets to establish robust empirical relationships. Both are often unavailable, especially when moving from single fields to regional scales. Furthermore, data gaps and the trade-off between spatial and temporal resolutions can aggravate coverage of physiological changes during main growth stages. Thanks to their twin-satellite setup, ESA’s Sentinel-2 satellites currently offer unprecedented possibilities for vegetation and crop monitoring.

Traditionally, agricultural yield prediction relies on simple but effective agrometeorological indices such as the ubiquitously used Growing Degree Days (GDD). Many indices have been proposed throughout the decades (e.g. Physiological Days or BioMeteorological Time), often being arguably more complex but still with very low data requirements. In many cases, only basic weather information is needed, for example daily temperature measurements that are available globally (though in different abundance and quality). This makes agrometeorological techniques particularly valuable in regions were detailed measurements or extensive agricultural statistics are unavailable. Some attempts have been made to merge agrometeorological and spectral indices to combine their respective strengths for crop growth analysis and crop yield prediction. For example, by attempting predictions about future vegetation index developments based on GDD time series or relating ratios of spectral indices and temperature sums to historical yields. So far, much of the efforts are focusing on the Normalized Difference Vegetation Index and GDD.

In our study, we want to expand on these efforts by exploring further agrometeorological and spectral indices to enable crop yield estimations. Building on established approaches, we will offer a technique capable of representing regional patterns and applicable to large scales, while relying on readily available earth observation and meteorological data. This ensures applicability even in regions where a dense network of observations is not available. By making use of seasonal index time series it can benefit from higher data availability and temporal resolution of systems such as Sentinel-2. Thanks to its simplicity, though, it will be transferrable to different sensors and a variety of meteorological products, including rasterized weather datasets.

Area A - Board 67

The Copernicus Sentinel missions have brought Earth Observation into a new era, especially its agricultural applications. The extremely short revisit cycle combined with high-resolution multispectral observations strongly suit the needs of vegetation monitoring with the help of red-edge bands. Assimilating such large volumes of data into any dynamic crop model to predict yield is, however, extremely challenging.

In this study, we 1) propose a data assimilation scheme that assimilates 10 meter resolution reflectance into a coupled crop growth–radiative transfer (RT) model, and 2) generate weather driver dataset by ECMWF numerical weather forecast, which in turn allows us to 3) predict wheat yield with a 10 meter resolution dynamically with quantified uncertainties. Data assimilation by optimization is a time-consuming method, particularly when conducting assimilation procedure in every 10 meter pixel. Hence, we divide a temporally stacked multi band dataset into blocks, and then conduct data clustering within each block. In this way, massive multiband pixels are assigned to a limited number of categories. Next, the average reflectance of each clustering category in each block is assimilated into a coupled WOFOST-PROSAIL model. Once the assimilation is completed, WOFOST model can predict yield of current season by using optimized parameters and weather forecast. In addition, we are able to update the yield prediction on a daily basis as new weather forecast and Sentinel-2 data becomes available, and we can provide the uncertainty of yield prediction with the help of ensemble weather prediction inputs.

We focus on North China Plain as our main study area, specifically Henan province, which is the largest wheat producing province in China. The yield prediction of 2017 shows that by assimilating Sentinel-2 reflectance data into coupled RT-crop models, the accuracy of yield prediction improves substantially. Compared to Landsat and MODIS data, Sentinel-2 data are more robust in data assimilation due to lower mixed-pixel effect resulting from its high spatial resolution. The results also suggest that higher observation frequency is beneficial for agricultural data assimilation.

Area A - Board 68

In Belgium as in other countries in Western Europe, nitrogen (N) recommendation methods for crops at field scale have been based for years on the provisional N balance sheet approach. The presence of a winter cover in the crop rotation has a strong impact on the N balance. In addition to its positive impact on soil, winter cover crops limit leaching of nutrients and supply up to 45 units of nitrogen per hectare to the following summer crops, depending on the winter cover crop type and its biomass production. Classically, this biomass production is estimated using 3 classes : 0-1, 1-3, 3+ tonnes of dry matter per hectares (DM t/ha).

According to the Belgian technical and extension service structures delivering N recommendation to farmers, accurately collecting such kind of information, often visually assessed by farmers, is complicated.

The recent launch of new satellites such as Sentinel 1a/1b or Sentinel-2a/2b, dedicated to agriculture monitoring at field scale, open new opportunities to collect free, accurate and operator-independent information contributing to relevant N recommendations.

The capabilities of these satellites to retrieve the winter cover crop biomass have been assessed. Biomass samples were collected in 1 m² quadrat for various types of winter cover fields (mustard, phacelia, oat and a range of mixtures) in 2016 and 2017 (yield between 0.1 to 5 tons of dry matter per hectare). Empirical relationships between the winter cover crop biomass and a wide range of vegetation indices (VI) derived from Sentinel-2 have been defined and the best VIs identified. Considering the importance of cloudy conditions hampering the use of optical data in Belgium, the empirical relationships have been also defined using SAR (sentinel-1) data. At last, each combination of the best indices have been tested to establish the best empirical model. First results show relatively good adequacies between the winter cover crop biomass and S2 vegetation indices such as the MTCI (R²=0.72 ; cross validation RMSE=0.57 DM t/ha) and the MSI (R²=0.73 ; cross validation RMSE=0.56 DM t/ha).

The added-value of using objective satellite-based estimation of winter cover biomass instead of subjective visual estimation from farmers has been also assessed by comparing both estimations with estimation from samples collected in the field, considered as reference data.

Area A - Board 69

Plant growth and yield models have become an important tool for supporting efficiency and sustainability related decision-making processes in agricultural land management systems. Simulating plant physiological processes in dependence of environmental and management related factors, plant growth models offer information on the development of crops over the growing period. On a global and regional scale this information can be used to understand crop reactions to climatic changes. On field-scale farmers can use this information to analyze site-specific crop development and identify possible problems at an early stage. Regardless of the scale of application, model outputs can be used to find mitigation options and adjust agricultural management measures accordingly.

With a worldwide production of 377 Mio. tonnes in 2016, potatoes are the number one non-grain food crop and fourth most important crop in general. Due to their high nutritional value, strong adaptive capacity and comparably uncomplicated and cheap way of cultivation, they have become an important food security crop. According to former director-general of the Food and Agriculture Organization of the United Nations (FAO), Jacques Diouf, potatoes are considered to be “on the frontline in the fight against world hunger and poverty”. However, despite their integral significance in the context of food security, potatoes have not received the same amount of attention as other crop types when it comes to quantitative modeling and many of the potato models that have been developed in the past are poorly validated. Furthermore, most of the existing potato models do not offer analyses on a spatial scale.

The PROcess and radiation Mass and Energy Transfer model (PROMET) is a spatio-temporal hydrological, physically based land surface model. It uses information on environmental drivers (temperature, precipitation, radiation, humidity, wind, atmospheric CO2), soil properties, terrain, and management actions (e.g. sowing date, fertilization events, irrigation) and applies biochemical, physical and plant physiological principles in order to simulate crop growth in hourly time steps. Information on the spatial variability of the crop dynamics on sub-field scale is retrieved by incorporating Earth Observation data into the modeling process. Satellite data is used to derive time series of plant parameters, e.g. Leaf Area Index (LAI), by SLC (soil-leaf-canopy) radiative transfer model inversion. The resulting parameter maps are then assimilated into the plant growth model. PROMET has proven its ability to accurately model spatially distributed plant growth and yield for several crops. It has been intensively tested and validated for winter wheat and maize and has furthermore successfully been applied and tested for rapeseed, barley, sugar beet and further crop types.

To investigate whether the PROMET modeling approach is applicable to potatoes, a model version parameterized for potato crops is tested in this study. PROMET plant growth simulations are conducted for 19 potato fields (variety Fontane) in Bavaria, Germany. Sentinel-2 (10 m spatial resolution) and Landsat 8 (30 m spatial resolution) imagery covering the whole vegetation period is used for SLC model inversion. Resulting LAI maps are assimilated into the plant growth simulations. To assess the accuracy of PROMET output, measured yield data for the respective fields (provided by the Kartoffel-Centrum-Bayern) are compared to modeled yield. Preliminary analyses show a strong correlation between modeled and measured yield levels (R²: 0.74, RMSE: 6.18 t/ha) and thus indicate that the PROMET model works well also for potatoes.

PROMET’s site-specific simulation results give insights on heterogeneities in crop development on sub-field level not usually available for potato crops. This opens up new possibilities to improve efficiency and sustainability in potato crop cultivation by adopting site-specific management practices.

Area A - Board 70

ABSTRACT

In this work, a new way to extract temporal polarimetric information from a stack of co-registered images is presented. This method considers not only polarimetric evolution of consecutive acquisitions but also includes polarimetric changes between every image with respect to the rest of images in the stack.

The methodology is tested for different crop types exploiting C-band quad-polarimetric RADARSAT-2 data over rice fields in Seville, South-West of Spain and the Indian Head in Canada as part of the Agrisar 2009 campaign.

1. INTRODUCTION

Traditionally, the extraction of polarimetric observables is done from the covariance or coherency matrices either at pixel or at parcel level. This process is repeated for each image acquired and the polarimetric observables are plotted as time varying variables.

The proposed methodology uses a polarimetric change detector as the core of its operation. It is based on the analysis of a pair of co-registered acquisitions by optimizing the pixel-wise difference of their covariance matrices. The optimization process uses the diagonalization of this difference matrix as described in [1,2], to find the set of unitary vectors that causes the largest polarimetric change between the pair of images.

The output provides information about the type and intensity of the scattering mechanisms that were added and/or removed from one image to the other. This can also be seen as finding the scattering mechanisms that increased and decreased between images.

The scattering mechanisms are represented as RGB composites where each element of the eigenvector takes a colour and the contrast of the image is modulated through its associated eigenvalue.

The methodology is initially applied to observe the temporal evolution of one crop type and subsequently the potential for further applications is introduced including usefulness for crop type mapping an crop parameters retrieval.

2. TEST SITE AND DATASETS

The methodology is first applied to rice fields located near Seville, in the south of Spain. Ground truth for six parcels was gathered for the season running from May to October of 2014, including phenological evolution, height of plants and length of the leaves, among others.

The satellite data used is composed of 16 C-band quad polarimetric Radarsat-2 Images acquired with three different incidence angles during the 2014 season, covering all the campaign.

3. INTERPRETATION OF CROP EVOLUTION BASED ON POLARIMETRIC CHANGES

The crop evolution can be described as a continuous change of the crop morphological parameters, including changes in the plant’s height, leaves size and the plant’s density and orientation, among others.

Since the change detection procedure compares pairs of images to find polarimetric changes between them (e.g. scattering mechanisms added and removed in the scene), we can associate the polarimetric changes with the changes in morphological parameters that drive the crop evolution. Previous
work introducing this concept has been presented in [3]

3.1 Multi-temporal change matrix

Since the information about the crop evolution is obtained from the analysis of pairs of images, it is possible not only to evaluate changes from consecutive images (as traditionally done) but is also possible to see the difference of an image with respect to all the other images in the stack. This process identifies the change of the scattering mechanisms throughout the season.

The information can be represented by a square matrix, where each element of the diagonal represents one of the acquisitions. The off-diagonal elements correspond to the evaluation of polarimetric changes between images.

The upper triangular part represents the increase in scattering mechanisms between a given pair of analysed images and the lower triangular part represents the corresponding decrease.
It is worth noticing that the unitary vector that represents a scattering mechanism is again represented as an RGB composite using its associated eigenvalue to modulate the contrast.

3.2 Rice crop change matrix

The result of this process is shown in the figure 1 attached. The figures (a) and (b) represent the change matrix for two different parcels and figure (c) represents a change matrix obtained from averaging the change matrices of the six parcels where ground truth is known.

A remarkable characteristic of the matrix is that its colours and intensity provide information of the scattering mechanisms during the season and they can be interpreted as a conventional RGB composite, that is, red for double bounce, green for volume and blue for surface scattering.

It can be seen from figure 1(a) and 1(c) that these change matrices have minor differences corresponding to the normal differences in the parcel’s evolution.

However, the general patterns as well as the colours and intensity are preserved. The parcel (b) on the other hand, presents atypical evolution of scattering mechanisms.

It is important to notice that because the matrix contains separately increase and decrease in time of each scattering mechanism, the analysis of parcel evolution can be done for each mechanism separately. This results in being able to compare for instance if the double bounce in a parcel due to the plants emergence above the water level is delayed in time with respect to a reference parcel.

The purpose of obtaining an average change matrix as in the figure 1(c) is to obtain a reference of a rice parcel such that one can compare any parcel against this reference and analyse that parcel’s evolution during the season, even if no ground truth is known for that parcel.

Furthermore, since each row and column of the matrix is associated to a day of the year, the change matrix immediately tells the scattering mechanism present at any given moment during the season (it can also be associated to a ground truth variable e.g. phenological stage at a particular day).

An initial method tested to analyse differences between parcel’s evolution is by determining the distance between the Frobenius norm of each parcel’s change matrix with respect to that of the other parcels as shown in figure 1(d).

As an interesting result, the distance of the Parcel B and the parcel F (not shown here due to lack of space) compared to the other parcels and the average parcels is the largest. This can be seen analyzing their change matrices, as the evolution was delayed. It can also be confirmed when looking at the yield from ground data of these parcels.

3.3 Change matrix for different crop types

The same methodology was applied to several crops including cereals such as barley, oats and wheat and other crop types such as canola, and field peas. The crops are located in the Indian head in Canada, where ground truth was gathered as part of the Agrisar 2009 campaign.

A stack of 10 Radarsat-2 quad pol images was analysed, and the resulting change matrices are shown in the figure 2.

It is possible to see by visual inspection how different the resulting matrices are for each of these three crop types and to that of the rice crop presented in figure 1.
These differences, described by the colors that represent the scattering mechanisms, the contrast of the image and the times when specific events occur, can be used as a feature for crop type classification.

As a basic preliminary test, again the Frobenius distance between 10 different crops is shown in figure the 2(d). It can be seen how the distance between the
same crop type is the lowest, e.g. Barley 1 to barley 2. Similarly, distance to crops of the same family such as cereal crops (Barley, Oats, Wheat) is lower compared to that of the canola or field peas. In fact, it remains below one.

On the other hand, distances from barley to Peas and canola are separable between them with Pea between 1 and 2 and with Canola having a distance of more than 2.

4. FURTHER APPLICATIONS
The following applications are being implemented but not included here due to lack of space:

• Different ways to extraction the information contained in the matrix are being developed to provide new features for agricultural monitoring.

• Initial results confirmed the correlation of the volume scattering mechanism contained within the matrix and the phenological stage provided by the ground truth. Further experiments are being developed to also corroborate correlation with the leaf area index.

• The effect of the incidence angle during the season can be clearly seen using the change matrix, explaining which scattering mechanism is affected at what point in time. This can be used to analyse incidence angle normalization techniques over agricultural fields.

• The matrix size increases with every new acquisition. Therefore, real time monitoring can be achieved by extracting information from the matrix when new data is obtained.

• This methodology is currently being transferred to the dual polarimetric case and being tested for Asparagus crops in South America using sentinel 1 data.

5. REFERENCES
[1] A. Marino and A. Alonso-Gonzalez,” Optimisations for different change models with polarimetric SAR,” EUSAR 2018; Aachen, Germany, 2018, pp. 1-5.

[3] C. Silva and A. Marino and J. M. Lopez-Sanchez and I. Cameron, Monitoring agricultural fields using an optimisation of the difference of covariance matrices for POLSAR, (IGARSS), 2018

Area A - Board 71

The use of satellite data for in-season crop monitoring is severely hindered in countries with extensive cloud cover. The Sentinel-1 constellation offers unprecedented SAR data, regardless of cloud coverage, every six days. The acquisition frequency of these data is essential for in-season crop status mapping and yield prediction, which is key for decision-makers within the food supply chain to provide instrumental information throughout the growing season. Agrimetrics was set up with initial funding from the UK Government to catalyse the agri-food sector by inspiring and empowering businesses to use data to solve global food challenges. Within this framework, we are developing a SAR-based tool to offer in-season, national crop maps for the UK.

An automatic, object-driven segmentation algorithm has been developed for the automated delineation of fields over the UK. The application of this algorithm mapped 1.45 million fields covering England and Wales for the cropping season of 2016-2017. SAR data were extracted on a field basis from Sentinel-1 imagery for the same period. Each field was characterised by the radar parameters derived from the extracted data, including terrain-corrected backscattering coefficient, interferometric coherence and polarimetric decomposition derived parameters like entropy and anisotropy. Machine learning algorithms were trained to classify crops at four stages in the 2016-2017 cropping season.

The overall accuracy of the results increased throughout the growing season from approximately 65% to 80%. Current efforts are devoted to training the classification algorithm with data more representative of the spatial and temporal variability of the crop types. In addition, we are working on methods to improve classification accuracy by incorporating ancillary data from non-Earth Observation sources.

At the conference we will present our results indicating the likelihood of the different crop types on a field-by-field basis. These estimates will be updated every three months throughout the growing season and published through our online platform.

Area A - Board 72

Lodging is one of the yield-reducing factors in wheat, causing losses of up to 80%. Lodging also has several knock-on effects such as deterioration in grain quality, slower harvest and increased drying costs. The severe effects of lodging can be minimised by detecting its incidence in time so that decisions related to expected crop yield, crop price or its insurance can be made timely. The traditional methods of lodging assessment require the on-site assessment to rate lodging state visually or through manual measurements of the inclination angle of the plant base from the vertical (lodging angle). However, such field estimates can be biased and subjective since they depend upon the skill of the observer or the complexity of the observation. A rapid and quantitative evaluation of crop lodging at field/farm scale can contribute to more accurate assessment of yield loss. The aim of this study is to apply remote sensing (RS) information to estimate the crop angle of inclination (CAI) and detect the stage of lodging (slight lodging, moderate lodging,…). Such quantitative estimates related to crop condition can facilitate prediction of yield losses. Studies show that yield losses incurred at 80o can be almost 2-4 times than those at 45o in wheat.

This study investiagtes the ability of multi-temporal Sentinel-1 along with multi-incidence angle and fully polarimetric Radarsat-2 SAR data for estimating CAI and to assess the lodging stage. Ground measurements of CAI, height and other structural/biophysical properties (e.g. LAI, biomass) were made on wheat in the Bonifiche Ferraresi farm (Jolanda di Savoia, Italy) from October 2017 to July 2018. The crop lodging condition (lodged/non-lodged and stages) was also recorded. A total of 35 Sentinel-1 and 20 corresponding time-series Radarsat-2 single look complex SAR images were acquired (coincident to the field measurements) covering all the growth stages of wheat. The Radarsat-2 data was acquired in four beam modes: FQ3, FQ8, FQ21 and FQ26. Both empirical and semi-empirical approaches were tested to test the angular responses of radar backscatter and SAR polarimetric parameters derived from Sentinel-1 and Radarsat-2 for the wheat canopy. Significant correlations were observed between SAR polarimetric parameters/backscatter ratios and the CAI of wheat from both sensors. From this analysis it was possible to assess crop condition in a scale comprising of five different lodging stages: upright (angle up to 4o from vertical); leaning crop/slight lodging (5-40o); moderate lodging (41-60o); severe lodging (61-75o); and very severe lodging (76-90o). This study is a step towards quantitative evaluation of lodging stages and further illustrates the capability of SAR data for lodging assessment.

Keywords: lodging, wheat, crop angle of inclination, Radarsat-2, Sentinel-1

Area A - Board 73

Considering the high importance of the agricultural sector, the European Union (EU) implements a set of regulations under the Common Agricultural Policy (CAP), aiming at the sustainable development and management of natural resources. Additionally, in case of a natural disaster, EU and each state member offer recovery assistance to disaster-affected farmers to mitigate the impact of the problems caused. The successful implementation of CAP and insurance coverage require precise and regular mapping of crops as well as monitoring of natural disasters effects by carrying out frequent and accurate controls. Until now, in Greece for instance, these controls are mainly performed by in-situ visits resulting in very high administrative costs. This procedure is not only time-consuming itself, but certain factors, i.e. adverse weather conditions, inaccessible areas, etc., can further delay subsidies/compensations. Some efforts have started recently to involve remote sensing data in these procedures. However, their usage is hindered by the limited availability and high cost of very high-resolution satellite images along with processing difficulties. Specifically, in these cases data with sub-meter spatial resolution are exploited by interpretation techniques, which are highly subjective and time demanding.
Free and open access policy of the Copernicus Sentinel data offers a big volume of data to the users on a constant and complete basis. Due to their various features and the frequent revisit time, Sentinels support efficiently many environmental applications, including monitoring of large agricultural fields and mapping of flooded/burnt areas extend. However, the effective handling of the available long time series is still at an early stage and the development of accurate automated techniques is limited.
To this end, the Disaster and Agriculture Sentinel Application (DiAS) project revises the existing knowledge on remote sensing methods for monitoring agricultural land and proposes improvements. The goal is to assist the responsible Greek services to take decisions on farmers’ subsidies and compensations. The developed methods will be implemented in a Decision Support System, which will be freely available and easy-to-use by non-experts. DiAS is expected to improve the efficiency of the decision-making process on agricultural aid and reimbursement. Exploiting free satellite data will reduce the number of in-situ visits, minimizing costs and delays of controls. The automation of the satellite image processing will contribute to the immediate decision making and respond of the responsible authorities towards the sustainability of agriculture.

Area A - Board 74

Key words: CAP, agriculture, crop detection, SITS, machine learning

Within the framework of EU Common Agriculture Policy (CAP), farmers can have access to subsidies from the European Union, that are provided through Paying Agencies operating at national or regional level. For the provision of the subsidies, Paying Agencies must operate several controls in order to verify the compliance of the cultivation with EU regulations. At present, the majority of the compliance controls are limited to a sample of the whole amount of farmers’ declarations due to the increased costs of acquiring high and very-high resolution satellite imagery. Moreover, they are often focused on a specific time window, not covering the whole lifecycle of the agriculture land plots during the year.
The free and open availability of Earth Observation data is bringing land monitoring to a completely new level, offering a wide range of opportunities, particularly suited for agricultural purposes, from local to regional and global scale, in order to enhance the implementation of Common Agricultural Policy (CAP). Nowadays, satellite image time series are increasingly used to characterize the status and dynamics of crops cultivated in different agricultural regions across the globe.
This paper proposes a fuzzy-based technique for crop detection and monitoring in Romania, based on combined free and open Sentinel-2 and Landsat-8 Earth Observation data, developed within the framework of the European Union’s Horizon 2020 DataBio project. The main goal of the approach is to provide services in support to the National and Local Paying Agencies and the authorized collection offices for a more accurate and complete farm compliance evaluation - control of the farmers’ declarations related to the obligation introduced by the current Common Agriculture Policy (CAP).
The analysis was performed for an area of interest situated in Southeastern Romania, with a surface of more than 10 000 square kilometers, including more than 10 000 parcels, which was chosen through a multi-criteria analysis performed in a GIS environment, based on the 3 criteria: surface, diversity and accessibility of the land plots.
The general methodology is based on the comparison between real crop behavior and the expected trends for each crop typology. It involves image processing, data mining and machine learning techniques and is based on different categories of input data: Sentinel-2 and Landsat-8 SITS covering the time period of interest, farmers’ declarations of intention with respect to crops types, as well as in-situ / field data. The processing chain involves a series of well-defined steps: image preprocessing (numerical enhancements for Sentinel-2 and Landsat-8 scenes, ingestion of external data, clouds and shadows masking), individual scene classification, followed by the use of unsupervised machine learning techniques in order to obtain the crop probability maps at scene level. In the end, time series analysis allows the generation of overall crop probability maps and derived products.
The system produces three main types of results, all provided at a 10-meters spatial resolution:
- Crop Mask Maps, which are pixel level maps, identifying some of the most important crop types;
- Parcel Use Maps, which are object-based maps, showing the most probable type of crop at plot level;
- Crop Inadvertencies Maps, which can be both pixel-based and object-based maps, revealing the areas for which the declared type of crop included in the LPIS appears to be different from the identified one. The pixel-based analysis states whether pixel values correspond or are different from typical spectral values of the declared crop types, whereas the object-based analysis reveals the plots for which the declared type of crop appears to be different from the one identified based on satellite imagery, based on a specific threshold. The parcel-based analysis involves a relative pixel count within the parcel boundary, in order to determine whether the parcel corresponds or not to the declared crop type.
The validation of the preliminary results against independent sources revealed promising results, with an accuracy higher than 90% for more than 10 crop types.
The highly-automated proposed approach allows the performing of big data analytics to various crop indicators, being reliable, cost- and time-saving and allowing a more complete and efficient management of EU subsidies, strongly enhancing their procedure for combating non-compliant behaviors. The developed technique is replicable at any scale level and can be implemented for any other area of interest.

Area A - Board 75

An agricultural subsidy is a governmental subsidy paid to farmers to supplement their income, manage the supply of agricultural commodities, and influence the cost and supply of such commodities. In the European Union, the agricultural policy is called the Common Agricultural Policy (CAP) and it implements a system of agricultural subsidies and other programmes.

On member state level, the CAP funds are managed by national paying agencies. These are responsible for both making the payments to the final beneficiaries and controlling that no undue payments are made. In Finland this duty is carried out by the Agency for Rural Affairts (Mavi). The controls in Finland are currently carried out as various cross checks between farmers’ applications and Mavi register data, but also as on-the-spot-checks (OTSC) in the field.

The current system for controls is in need of development. The new system should not only be more cost efficient but also faster, in order to enable Mavi to make payments on time. This study will aid these objectives by providing information based on remote sensing for the control process. The more specific aims are the classification of tillage practices, the classification of crop groups and locating anomalous parcels.

The study area covered the entire Satakunta province in South-Western Finland. Ground truth data consisted of over 5000 parcels which were checked in-situ. Remote sensing data included Sentinel-1 (S1) SAR-images (12 day difference, 11 dates from 2016 and 19 dates 2017) and Sentinel-2 (S2) image index (NDVI, NDTI, NDMI, NDBI, NDSI) mosaics from autumn 2016 to autumn 2017.

Here are some examples of preliminary results. Different tillage practices were classified into four classes: 1. grass, 2. ploughed, 3. lightly tillaged, and 4. autumn cereals. The classification was performed using the k-nearest neighbor classifier with k varying from 3 to 15 and using leave-one-out error estimation. As basis for the classification different combinations of S1 images and S2 image indices from autumn 2016 and 2017 were used. So far the best overall accuracy is about 95% when using S2 image indices from autumn 2016 and spring 2017 (April and May) in combination with S1 images .

Crop group classification (1. grass, 2. peas, 3. oil plants, 4. spring cereals, 5. root plants, 6. broad beans, 7. fallow, 8. potato, 9 autumn cereals) were performed using S1 time series for the year 2017. The figure 1 illustrates the behavior of overall accuracy as a function of time when using both VV- and VH-polarized parcel mean values in the classification. The classification was performed using the k-nearest neighbor classifier with k varying from 3 to 15 (differently colored lines in the figure) and using leave-one-out error estimation. At the start of the growing season overall accuracy is about 60% and increases as more images are added to the classification. 85% overall accuracy is reached at late June and 90% at late July. The next step will be to study the classification accuracies obtained when using S2 image indices and combine the S1 and S2 time series in one classification.

One way to find anomalous parcels is to compare the average value of a parcel to the average value of all parcels declared by the farmers to be of the same crop. In other words, the average values and standard deviations of all crops are computed for the whole study area. Then, the mean value of each individual parcel is compared to the mean and standard deviation of the same crop. In the figure 2, parcels colored in red indicate parcels where NDVI is considerable (2*std) lower than the mean of that crop, yellow slightly smaller and blue slightly higher. Thus, on-the-spot-checks could be targeted to red parcels in order to investigate the reason for the large deviation from the crop mean.

Area A - Board 76

As part of the upcoming reform of the Common Agricultural Policy (CAP) in 2020, the satellite Earth Observation for crop monitoring will play an increasing role. In this context of future agriculture monitoring by Remote Sensing, accurate and up-to-date digital field boundary maps are required to be incorporated in further analyses, such as crop type classification or ploughing detection.
The present study proposes a simple, operational and automated method for detecting and updating changes in the delineation of agricultural parcels over time as an important pre-requisite for accurate and efficient crop monitoring system. Being identified using Sentinel 1 and 2 satellites, the level of detail of detected changes is directly compatible with next steps of monitoring.
The study site is located in the Walloon Region, South part of Belgium. In 2018, it counts 286288 parcels declared by farmers for CAP aids that covers 772000 hectares. The main crops are grassland, winter wheat, maize, sugar beet, and potatoes, which correspond to 41%, 16%, 7%, 5.7%, and 5.6% of the total agricultural area respectively.
Using Sentinel-2 derived NDVI and Sentinel-1 backscatter coefficient time series, the current agricultural parcels – defined as plots cultivated with only one type of crop and managed by only one farmer – that don’t match with the last declaration of the farmers in the Walloon Region can be targeted and their boundaries updated.
The changes, i.e. the differences in the boundaries of the agricultural parcels over time, are detected in two steps.
First, an automatic method based on detection of discrepancies in time is used. The discrepant parcels are defined as fields presenting inconsistencies between a vector delineating the field boundaries digitized at one moment of the year and a raster data acquired later. This method directly compares a vector layer delineating the agriculture parcels declared by farmers with Sentinel-2 derived NDVI images in order to detect the inconsistencies between them. The workflow includes firstly the use of two simple algorithms based on the analysis of NDVI pixels statistics inside each parcel and one algorithm enhancing the local contrast to target the plots that present discrepancies with a high probability. Then, changes in parcels delineation are detected and the boundaries updated using a segmentation algorithm, where discrepancies are flashed.
Secondly, an object-based classification algorithm using Sentinel-1 and Sentinel-2 images is used in order to detect crop type of the new expanded plots.

Area A - Board 77

For many years, the correct management and compliance control of area-based payments to the European farmers, provided through the EU Common Agricultural Policy (EU CAP), have relied on two main processes: (1) administrative (alphanumeric and geospatial) cross-checks of all annual farmer applications and (2) a sample of checks “on-the-spot” to verify the conditions under which aid is granted. These eligibility conditions relate to the (1) correct location and quantification of the agricultural land cover, available for payment and (2) the presence of the agricultural activities (land use) required by the particular Regulations. On-the-spot checks were done either by a visit in the field, or by using remote sensing. The latter implies the use of satellite imagery of high and very high resolution and a common control methodology, based on computer-assisted image analysis. Although procedurally different, both methods rely on a visual and cognitive interpretation of a given land phenomenon by a trained operator of a number of discrete observations in time (number of field visits or acquired images). The decision, whether the eligibility conditions and the farmer obligations are met on the spot, depends on expert judgment taken by inspector.

The extension of the scope of the EU CAP towards environment and climate performance, by introducing the “greening” measures, increased the complexity of the administration and management of direct aids. It poses particular challenges to the on-the-spot control setup, which despite its effectiveness to capture and quantify the agricultural land cover, was far less performant to observe the continuous dynamics of the farming activities. The number of satellite images, even combined with rapid field visits, were found insufficient to capture any temporal agronomic phenomenon (sampling problem). Also, even if acquired within a carefully selected set of acquisition windows, the acquired images often failed to evidence important dynamics of observed land phenomena occurring during the agricultural season (quantization problem).

The arrival of the Copernicus Sentinels and the new data capturing and processing technologies, such as geotagged photos and cloud computing, acted as a true game changer for the EU CAP community and the compliance control setup. The continuous and regular data flow from Sentinels available at country scale, complemented by field data, could finally overcome the shortcomings of the on-the-spot checks by providing the necessary set of observations for a correct interpretation of the characteristics and temporal behaviour of the land phenomena. It also allows the application of compliance checks over all annual farm applications. Certainly, it would require a new approach towards data processing and analysis, since it would no longer be possible to rely on experts’ photo-interpretation of hundreds of images on thousands of agricultural parcels. Instead, the process would highly rely on automation. However, the application of the traditional automated data extraction methods focused on the production of thematic mapping products, would not suffice. The EU CAP compliance control would require the near real-time capturing of particular events that can be associated to particular farming activities. There is a need for the definition of new semantic concepts for characterization of the land phenomenon in standardized and machine – readable manner, in order to establish the link between the behaviour of the land phenomenon and its manifestation through the Sentinel “signal”. This could be enabled through the introduction of the “marker” concept.

Markers can be considered as a unique collection (combination) of property values of a data signal that evidence the existence or presence of a particular continuous state or a change of state of the land phenomenon. Markers describe the telltale behaviour of the land as picked up by a particular signal. A dictionary defines marker as something that demonstrates the existence or presence of a particular quality or feature. In this technical context of agricultural land monitoring, such definition can be interpreted to be a “structured record of an observation on the signal”, with ISO 19156 defining observation as “act of measuring or otherwise determining the value of a property”.

Markers can be seen as an analogue to the photo-interpretation key in classical remote sensing: they explain and add meaning to an objective observation of the signal. The central component of the marker, called core, is a pure remote sensing observation of a land cover temporal event i.e. behaviour detectable within a certain signal (e.g. a continuity, a change, a transition, an appearance or a disappearance of vegetation or other land cover). Essentially, a core describes in which signal a particular event is best detected, e.g. the bare soil index signal reveals the presence of bare soil when positive for two consecutive data captures.

For the marker to become operational, i.e. act with the precision needed for the local context, restrictions can be applied on the core through the use of parameters. The intensity, the starting date or the duration of a particular behaviour is not uniform over the globe. A phenomenon like “fast vegetation growth” will appear differently on a signal taken in the Mediterranean than on one taken in the Baltics. In mountainous areas, a phenomenon like “appearance of vegetation” will be delayed for parcels in elevated positions compared to those located in the valley below. The marker values (or amplitude of the signal) of the same crop can vary inside and between regions and can change from one year to another. The parameters and threshold values of the marker allow the core to be fine-tuned to accommodate these differences and thus make any marker applicable in different regions.

Parametrization of the markers is not a straightforward task as it requires combination of data and observations from various sources, e.g. the crop calendar in the particular agro-climatic region and the statistical information derived from the image time-series from previous years e.g. with machine learning algorithms based on neural networks.

The proposed paper introduces the concept of the marker and describes the design of the marker components (marker core and its parameters) through the semantics of the Land Cover Meta Language (ISO 19144-2) and the TEGON strata structure to better tackle the dimension of the physical feature observed and for automation of the semantic classification.

The paper further proposes a method for instantiation and parametrization of the markers based on real case examples of Sentinel 1 and Sentinel 2 data. The methodology focuses on application of the concept for each of the main crop groups and associated farm activities, relevant from the point of CAP operational eligibility checks by monitoring, at field level and on national scale.

Area A - Board 78

Time series of satellite images offer opportunity to evaluate the land use dynamic and monitoring the crops with different spatial and temporal resolutions. Remote sensing images provide precise information on crops with low time, cost and resources. Spectral variables from different satellites (e.g, Landsat, Modis, SPOT) have been used to forecast crop yield, mainly with vegetation indices, using regression models and more recently machine learning algorithms. We used Landsat images and random forest algorithm to forecast sugarcane yield in a small region in Brazil. The study area is located on the west of São Paulo state within an area of 50,000 hectares. Landsat images from sensors OLI, ETM+ and TM were combined to create monthly composite image to eliminate clouds and cloud shadows. The monthly composite, made on Google Earth Engine (GEE), used the median values of the available pixels with high quality. Six spectral indices and four spectral bands were computed. Field data from 2012 to 2016, obtained from a sugarcane company, included information related to soil type, sugarcane variety, planting date, harvest date and yield at field scale within a GIS. A global model, calibrated on the 5 years (2012 to 2016), was created using random forest algorithm and applied on independent data previously removed from the total dataset. For each year i, 4 month time series were used from december of year i-1 to march of year i, which corresponds to the wet season. The 4 month time-series of the 10 spectral variables and 16 field data variables were used as explanatory variables (56 variables in total). A cross-validation test was used to evaluate the model prediction based on coefficients of determination (R²), root mean square error (rmse) and the index of agreement modified (dmod). Relative importance of the predictor variables was calculated. The global model showed good yield prediction ability with on average of R²= 0.76, rmse = 7.5 ton/ha and dmod = 0.74. Such prediction is comparable to the ones found in literature (R² from 0.65 to 0.78 and rmse above 7 ton/ha), and to the model used by the company for its own prediction. The most important variables were the development stage of sugarcane areas, followed by the spectral variables from Landsat images. The vegetation index and bands from SWIR and NIR spectral regions were the more important spectral variables, especially the indices NDMI, EVI and NDWI1. Months of december of the year i-1 and january and february of the year i were the most important ones. Different prediction models allowed to test the improvement of combining satellite images, climate dataset, soil dataset and crop characteristics: the use of satellite images always improved the yield prediction. The potential for forecasting sugarcane yield with Landsat images can be improved with Sentinel-2 images mainly for the better spatial, temporal and spectral resolution and that are not cover by the Landsat sensors. The global model provided good results for forecast sugarcane yield in local scale and has potential to be applied on future years, helping the management and monitoring of crop production, essential for industries, producers and policy makers.

Area A - Board 79