The European grassland-based farming systems face numerous challenges to fulfill (and strengthen) its multiple functions as a supplier of forage and ecosystem services such as ensuring the maintenance or enhancement of biodiversity in a context of social and environmental change. There is a strong need to rethink the existing management of permanent grasslands across Europe to halt the loss of biodiversity and still maintain or restore livelihoods that depend on these systems. This would require the integration of novel knowledge and technologies for farm-level management of biodiversity and ecosystem services with the existing regional scale tools for monitoring and decision-making.

Copernicus Sentinels together with other EO-derived open data provide the capability for monitoring of natural grasslands with an unprecedented degree of temporal resolution. The consistent availability of historic (archive) Sentinel data from the last couple of years also created the source for observing trends that occurred in natural grasslands regarding their biodiversity status and their capacity to provide the relevant provisioning and regulating ecosystem services. Previous R&D projects developed prototypes for downstream services based on Sentinel precursors for monitoring of farmland abandonment or assessment the status of high nature value farmland, that could capitalize from the availability of the Sentinel data to deploy operational solutions.

Certainly, the accurate assessment of the changes occurring on natural grasslands, as well as the assessment of the long-term trends, would require the establishment of a correct baseline, derived from existing reference datasets. This is particularly relevant for those natural grasslands that consist of an intrinsic mix of herbaceous and scattered woody vegetation, since they are more prone to loss of biodiversity due to encroachment of unwanted vegetation, if abandoned. Such grasslands would require also more complex monitoring techniques, due the complexity of the vegetation physiognomy and phenology.

As the majority of these natural grasslands are within the scope of the area-based aids of the EU Common Agricultural Policy their spatial extent and typology is recorded in the relevant geospatial registers of the integrated system for administration and control – known as the Land Parcel Identification System (LPIS). Furthermore, due to the need for correct quantification of the eligible area available for payment, the nature of the scattered woody vegetation and its distribution within the grassland area is well documented following a common standardized methodology developed in-house by the Joint Research Centre.

Natural grassland with scattered non-agricultural (ineligible) features poses mapping challenges. The technical guidance developed by JRC proposes a mapping methodology based on photointerpretation of up-to-date orthophoto to reduce the risk of an incorrect quantification of eligible area and to provide simplification, accountability and precision. It sets out a dual framework wherein, pending on the local conditions, either one of two methodologies can be applied:

● systematic approach when the deduction of the scattered features can be determined based on known characteristics of well identified land cover classes.
● sporadic approach for conditions where only an individual, direct assessment of the proportion of grass is available.

Experience gained from the operational implementation of the methodology called for a revision of the methodology in order reflect the feedback and recommendations of the expert community, as well as to take advantage of the availability the Copernicus Sentinels. To address these challenges JRC conducted a pilot study.

The paper proposed consists of three parts. The first part provides a brief description of the JRC methodology for assessment of natural grassland with scattered woody vegetation. The second part presents the design of the pilot study and its results. The third part outlines revisions planned in the JRC guidance and the possible role of Copernicus Sentinels. It proposes that Sentinels could assess on annual basis the ratio between the wooded and herbaceous vegetation, which acts as a proxy indicator for biodiversity and ecosystem services capacity.

Area A - Board 16

The protection of farmland biodiversity has become a key issue in Europe. Particularly, extensively managed semi-natural grasslands constitute key habitats for many species. Over the past decades, semi-natural grasslands have been increasingly threatened by agricultural intensification and abandonment.

In order to sustain the remaining semi-natural grasslands, up-to-date information on their status and their extent is urgently needed. So far, agricultural statistics do not discriminate between different types of grassland, and conservation oriented schemes, such as Natura 2000, focus exclusively on those grassland habitat types with high conservation priority.

As field data collection is time and resource intensive, using remote sensing data to map natural and semi-natural vegetation is gaining importance. However, the increased availability of remote sensing data from different sensors, with higher spatial, spectral and temporal resolution, requires informed decisions on which sensors to use and how to process the data. Therefore, this study aims to assess the potential of Sentinel-2 and RapidEye remote sensing imagery combined with ancillary data on topography and vegetation structure to differentiate semi-natural grasslands in upland areas.

As ground truth data, we use an extensive field survey of grassland plant communities in the Thuringian forest, an upland area in central Germany that is characterized by a complex terrain and a fine-scale mosaic of different vegetation types. A total of 1142 managed grassland sites were surveyed within the study area using a line transect approach, covering 2,701 ha. The main management activity carried out in the surveyed grasslands was low-intensity grazing, most frequently by cattle. Based on a phytosociological approach and expert knowledge, grasslands were classified into different types.

Several remote sensing data sources are analyzed for their performance in differentiating these grassland types, namely Sentinel-2, RapidEye and LiDAR data. The Sentinel-2 and RapidEye time series capture the phenological development of the grasslands, and provide a high spectral (Sentinel-2) and spatial (RapidEye) resolution. LiDAR data are used to include information on the vegetation structure and surface roughness.

We analyze to which extent the grassland types are distinguishable with these data sets. We further explore which spectral features are most relevant for the differentiation of grasslands, if multitemporal data and phenological indices increase the classification accuracy, and how data on topography and vegetation structure contribute to the differentiation. First results of this work in progress substantiate the great potential of multispectral remote sensing for mapping grassland types, and the benefits of combining different data sources.

Area A - Board 17

Permanent grasslands play an important role in agricultural landscapes of the temperate zones. In Europe, permanent grassland accounts for approximately one-third of all agricultural areas. Besides their landscape-shaping character, they provide a range of other important ecosystem services. Grasslands contribute to the regulation of the natural environment for example by carbon sequestration or rainwater filtration, as well as by the provision of habitats to numerous plant and animal species. Further, grasslands are important for the production of milk and meat as they provide a high-quality basic feed without huge costs of production and transportation. To maintain all of these different aspects of permanent grasslands, site-specific, environmentally friendly but also market-oriented management is crucial. In order to provide practitioners with spatially explicit decision support information, it is necessary to increase our general understanding of grasslands and how relevant information can be regularly derived in a synoptic manner.
For this purpose, the satellites of the Copernicus program seem promising as they provide data at unprecedented spatial and temporal resolutions that are necessary to capture the short-lived management interventions in temperate grasslands. For this purpose, we used dense time series of Sentinel-1 and Sentinel-2 data with a regional focus on permanent grasslands in Germany. The dense time series allowed us to derive generic phenological metrics, such as start, end, and peak of season, but also to develop grassland-specific metrics that were specifically designed to describe biomass development and management events. As cutting is a major determinant of grassland use intensity, we used the number and timing of mowing events as a proxy for land use intensity, but also as a starting point to derive further metrics. Commonly used time series metrics to assess the productivity of vegetation are seasonal integrals, which are less suited for intensively managed grasslands, as management events lead to several partial integrals with varying base values throughout a season. Therefore, we developed new productivity metrics that are based on the accumulated increase of a vegetation index through time, in relation to the previous mowing event.
The combination of all derived metrics allowed us to map grasslands and to categorize them in terms of their use intensity and productivity using a cluster analysis. Based on these information practitioners may already be able to make information-based management decisions for the upcoming season. However, as our results stress benefits as well as limitations of Sentinel-1 and Sentinel-2 data for mapping grassland specific metrics, they indicate future research pathways that may focus on the development of near-real-time applications.

Area A - Board 18

Collection of information in the context of agricultural grassland through fieldwork can be costly and time-consuming. Therefore, remote sensing should be incorporated into monitoring programs as a cost/time-effective alternative. Previous studies have taken advantage of the spatiotemporal dynamics of satellite imagery for agricultural cropland yield estimation. Nevertheless, due to the frequent cloud cover in space-borne optical imagery throughout the growing season, time series analysis is difficult. The mentioned limitations can be handled by using high resolution multi-temporal weather- and illumination-independent Synthetic Aperture Radar (SAR) data.

We demonstrate a comprehensive comparison between the two most commonly used feed-forward neural network models, MultiLayer Perceptron (MLP) and Radial Basis Function (RBF), as well as the Support Vector Machine (SVM) algorithm for automatic grassland monitoring using SAR and weather data. Furthermore, robustness of different machine learning algorithms for grassland monitoring (yield estimation and cutting date detection) is evaluated.
The input variables for all models include 22 weather parameters, VV/VH polarization backscattering intensity, and second-order texture metrics for the detection date calculated for both VV/VH channels. Biomass estimations were performed using SAR Intensity data alone, and sequentially adding second-order texture metrics, temperature, precipitation, Global Radiation, and evapotranspiration to assess the influence of different sets of inputs on model accuracies.

Among the presented models, MLP and RBF NNs proved to be more stable and accurate than the SVM algorithm. The MLP NNs presented a somewhat better performance than RBF NNs. Training of MLP NNs was more time-consuming in comparison to RBF. Initial location selection of radial basis centers in RBF NNs and adjustment of C (rate of classification) and γ (inverse of the radius of influence of samples) in SVM, however, proved to be very difficult. We have achieved the best results for input sets containing SAR intensity, second-order texture metrics, temperature, precipitation, global radiation, and evapotranspiration parameters. Consistent with the obtained results, a decrease in the number of input parameters led to a strong reduction in the data dimensionality. Overall, the neural network’s ability to successfully generate nonlinear estimation models, is a great advantage over commonly used statistical or biophysical approaches.

Area A - Board 19

Alpine meadows and pastures fulfil numerous key environmental processes such as carbon and water storage, soil stability or flora and fauna habitat. Recently, alpine grasslands are exposed to changes in climatic conditions such as later frost events in spring or drought periods, higher temperatures and less precipitation in summer. At the same time grassland ecosystems are exposed to diverse land use and management processes such as intensification, leading to multiple and earlier harvesting processes. These processes often occur within a short time span and - especially in the alpine environment - differ locally due to various topographic, meteorological and geographic factors. It becomes necessary to implement innovative and stable approaches for continuously monitoring changes within alpine grassland communities and to overcome sensor-specific drawbacks such as temporal, spectral or spatial resolutions. The technologies and subsequently the available sensors for monitoring alpine grassland communities have increased over the past decades. The enhanced internet connectivity and the development of low-cost sensors have led to the installation of numerous sensor networks. At the same time, the Copernicus initiative comprising the operationalization of the Sentinel satellites has increased the availability of imagery for monitoring grassland. However, their potential of synergetic use for deriving spectral or biophysical information across scales is by far not yet fully exploited. This study is focussing on a synergetic use of four different sensors detecting on separate scales ranging from ground measurement to remote sensing imagery for assessing the Phytomass development and harvesting periods by combining a spectral index (NDVI) and a modelled biophysical variable (LAI).
We used a ground spectrometer, station-based spectral reflectance sensor (SRS) measurements together with fixed installed digital imagery (Phenocam) and Sentinel-2 MSI acquisitions to derive NDVI observations on different scales. Researching on alpine meadows we laid the main focus on understanding possible drawbacks resulting from topographic and geographic properties including five different grassland sites spread across South Tyrol and split into meadows and pastures. Additionally, we modelled LAI maps based on Sentinel-2 imagery using the Radiative Transfer modelling of PROSAIL as implemented in the ARTMO Toolbox. The resulting information are statistically confronted to in-situ measurements of both Phytomass and LAI taken in 2017 (n=126) and 2018 (n=38) using Random Forest Regression. By combining spectral NDVI and modelled LAI variables we aim at evaluating and further redefine the derived LAI maps and to extensively study the detectability of LAI and Phytomass for alpine grasslands.
By comparing the sensor specific NDVI measurements we noticed that the spectral signatures among sensors behave similar but have bigger differences when it comes to offsetting and saturation of sensors as well as the detectability of short-term events such as snow or harvesting periods. This is mostly due to the spectral properties of the sensors as well as the viewing angles. The NDVI signal reacts differently with the clearest change of signal in SRS and Sentinel-2 signal, depending on the number and frequency of acquisitions. First modelling of the LAI indicates that the over- and underestimation of the LAI represents a big challenge, especially when deriving biomass uptake and the exact occurrence of harvest dates. Also, steepness and exposure of the grassland generate uncertainties in the LAI maps and has to be specifically addressed and parametrized. Even though the NDVI tends to over and underestimate the actual measured LAI and phytomass, first Random Forest statistics reveal a good estimation of LAI values on alpine meadows (R2 ~0.7). The Sentinel-2 MSI sensor and the SRS tend to have the highest impact on the models whereas the Phenocam tend to have the least significance for the Random Forest model.
This study shows that the combinability of spectral information in a unified framework with diverse sensors on diverse spatial scales relies heavily on their spectral, geometric and temporal resolution but also on their viewing geometry. Especially when analysing short term changing events these factors have to be closely taken into account. Nevertheless, first results indicate that the combined use led to a good predictability of both LAI and Phytomass for alpine grassland sites.

Area A - Board 20

Agriculture is a critical source of employment in rural Colombia and is one of the sectors most affected by climate and climate change and where solutions to key challenges affecting the productivity and sustainability of forages and the livestock sector are required. Brachiaria grasses are the most economically important forage in Colombia, where they are planted on 4 Mha, accounting for approximately 90 percent of all sown pasture land. This research aims to develop remote sensing-based approaches for forage monitoring and biomass prediction at local and regional levels. Local access to such information can help improve decision making and increase productivity and competitiveness while minimising impacts on the environment. A multi-source approach is adopted as persistent seasonal cloud cover drastically reduces the number of multispectral spaceborne observations that are available and incorporating the use of multi-temporal Synthetic Aperture Radar (SAR) data offers an attractive solution for providing recurrent and up-to-date information on forages over large and remote areas due to their independence of weather conditions. Unmanned Aerial Vehicle (UAV) acquisitions were also incorporated into the analysis to determine the interrelationship between in situ, aerial and spaceborne observations and their sensitivities to forage quantity and quality metrics.

Between July and November 2018, an intensive field sampling campaign was carried out at a trial plot in Palmira, Valle De Cauca department. 36 forage samples across 9 subplots were acquired on a weekly basis with corresponding UAV acquisitions. This was supplemented by additional sampling in September 2018 from five smallholder farms at Patía, Cauca department, including: Brachiaria brizantha cv. Toledo, cv. Marandú, Brachiaria hybrid cv. Caymán, Mulato II and Megathyrsus maximus cv. Mombasa forage types. These areas all differed in terms of local climate, forage types and management practices. Sentinel-1 and -2 acquisitions on or near (within 5 days) the day of field measurements were acquired and a series of regression models (e.g. Random Forest, Lasso) were applied with varying input parameter permutations. As part of the study, the investigation was carried out in cooperation with smallholder farmers to determine their attitudes and potential constraints to mainstreaming such technologies and their outcomes on the ground. Through improving communication between earth observation and agricultural communities and the successful integration of satellite-based technologies, future strategies can be implemented for increasing production and improving forage management while maintaining ecosystem attributes and services across tropical grasslands.

Area A - Board 21

Permanent grasslands (meadows and pastures) are the most common agricultural land use type covering 34% (0.65 million hectares) of agricultural land in Latvia. The Common Agriculture policy (CAP) stipulates that the EU Member States have to designate permanent grasslands, ensure that farmers do not convert or plough them and that the ratio of permanent grasslands to the total agricultural area does not decrease by more than 5% in order to receive support payments. However, semi-natural grassland habitats require appropriate management activities to ensure their long-term conservation. The European Commission report (2015) required by the Birds and Habitats directives concludes that ‘grasslands and wetlands have the highest proportion of habitats with an unfavourable-bad and deteriorating status’ in the EU, while the midterm review of EU biodiversity strategy 2010-2020 highlighted that grassland habitat change presents a high risk to biodiversity. Latvia’s rural development programme (2014-2020) has identified only 47 thousand hectares of biologically valuable grasslands. These grasslands are semi-natural meadows and pastures that include species and habitat types of EU importance. 70-90% of EU importance grassland habitats in Natura 2000 sites were in poor condition in Latvia during 2012. There is a clear interest from a number of end-users (e.g. the Nature Conservation Agency, the Rural Support Service,) for grassland mapping and management practice monitoring solutions.
Development of Sentinel-2 optical and Sentinel-1 radar satellite data based approach for monitoring of permanent grasslands was the goal of the SentiGrass project (2016-2018) implemented within the ESA PECS for Latvia program. Mapping of permanent grasslands is of primary interest and such data product represents areas covered with non-disturbed grass type vegetation during the whole observation period. It requires at least one satellite data scene during spring and autumn seasons when arable agricultural lands usually get ploughed, thus allowing detection clear separation between grasslands and agricultural lands. High classification accuracy 90...95% could be achieved by using Sentinel-2 optical or Sentinel-1 radar data. Frequently cloudy sky in the Northern Europe limits the use of optical data, however our studies from 2015-2018 show that at least one cloud free scene per spring and autumn season could be obtained, thus allowing monitoring of permanent grasslands at least twice a year. Ploughing events and other disturbances resulting in grassland cover loss could be clearly distinguishable using also more frequent Sentinel-1 radar data - repeated acquisition every 6 days. Therefore, Sentinel-1 radar data could be used for permanent grassland monitoring on a weekly bases.
Grasslands are also one of agricultural land use types and could be declared as permanent or cultivated ones in order to receive subsidy payments. There was an interest from the Rural Support Service (responsible institution for implementation of the CAP in Latvia) to test Sentinel data-based cross-check of declared agricultural land use types in 2017. In such case, the interest is to verify agricultural land use type (including grasslands as one of the types) within known polygons. The pilot study was performed in a territory corresponding 35VMD Sentinel-2 tile or 17% of the territory of Latvia. Summer 2017 was extremely cloudy and rainy, therefore only two low cloud-cover Sentinel-2 data scenes were available – one from the spring and another one from the end of summer, thus clearly showing problematics working with optical data in the Northern Europe. On the contrary, continuous nearly 6-day data flow was provided by Sentinel-1 radar satellites. Nevertheless, Sentinel-2 optical data showed better grassland classification accuracy (96%) in comparison to Sentinel-1 radar data from two different obrites (85...89%). However, combined use of Sentinel-2 and Sentinel-1 data showed slightly improved performance in grassland classification (97% accuracy) and significantly improved classification of other crops, resulting in >85% accuracy for most of crop types.
Sentinel-1 data is useful to frequent monitoring of grasslands when it is necessary to detect a certain event (ploughing or mowing). However, the advantage of spectral optical data is its discrimination ability between vegetation groups, allowing detection of species invasion. In the case of Latvia, Heracleum Sosnowskyi is an invasive specie and tends to occupy large areas, therefore could be detectable using 10...20 m/pix spatial resolution data from Sentinel-2 satellite. Our study shows that detection accuracy of Heracleum Sosnowskyi could be >90% in open areas for relatively dense stands.

Area A - Board 22

Land use / Land cover (LULC) assessment in the Sahel is essential for many reasons, including cropland mapping, forage monitoring, carbon, water and nitrogen cycles understanding and many other environmental diagnostics. LULC mapping is notoriously difficult in this region because rain-fed crops and rangeland phenology changes from year to year following the rainfall season variability and because fallow, rain-fed crops and rangeland are not easily distinguished. Moreover, bushes and trees may be scattered all over the fields and fallows and LULC changes rapidly from one year to another. Here we evaluate the capability of Sentinel2 data processed with the IOTA2 (https://framagit.org/inglada/iota2) land cover map production chain to derive LULC in a typical Sahelian landscape.

The study site located in the Fakara area (east of Niamey, Niger) covers an area of 3700 km² too large to be extensively observed by field surveys. Multi-temporal satellite remote sensing approaches can capture agro-pastoral dynamics provided that the temporal and spectral signatures are distinct at least during part of the year.
Sentinel-2 data have characteristics adapted to the description of this environment (5-day revisit time, 10 m spatial resolution for wavelengths suited to vegetation analysis). A time series of 22 images acquired between September 2017 and January 2018, from the end of the rain season to the end of the dry season, with cloud cover below 50% have been processed using IOTA2 and data from a long-term field survey (AMMA-CATCH observatory).
The analysis evaluates the contribution of multi-temporal data sets according to the number of images used and the 19 ground truth themes (combination of vegetation type and soil background characteristics) with a random forest classification. The percentage of correctly classified pixels is between 40% and 90% depending on the theme and the number of images used. The prospects for large scale classification of Sahelian LULC are discussed.

Area A - Board 23

Spatially-explicit information about grassland forage quality and quantity is of pivotal importance for the sustainable management of wildlife grazing by herbivores such as red deer (Cervus elaphus). The availability and spatial distribution of potential forage areas affect the spatial distribution and activities of herbivores, and thus their influence on the ecosystem by e.g. trampling, grazing and nutrient transfer. Remote sensing offers unique possibilities to derive biophysical information about vegetation conditions over large areas. The radar and multi-spectral imaging capabilities of the Sentinel-1 and Sentinel-2 constellations deliver data with a very high frequency and high spatial and spectral resolution. This study explores the synergetic use of these sensors to monitor and predict grassland quality and quantity over the Grafenwoehr military training area (GTA) in Bavaria, Germany (49°40’ 56’’ N, 11° 47’ 20’’ E; approx. 230 km²) between 2015 and 2017. About 85% of the GTA is part of the Natura 2000 network and contains numerous rare and protected habitat types. Semi-natural grasslands cover roughly 25% of the GTA. The whole area is grazed by abundant red deer.

The study investigates the regional mapping of grassland crude protein (CP) and organic acid detergent fiber (oADF) concentration, to indicate forage quality. In addition, compressed sward height (CSH) and standing biomass (DM) of grassland were mapped to determine grassland forage quantity. To map oADF, CP, CSH, and DM, vegetation samples were collected and CSH (range = 2.9–25.5 cm) was measured at 15 locations at eight dates during the growing season of each year. For this, permanent sampling plots of 15 × 15 m were established. Nitrogen concentrations were measured according to the Dumas combustion method in a CN elemental analyser and converted to CP (range = 5.9%–21.3%). oADF concentrations were determined using near-infrared spectroscopy (range = 23.4%–48.0%). DM was estimated based on a generalised least squares model using CSH measurements and calibration cuts (range = 117.6–630.1 g/m²). Sentinel-1 (ascending and descending) and Sentinel-2 data were acquired as close as possible to the respective field sampling date (max difference to mean sampling date = 5). Sentinel-1 data were processed to a ground resolution of 10 m using SNAP and included the application of the respective orbit file, geometric calibration, terrain correction, speckle filtering (Lee filter), and conversion to dB using a range-doppler approach. Sentinel-2 level-1C top-of-atmosphere reflectance values were converted to surface reflectance using Sen2Cor. All 20 m Sentinel-2 bands were resampled to 10 m using the super-resolution plugin in SNAP. In addition, 77 vegetation indices (e.g. NDVI) and five biophysical parameters (e.g. LAI) were derived from the Sentinel-2 data. Several Synthetic Aperture Radar (SAR) indices, such as simple ratios were also derived. In total, 102 potential predictor variables were included to map semi-natural grassland quality and quantity. The Random Forest regression algorithm within the statistical programming environment R was used. The model performance was estimated by the absolute mean square error (RMSE) and coefficient of determination (R²), using a 10-fold cross-validation approach. Starting from a model with all predictor variables, the least important variable was iteratively removed, determined by a permutation-based change in RMSE by excluding one variable and keeping the rest in the model.

The results demonstrate the added benefit of combined Sentinel-1 and Sentinel-2 data for prediction of both forage quality and quantity. The maximum R² for all tested models using Sentinel-2 data only was 0.78 and for the combined Sentinel-1 and Sentinel-2 data set 0.79. The best results for the prediction of oADF (R² = 0.79, sd = 0.12; RMSE = 2.31%, sd = 0.43%) and CP (R² = 0.72, sd = 0.14; RMSE = 1.7%, sd = 0.34%) were observed applying 14 and 18 predictor variables, respectively. For CSH (R² = 0.59, sd = 0.21; RMSE = 2.77 cm, sd = 0.63 cm) and DM (R² = 0.44, sd = 0.24; RMSE = 91.50 g/m², sd = 17.42 g/m²) the Random Forest regression algorithm performed best using nine predictor variables in both cases. Only a few studies concerned with the prediction of vegetation forage quality and quantity using Sentinel-1 and Sentinel-2 exist so far. This study demonstrated that there is a major benefit in utilising the temporal potential of these high-resolution sensors. Remote sensing-based regression techniques have a great potential for mapping grassland forage quality and quantity to support the conservation management and monitoring of grassland areas.

Area A - Board 24

Grasslands vary cosiderably across the globe in their richness, biodiversity and the systems and services they support and they are widely influenced by human management. The heterogeneity and complexity of grassland management systems and their short-term variation make the mapping of grassland use intensity extremely challenging.

Here we present an assessment of grassland management type and intensity (namely mowing and grazing) using a complex hierarchical rule set applied over a NDVI time series derived from high temporal and spatial resolution satellite imagery. We applied the method in the lowlands of Switzerland across the Canton of Aargau, a largely agricultural area featuring diverse land use management intensities. We address the following aims:

- identification of grassland within the landscape

- development of a framework for mapping grassland management activities using phenological curves

- assessment of grassland management throughout the region using a combination of different satellite sensors

The base data source for our assessment of phenological time series was Planetscope CubeSat imagery (with 3m and ~1day, spatial and temporal resolution, respectively). In the first instance, the Planetscope imagery was processed to the 2A level by applying the 6SV model (GRASS GIS) and then cloud masking was performed by thresholding of the blue band. The pre-processed cloud-free imagery was then used to delineate the agricultural parcels (homogenous fields with respect to RGBN bands) by using object-oriented image analysis (OBIA). Non-agricultural areas were then masked out.

For every agricultural pixel in the study area, we obtained more than 50 cloud-free NDVI values between the period 01/2017-12/2017. The NDVI time series developed (using the median of a segmented parcel), however, is affected by noise caused by properties of the Planetscope CubeSat with imagery sourced from a constellation of numerous satellites. To limit noise effects and increase the density of the TS we used machine learning techniques to fuse the Planet NDVI TS with additional NDVI TS derived on the same geographical locations using the Sentinel and Landsat satellites (both processed to 2A level). The fusion model was developed on image pairs of overlapping days of Planet and Sentinel (or Landsat), and then applied to all Planet, Sentinel and Landsat scenes. Thus, in addition to the original Planetscope TS, the resulting TS covered also cloud-free pixels of overpassing Sentinel 2 and Landsat satellites.

Using this fusion methodology, the TS of median NDVI values of a unique segment of agricultural area provided robust results and were not very sensitive to mixed pixels resulting from the different resolutions of the satellites used.

The methodology developed allowed us to effectively identify abrupt temporal changes of NDVI which indicated mowing activities. Temporal densification of the time series by fusing different satellite imageries also decreased the time lag effect, where gaps in a time series or cloud cover cause a lag between mowing event occurrence and detection. Time lags effect are still likely to be present when applying the proposed methodology on less dense TS as derived from individual satellites. Compared to the detection of mowing, the identification of the exact length of grazing on grassland was less accurate (mostly omission errors). Grazing is a continuous process which is considerably more dependent on the type of climate and livestock.

We believe that the method developed here for semi-automatic assessment of signals from remote sensing instruments present highly accurate results for identifying grassland management in areas without any a priori knowledge of the local management and the grassland condition.

Area A - Board 25

Growing global demand for agricultural products results in agricultural expansion at the expense of environmental frontiers, and such demand is likely to increase in future. At the same time, agricultural land abandonment is widespread globally and affects biodiversity, global biogeochemical cycles and societal well-being. There has been a strong focus on monitoring cropland or agricultural land abandonment-combined abandoned croplands and grasslands used for hay cutting and livestock grazing. Unfortunately, though, monitoring grassland abandonment alone is challenging due to the subtle change of electromagnetic signal/reflectance once managed grasslands become abandoned. Former Soviet Union is a prime example of widespread cropland and managed grassland abandonment, which is tightly linked with precipitous livestock decline after the breakup of the Soviet Union. Our ultimate goal was to assess the suitability of commonly used machine-learning classifiers, such Support Vector Machines, Random Forest, Naïve Bayes, and convolutional neural networks (CNN) to classify agricultural land-use change from 1990 to 2018 and more explicitly mowing patterns for circa 2018 with Landsat and Sentinel-2 multi-seasonal imagery. Our second goal was to test how well selected classifiers hold misclassified training pixels, with up to 30% artificially injected errors in labeling (a replication of potential issues related to utilization of LUCAS as training data). Three Landsat footprints were selected in different environmental and socioeconomic settings, namely, 1) Lithuania/ Belarus cross-border, a temperate region (WRS-2 path 186 row 22), 2) Ryazan province in Russia with boreal to forest-steppe transition (path 176/ row 22) and 3) Orenburg province in Russia with forest-steppe to dry-steppe transition (path 165 row 24). The results showed the diverging land-use change trajectories across three study sites 1) with little cropland and grassland abandonment in Lithuania/ Belarus cross-border from 1990 to 2000, and ongoing recultivation by 2018; 2) with massive abandonment in Ryazan and little recultivation by 2018; 3) little cropland and negligible grassland abandonment in Orenburg and little recultivation by 2018. CNN was the most accurate classifier for agricultural land-use change detection across three study sites (OA above 90%). At the same time, Support Vector Machines and Random Forest were very close to CNN regarding the classification accuracy. CNN and SVM hold well if some training pixels were artificially labeled with wrong classes, followed by RF, but not Naïve Bayes classifier. User’s and Producer’s accuracies were below 80% when we employed CNN, SVM and RF and major misclassification was with intensively grazed areas in the vicinity to settlements (Orenburg). Our results demonstrate the suitability of Landsat| and Sentinel-2 imagery to monitor agricultural land-use change and agricultural land abandonment, while more efforts and better training data are needed to accurately classify mowing and grazing patterns.

Area A - Board 26