Search

Field boundaries are at the core of many agricultural applications and are a key enabler for the operational monitoring of agricultural production to support food security. Recent scientific progress in deep learning methods has highlighted the capacity to extract field boundaries from satellite and aerial images with a clear improvement from object-based image analysis (e.g. multiresolution segmentation) or conventional filters (e.g. Sobel filters). However, these methods need labels to be trained on. So far, no standard data set exists to easily and robustly benchmark models and progress the state of the art. The absence of such benchmark data further impedes proper comparison against existing methods. Besides, there is no consensus on which evaluation metrics should be reported (both at the pixel and field levels). As a result, it is currently impossible to compare and benchmark new and existing methods. To fill these gaps, we introduce AI4Boundaries, a data set of images and labels readily usable to train and compare models on field boundary detection. AI4Boundaries includes two specific data sets: (i) a 10 m Sentinel-2 monthly composites for large-scale analyses in retrospect and (ii) a 1 m orthophoto data set for regional-scale analyses, such as the automatic extraction of Geospatial Aid Application (GSAA). All labels have been sourced from GSAA data that have been made openly available (Austria, Catalonia, France, Luxembourg, the Netherlands, Slovenia, and Sweden) for 2019, representing 14.8 M parcels covering 376 K km2. Data were selected following a stratified random sampling drawn based on two landscape fragmentation metrics, the perimeter/area ratio and the area covered by parcels, thus considering the diversity of the agricultural landscapes. The resulting “AI4Boundaries” dataset consists of 7831 samples of 256 by 256 pixels for the 10 m Sentinel-2 dataset and of 512 by 512 pixels for the 1 m aerial orthophoto. Both datasets are provided with the corresponding vector ground-truth parcel delineation (2.5 M parcels covering 47 105 km2), and with a raster version already pre-processed and ready to use. Besides providing this open dataset to foster computer vision developments of parcel delineation methods, we discuss the perspectives and limitations of the dataset for various types of applications in the agriculture domain and consider possible further improvements. The data are available on the JRC Open Data Catalogue: http://data.europa.eu/89h/0e79ce5d-e4c8-4721-8773-59a4acf2c9c9 (European Commission, Joint Research Centre, 2022).

Crop yield forecasting at national level relies on predictors aggregated from smaller spatial units to larger ones according to harvested crop areas. Such crop areas come from land cover maps or reported statistics, both of which can have errors and uncertainties. Sub-national or regional crop yield forecasting minimizes the propagation of these errors to some extent. In addition, regional forecasts provide added value and insights to stakeholders on regional differences within a country, which would otherwise compensate each other at national level. We propose a crop yield forecasting approach for multiple spatial levels based on regional crop yield forecasts from machine learning. Machine learning, with its data-driven approach, can leverage larger data sizes and capture nonlinear relationships between predictors and yield at regional level. We designed a generic machine learning workflow to demonstrate the benefits of regional crop yield forecasting in Europe. To evaluate the quality and usefulness of regional forecasts, we predicted crop yields for 35 case studies, including nine countries that are major producers of six crops (soft wheat, spring barley, sunflower, grain maize, sugar beets and potatoes). Machine learning models at regional level had lower normalized root mean squared errors (NRMSE) and uncertainty than a linear trend model, with Wilcoxon p-values of 3e-7 and 2e-7 for 60 days before harvest and end of season respectively. Similarly, regional machine learning forecasts aggregated to national level had lower NRMSEs than forecasts from an operational system in 18 out of 35 cases 60 days before harvest, with a Wilcoxon p-value of 0.95 indicating similar performance. Our models have room for improvement, especially during extreme years. Nevertheless, regional crop yield forecasts from machine learning and aggregated national forecasts provide a consistent forecasting method across spatial levels and insights from regional differences to support important policy decisions.

Accurate, automated cloud and cloud shadow detection is a key component of the processing needed to prepare optical satellite imagery for scientific analysis. Many existing cloud detection algorithms rely on temperature information to identify clouds, making detection difficult for imagers that lack a thermal band, like Sentinel-2. To get maximum benefit from Sentinel-2 products it is critical to understand which algorithms best identify clouds and their shadows in images. We examined the relative performance of five different cloud-masking algorithms (Sen2Cor, MAJA, LaSRC, Fmask and Tmask) in 6 Sentinel-2 scenes (28 total images) distributed across the Eastern Hemisphere. Expanding on these comparisons, we tested ensemble approaches to improve results. We tested three ensemble approaches to cloud and shadow classification based on the outputs of the five initial algorithms using the cloud masks in: (1) a majority prediction model; (2) a random forests model; and (3) a conditional logic model. Accuracy assessments show a trade-off between omission and commission errors in cloud detection for individual algorithms across all sites, and some algorithms are better at detecting either clouds or cloud shadows. No single algorithm outperforms the others for both clouds and shadows. Aggregating the results from multiple algorithms produces fewer undetected clouds and higher overall accuracy than any single algorithm, with as high as 2.7% improvement over the top-performing algorithm, suggesting an ensemble approach may be the most useful for processing of Sentinel-2 data.

A PRELIMINARY INVESTIGATION ON COMPARISON AND TRANSFORMATION OF SENTINEL-2 MSI AND LANDSAT 8 OLI Timely and accurate earth observation with short revisit interval is usually necessary, especially for emergency response. Currently, several new generation sensors provided with similar channel characteristics have been operated onboard different satellite platforms, including Sentinel-2 and Landsat 8. Joint use of the observations by different sensors offers an opportunity to meet the demands for emergency requirements. For example, through the combination of Landsat and Sentinel-2 data, the land can be observed every 2–3 days at medium spatial resolution. However, differences are expected in radiometric values (e.g., channel reflectance) of the corresponding channels between two sensors. Spectral response function (SRF) is taken as an important aspect of sensor settings. Accordingly, between-sensor differences due to SRFs variation need to be quantified and compensated. The comparison of SRFs shows difference (more or less) in channel settings between Sentinel-2 Multi-Spectral Instrument (MSI) and Landsat 8 Operational Land Imager (OLI). Effect of the difference in SRF on corresponding values between MSI and OLI was investigated, mainly in terms of channel reflectance and several derived spectral indices. Spectra samples from ASTER Spectral Library Version 2.0 and Hyperion data archives were used in obtaining channel reflectance simulation of MSI and OLI. Preliminary results show that MSI and OLI are well comparable in several channels with small relative discrepancy (

The Harmonized Landsat and Sentinel-2 (HLS) project is a NASA initiative aiming to produce a Virtual Constellation (VC) of surface reflectance (SR) data acquired by the Operational Land Imager (OLI) and Multi-Spectral Instrument (MSI) aboard Landsat 8 and Sentinel-2 remote sensing satellites, respectively. The HLS products are based on a set of algorithms to obtain seamless products from both sensors (OLI and MSI): atmospheric correction, cloud and cloud-shadow masking, spatial co-registration and common gridding, bidirectional reflectance distribution function normalization and spectral bandpass adjustment. Three products are derived from the HLS processing chain: (i) S10: full resolution MSI SR at 10 m, 20 m and 60 m spatial resolutions; (ii) S30: a 30 m MSI Nadir BRDF (Bidirectional Reflectance Distribution Function)-Adjusted Reflectance (NBAR); (iii) L30: a 30 m OLI NBAR. All three products are processed for every Level-1 input products from Landsat 8/OLI (L1T) and Sentinel-2/MSI (L1C). As of version 1.3, the HLS data set covers 10.35 million km2 and spans from first Landsat 8 data (2013); Sentinel-2 data spans from October 2015. The L30 and S30 show a good consistency with coarse spatial resolution products, in particular MODIS Collection 6 MCD09CMG products (overall deviations do not exceed 11%) that are used as a reference for quality assurance. The spatial co-registration of the HLS is improved compared to original Landsat 8 L1T and Sentinel-2A L1C products, for which misregistration issues between multi-temporal data are known. In particular, the resulting computed circular errors at 90% for the HLS product are 6.2 m and 18.8 m, for S10 and L30 products, respectively. The main known issue of the current data set remains the Sentinel-2 cloud mask with many cloud detection omissions. The cross-comparison with MODIS was used to flag products with most evident non-detected clouds. A time series outlier filtering approach is suggested to detect remaining clouds. Finally, several time series are presented to highlight the high potential of the HLS data set for crop monitoring.

High-resolution sensor Surface Reflectance (SR) data are affected by surface anisotropy but are difficult to adjust because of the low temporal frequency of the acquisitions and the low angular sampling. This paper evaluates five high spatial resolution Bidirectional Reflectance Distribution Function (BRDF) adjustment techniques. The evaluation is based on the noise level of the SR Time Series (TS) corrected to a normalized geometry (nadir view, 45° sun zenith angle) extracted from the multi-angular acquisitions of SPOT4 over three study areas (one in Arizona, two in France) during the five-month SPOT4 (Take5) experiment. Two uniform techniques (Cst, for Constant, and Av, for Average), relying on the Vermote–Justice–Bréon (VJB) BRDF method, assume no variation in space of the BRDF shape. Two methods (VI-dis, for NDVI-based disaggregation and LC-dis, for Land-Cover based disaggregation) are based on disaggregation of the MODIS-derived BRDF VJB parameters using vegetation index and land cover, respectively. The last technique (LUM, for Look-Up Map) relies on the MCD43 MODIS BRDF products and a crop type data layer. The VI-dis technique produced the lowest level of noise corresponding to the most effective adjustment: reduction from directional to normalized SR TS noises by 40% and 50% on average, for red and near-infrared bands, respectively. The uniform techniques displayed very good results, suggesting that a simple and uniform BRDF-shape assumption is good enough to adjust the BRDF in such geometric configuration (the view zenith angle varies from nadir to 25°). The most complex techniques relying on land cover (LC-dis and LUM) displayed contrasting results depending on the land cover.

This paper presents the SPOT-4 (Take 5) experiment, aimed at providing time series of optical images simulating the repetitivity, the resolution and the large swath of Sentinel-2 images. The aim was to help users set up and test their applications and methods, before Sentinel-2 mission data become available. In 2016, when both Sentinel-2 satellites are operational, and for at least fifteen years, users will have access to high resolution time series of images systematically acquired every five days, over the whole Earth land surfaces. Thanks to Sentinel-2’s high revisit frequency, a given surface should be observed without clouds at least once a month, except in the most cloudy periods and regions. In 2013, the Centre National d’Etudes Spatiales (CNES) lowered the orbit altitude of SPOT-4, to place it on a five-day repeat cycle orbit for a duration of five months. This experiment started on 31 January 2013 and lasted until 19 June 2013. SPOT-4 images were acquired every fifth day, over 45 sites scattered in nearly all continents and covering very diverse biomes for various applications. Two ortho-rectified products were delivered for each acquired image that was not fully cloudy, expressed either as top of atmosphere reflectance (Level 1C) or as surface reflectance (Level 2A). An extensive validation campaign was held to check the performances of these products with regard to the multi-temporal registration, the quality of cloud masks, the accuracy of aerosol optical thickness estimates and the quality of surface reflectances. Despite high a priori geo-location errors, it was possible to register the images with an accuracy better than 0.5 pixels in the large majority of cases. Despite the lack of a blue band on the SPOT-4 satellite, the cloud and shadow detection yielded good results, while the aerosol optical thickness was measured with a root mean square error better than 0.06. The surface reflectances after atmospheric correction were compared with in situ data and other satellite data showing little bias and the standard deviation of surface reflectance errors in the range (0.01–0.02). The Take 5 experiment is being repeated in 2015 with the SPOT-5 satellite with an enhanced resolution.

The Advanced Very High Resolution Radiometer (AVHRR) sensor provides a unique global remote sensing dataset that ranges from the 1980s to the present. Over the years, several efforts have been made on the calibration of the different instruments to establish a consistent land surface reflectance time-series and to augment the AVHRR data record with data from other sensors, such as the Moderate Resolution Imaging Spectroradiometer (MODIS). In this paper, we present a summary of all the corrections applied to the AVHRR surface reflectance and NDVI Version 4 Product, developed in the framework of the National Oceanic and Atmospheric Administration (NOAA) Climate Data Record (CDR) program. These corrections result from assessment of the geolocation, improvement of cloud masking, and calibration monitoring. Additionally, we evaluate the performance of the surface reflectance over the AERONET sites by a cross-comparison with MODIS, which is an already validated product, and evaluation of a downstream leaf area index (LAI) product. We demonstrate the utility of this long time-series by estimating the winter wheat yield over the USA. The methods developed by Becker-Reshef et al. (2010) and Franch et al. (2015) are applied to both the MODIS and AVHRR data. Comparison of the results from both sensors during the MODIS-era shows the consistency of the dataset with similar errors of 10%. When applying the methods to AVHRR historical data from the 1980s, the results have errors equivalent to those derived from MODIS.

The surface reflectance, i.e., satellite derived top of atmosphere (TOA) reflectance corrected for the temporally, spatially and spectrally varying scattering and absorbing effects of atmospheric gases and aerosols, is needed to monitor the land surface reliably. For this reason, the surface reflectance, and not TOA reflectance, is used to generate the greater majority of global land products, for example, from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) sensors. Even if atmospheric effects are minimized by sensor design, atmospheric effects are still challenging to correct. In particular, the strong impact of aerosols in the Visible and Near Infrared spectral range can be difficult to correct, because they can be highly discrete in space and time (e.g., smoke plumes) and because of the complex scattering and absorbing properties of aerosols that vary spectrally and with aerosol size, shape, chemistry and density. This paper presents the Landsat 8 Operational Land Imager (OLI) atmospheric correction algorithm that has been developed using the Second Simulation of the Satellite Signal in the Solar Spectrum Vectorial (6SV) model, refined to take advantage of the narrow OLI spectral bands (compared to Thematic Mapper/Enhanced Thematic Mapper (TM/ETM+)), improved radiometric resolution and signal-to-noise. In addition, the algorithm uses the new OLI Coastal aerosol band (0.433-0.450μm), which is particularly helpful for retrieving aerosol properties, as it covers shorter wavelengths than the conventional Landsat, TM and ETM+ blue bands. A cloud and cloud shadow mask has also been developed using the "cirrus" band (1.360-1.390 μm) available on OLI, and the thermal infrared bands from the Thermal Infrared Sensor (TIRS) instrument. The performance of the surface reflectance product from OLI is analyzed over the Aerosol Robotic Network (AERONET) sites using accurate atmospheric correction (based on in situ measurements of the atmospheric properties), by comparison with the MODIS Bidirectional Reflectance Distribution Function (BRDF) adjusted surface reflectance product and by comparison of OLI derived broadband albedo from United States Surface Radiation Budget Network (US SURFRAD) measurements.

Sentinel-2 (S2) earth observation satellite mission, launched in 2015, is foreseen to promote within-field decisions in Precision Agriculture (PA) for both: (1) optimizing crop production; and (2) regulating environmental impacts. In this second scope, a set of Leaf Area Index (LAI) derived from S2 type time-series (2006-2010, using Formosat-2 satellite) is used to spatially constrain the within-field crop growth and the related nitrogen contamination of surface water simulated at a small experimental catchment scale with the distributed agro-hydrological model Topography Nitrogen Transfer and Transformation (TNT2). The Soil Water Holding Capacity (SWHC), represented by two parameters, soil depth and retention porosity, is used to fit the yearly maximum of LAI (LAX) at each pixel of the satellite image. Possible combinations of soil parameters, defining 154 realistic SWHC found on the study site are used to force spatially homogeneous SWHC. LAX simulated at the pixel level for the 154 SWHC, for each of the five years of the study period, are recorded and hereafter referred to as synthetic LAX. Optimal SWHCyear_I,pixel_j, corresponding to minimal difference between observed and synthetic LAX(year_I,pixel_j), is selected for each pixel, independent of the value at neighboring pixels. Each re-estimated soil maps are used to re-simulate LAX(year_I). Results show that simulated and synthetic LAX(year_I,allpixels) obtained from SWHCyear_I,allpixels are close and accurately fit the observed LAX(year_I,allpixels) (RMSE = 0.05 m(2)/m(2) to 0.2 and R-2 = 0.99 to 0.94), except for the year 2008 (RMSE = 0.8 m(2)/m(2) and R-2 = 0.8). These results show that optimal SWHC can be derived from remote sensing series for one year. Unique SWHC solutions for each pixel that limit the LAX error for the five years to less than 0.2 m(2)/m(2) are found for only 10% of the pixels. Selection of unique soil parameters using multi-year LAX and neighborhood solution is expected to deliver more robust soil parameters solutions and need to be assessed further. The use of optical remote sensing series is then a promising calibration step to represent crop growth within crop field at catchment level. Nevertheless, this study discusses the model and data improvements that are needed to get realistic spatial representation of agro-hydrological processes simulated within catchments.

The growing availability of high-resolution satellite image series offers new opportunities in agro-hydrological research and modeling. We investigated the possibilities offered for improving crop-growth dynamic simulation with the distributed agro-hydrological model: topography-based nitrogen transfer and transformation (TNT2). We used a leaf area index (LAI) map series derived from 105 Formosat-2 (F2) images covering the period 2006–2010. The TNT2 model (Beaujouan et al., 2002), calibrated against discharge and in-stream nitrate fluxes for the period 1985–2001, was tested on the 2005–2010 data set (climate, land use, agricultural practices, and discharge and nitrate fluxes at the outlet). Data from the first year (2005) were used to initialize the hydrological model. A priori agricultural practices obtained from an extensive field survey, such as seeding date, crop cultivar, and amount of fertilizer, were used as input variables. Continuous values of LAI as a function of cumulative daily temperature were obtained at the crop-field level by fitting a double logistic equation against discrete satellite-derived LAI. Model predictions of LAI dynamics using the a priori input parameters displayed temporal shifts from those observed LAI profiles that are irregularly distributed in space (between field crops) and time (between years). By resetting the seeding date at the crop-field level, we have developed an optimization method designed to efficiently minimize this temporal shift and better fit the crop growth against both the spatial observations and crop production. This optimization of simulated LAI has a negligible impact on water budgets at the catchment scale (1 mm yr−1 on average) but a noticeable impact on in-stream nitrogen fluxes (around 12%), which is of interest when considering nitrate stream contamination issues and the objectives of TNT2 modeling. This study demonstrates the potential contribution of the forthcoming high spatial and temporal resolution products from the Sentinel-2 satellite mission for improving agro-hydrological modeling by constraining the spatial representation of crop productivity.

International audience; Keywords: LAI GAI FAPAR Essential Climate Variables Green Area Index Radiative transfer model PROSAIL MODIS Formosat-2 Crop Validation Surface reflectance BRDF This study aims at validating Leaf Area Index (LAI) and Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) products derived from MODIS surface reflectance (MOD09CMG) at coarse resolution (0.05°) over crops. These Essential Climate Variables (ECVs) are estimated by using the inversion of the PROSAIL radiative transfer (BV-NNET tool) applied on MODIS BRDF (Bidirectional Reflectance Distribution Function) corrected surface reflectances and non-corrected. ECV estimates and the corresponding MCD15A3 Collection 5 and GEOLAND-2 (GEOv1) products are compared with ECV reference maps derived from BV-NNET applied on 105 high spatial resolution images (Formosat-2, 8 m) which were acquired from 2006 to 2010 in Southwest France. These latter are compared with local scale in situ measurements. The validation shows an uncertainty of 0.35 and 0.07 for LAI and FAPAR, respectively. The comparison shows that the ECV estimates from the three products properly capture the crops phenology in agreement with reference maps. Results indicate that MCD15A3 uncertainties (0.23 and 0.07 for LAI and FAPAR, respectively) are similar to previous intercomparison studies. GEOv1 shows a systemic positive bias for both LAI and FAPAR. The best agreement with the reference maps is found for MODIS BV-NNET products with r 2 higher than 0.9 and relative uncertainties lower than 17%. The use of BRDF-corrected surface reflectances as input of BV-NNET tool improves the uncertainty of LAI estimates (0.11, compared to 0.17 when directional surface reflec-tances are used as input) but not the uncertainty of FAPAR estimates. The deviation between FAPAR products which mostly affects low winter FAPAR, is related to the discrepancy of the soil directional assumption in PROSAIL model and BRDF correction method. The temporal stability of the daily MODIS BV-NNET products is better than the 4-day composite MCD15A3 products. Finally, BV-NNET tool applied at finer resolutions demonstrates that the increase of the resolution results in a decrease of the LAI and FAPAR uncertainties and a conservation of the biases.

The recent availability of high spatial and temporal resolution (HSTR) remote sensing data (Formosat-2, and future missions of Ven mu s and Sentinel-2) offers new opportunities for crop monitoring. In this context, we investigated the perspective offered by coupling a simple algorithm for yield estimate (SAFY) with the Formosat-2 data to estimate crop production over large areas. With a limited number of input parameters, the SAFY model enables the simulation of time series of green area index (GAI) and dry aboveground biomass (DAM). From 2006 to 2009, 95 Formosat-2 images (8 m, 1 day revisit) were acquired for a 24 x 24 km(2) area southwest of Toulouse, France. This study focused on two summer crops: irrigated maize (Zea mays) and sunflower (Helianthus annuus). Green area index (GAI) time series were deduced from Formosat-2 NDVI time series and were used to calibrate six major parameters of the SAFY model. Four of those parameters (partition-to-leaf and senescence function parameters) were calibrated per crop type based on the very dense 2006 Formosat-2 data set The retrieved values of these parameters were consistent with the in situ observations and a literature review. Two of the major parameters of the SAFY model (emergence day and effective light-use efficiency) were calibrated per field relative to crop management practices. The estimated effective light-use efficiency values highlighted the distinction between the C4 (maize) and 0 (sunflower) plants, and were linked to the reduction of the photosynthesis rate due to water stress. The model was able to reproduce a large set of GAI temporal shapes, which were related to various phenological behaviours and to crop type. The biomass was well estimated (relative error of 28%), especially considering that biomass measurements were not used for the calibration. The grain yields were also simulated using harvest index coefficients and were compared with grain yield statistics from the French Agricultural Statistics for the department of Haute-Garonne. The inter-annual variation in the simulated grain yields of sunflower was consistent with the reported variation. For maize, significant discrepancies were observed with the reported statistics.