Source-linked AI summary
From parcel to continental scale -- A first European crop type map based on Sentinel-1 and LUCAS Copernicus in-situ observations
Raphaël d'Andrimont, Astrid Verhegghen, Guido Lemoine, Pieter Kempeneers, Michele Meroni, Marijn van der Velde
TL;DR
EU-wide crop information is needed to evaluate agricultural policies, while Sentinel-1’s continental-scale potential had not been fully explored. The study develops and rigorously assesses a 10-m continental crop-type mapping framework using Sentinel-1 and extensive in-situ data, finding consistent mapping across landscapes and F-scores above 0.9 for several countries.
Problem
EU-wide crop information is essential for evaluating agricultural policies and supporting timely crop-yield and production decisions.
Method
The study combines Sentinel-1 observations with LUCAS 2018 in-situ data and assesses the resulting crop map using area-weighted accuracy metrics and comparisons with farmer-declared parcels.
Results
The resulting map is consistent across different landscapes for major land-cover classes, while F-scores exceed 0.9 in Austria, Czech Republic, Germany, Lithuania, Luxembourg, Latvia, Poland, and Slovakia.
Takeaways & Limitations
The framework opens avenues for consistent agricultural monitoring at fine spatial and temporal scales over large areas, serving scientific modellers and policy makers.
Takeaways & Limitations
National or subnational area statistics were not estimated, and direct pixel counting would not account for bias from mixed and misclassified pixels.
Abstract
from arXiv · showhide
Detailed parcel-level crop type mapping for the whole European Union (EU) is necessary for the evaluation of agricultural policies. The Copernicus program, and Sentinel-1 (S1) in particular, offers the opportunity to monitor agricultural land at a continental scale and in a timely manner. However, so far the potential of S1 has not been explored at such a scale. Capitalizing on the unique LUCAS 2018 Copernicus in-situ survey, we present the first continental crop type map at 10-m spatial resolution for the EU based on S1A and S1B Synthetic Aperture Radar observations for the year 2018. Random forest classification algorithms are tuned to detect 19 different crop types. We assess the accuracy of this EU crop map with three approaches. First, the accuracy is assessed with independent LUCAS core in-situ observations over the continent. Second, an accuracy assessment is done specifically for main crop types from farmers declarations from 6 EU member countries or regions totaling >3M parcels and 8.21 Mha. Finally, the crop areas derived by classification are compared to the subnational (NUTS 2) area statistics reported by Eurostat. The overall accuracy for the map is reported as 80.3% when grouping main crop classes and 76% when considering all 19 crop type classes separately. Highest accuracies are obtained for rape and turnip rape with user and produced accuracies higher than 96%. The correlation between the remotely sensed estimated and Eurostat reported crop area ranges from 0.93 (potatoes) to 0.99 (rape and turnip rape). Finally, we discuss how the framework presented here can underpin the operational delivery of in-season high-resolution based crop mapping.
1. Introduction
The study addresses the need for detailed, timely EU-wide crop information to evaluate agricultural policies and support crop production decisions. It combines Sentinel-1 observations with LUCAS Copernicus in-situ data to establish a 10-m continental crop-mapping benchmark.
- Motivation: EU-wide crop-type information is essential for evaluating agricultural policies and supporting independent, timely crop-yield and production decisions.The motivation is linked to the scale of EU agriculture and the importance of the Common Agricultural Policy.
- Sentinel-1 opportunity: Sentinel-1 provides frequent, all-weather C-band SAR observations that the study exploits for continental agricultural monitoring at 10-m resolution.Its microwave observations are less affected by atmospheric conditions than optical imagery.
2. Materials and Methods
The study covers the EU-28 as defined in 2018 and organizes the analysis through a sequence of data preparation, legend definition, feature and parameter selection, continental classification, and accuracy assessment.
- Study area: The study area is the EU-28 in 2018, covering 4,469,169 km2, with a 10-km buffer added around its borders.The buffer was applied to avoid mapping issues at the EU-28 boundary.
- Workflow: The workflow progresses from Sentinel-1 and in-situ inputs through legend definition, feature and parameter selection, continental classification, and three accuracy assessments.The accuracy assessments are presented in Section 2.6.
2.1. Sentinel-1 data
Sentinel-1 data are processed into spatially and temporally consistent VV, VH, and cross-polarization-ratio features for continental crop classification. The processing uses geocoded, calibrated 10-m backscatter and 10-day composites across 2018.
- Sensor data: Sentinel-1 IW acquisitions over Europe provide VV and VH polarized backscatter from a 5.404 GHz C-band signal with a 5.55 cm wavelength.The constellation acquires data in ascending and descending orbits without relying on solar illumination.
- Preprocessing: Sentinel-1 GRD data are processed in Google Earth Engine to remove thermal noise and generate geocoded, radiometrically calibrated σ0 at 10-m pixel spacing.The workflow does not apply terrain flattening and instead uses σ0, while masking steep slopes and relying on the predominantly flat distribution of crops.
- Temporal compositing: The workflow masks scene edges, averages ascending and descending acquisitions over successive 10-day periods, converts averaged σ0 to decibels, and computes the VH/VV cross-polarization ratio.The cross-polarization ratio is computed for each scene and averaged to the same 10-day periods.
- Feature archive: The resulting archive contains 36 regularly timed 10-day composites of VV, VH, and CR for 2018-01-01 through 2018-12-31.The feature set is independent of the number of actual acquisitions within each 10-day period.
2.2. LUCAS 2018 in-situ data
The LUCAS 2018 survey provides the in-situ information used to train the classification models and assess the crop map. Its observations include core-point variables and an EO-tailored Copernicus polygon dataset.
- Data use: The LUCAS 2018 data are used for both training the classification models and assessing crop-map accuracy.The section explicitly describes the survey data as serving both purposes.
- LUCAS core survey: LUCAS 2018 collected 97 variables at 337,854 core points across the survey.Most surveyed points fall within homogeneous areas with a minimum mapping unit of about 7 m2.
- LUCAS Copernicus module: The Copernicus module collected land-cover extent up to 51 meters in four cardinal directions around observation points, producing EO-compatible in-situ data.The level-2 dataset contains 63,287 polygons, while filtering for level-3 legends leaves 58,426 polygons with detailed land-cover and land-use information.
2.3. Legend definition
The study reorganizes the LUCAS legend into three broad vegetation classes and further divides arable land into 19 crop-type or crop-group classes.
- The study groups LUCAS vegetation classes into arable land, woodlands and shrubland, and grasslands.
- Arable land is subdivided into 19 specific crop types or crop groups for EU-28 mapping.
- Temporary grasslands are assigned to grasslands, while permanent crops are assigned to woody vegetation.
- Woodlands and shrublands are regrouped into one class, while cropland is separated from grasslands and permanent crops.
2.4. Crop type classification
Crop classification uses LUCAS Copernicus polygons, broad biome stratification, hierarchical supervised models, and Sentinel-1 time-series features selected for seasonal accuracy.
- 58.178 of 63.287 available LUCAS Copernicus polygons are selected to train the classifiers.
- The EU-28 is divided into northern continental and southern Mediterranean strata to account for climatic and ecological gradients.
- The authors choose broad stratification as a trade-off between stratification detail and the sample size available for resulting strata.
- A two-phase procedure first classifies five broad land-cover classes, then separates arable land into 19 crop types or groups.
- Random forest models trained on Sentinel-1 time series map both hierarchical levels in each stratum.
- Feature selection evaluates 10-day averaged Sentinel-1 indices, seasonal periods from January to December, overall accuracy, and crop-type F-scores.
2.5. Classification at scale
The selected Sentinel-1 features support a continental EU-28 crop map, produced from four stratified hierarchical random-forest models and post-processed for distribution.
- The final EU crop map uses Sentinel-1 information from January through the end of July.
- Four supervised models are trained by combining two strata with two hierarchical classification levels.
- Randomized hyperparameter search evaluates seven parameters with 3-fold cross-validation across 100 combinations, totaling 300 fits.
- The classification is performed at continental scale on the JRC Big Data Analytics Platform using an HTCondor environment.
- Areas above 1000 m, slopes exceeding 10°, and poorly represented non-arable classes are masked during distribution processing.
- The map is re-projected to the ETRS89-LAEA Lambert azimuthal equal-area projection for analysis.
2.6. Accuracy assessment
Accuracy is assessed against independent LUCAS observations, farmer-declared parcels, and official subnational crop-area statistics using complementary classification and area-agreement measures.
- Three validation approaches use independent LUCAS core points, GSAA farmer declarations, and Eurostat subnational statistics.
- LUCAS-based assessment reports confusion matrices, overall accuracy, user accuracy, and producer accuracy at a 95% confidence level.
- GSAA comparisons cover six EU regions and select crop classes representing at least 1% of each region’s cumulative GSAA area.
- For selected GSAA parcels, the majority predicted pixel class is compared with the declared crop label to calculate producer and user accuracy.
- Eurostat comparisons use NUTS 2 area statistics where available, with NUTS 1 data for the UK and Germany, and calculate Pearson correlations.
3. Results
The EU crop map uses Sentinel-1 VV and VH time series from January through July and achieves strong accuracy across broad land-cover and crop classes. Performance varies by crop, stratum, and administrative region, with the strongest results for rape and turnip rape and remaining confusion among similar classes.
- 3.1. Features selection: 79.89% overall accuracy was obtained by combining VV and VH backscatter, which were selected for the final EU crop map.The training series covered 22 ten-day periods from 1 January through 31 July.
- 3.2. Visual assessment: The map covers 91 Mha of cropland at 10-m resolution and represents major crop types and vegetation classes across the EU.It contains 9,174 million 10-m pixels and is available for download and visualization.
- 3.2. Visual assessment: The mapped areas are consistent across landscapes at broad class level, while parcel sizes and diverse agricultural patterns are visibly represented.Common wheat, maize, and barley are the three most detected crop types by area.
- 3.2.2. F-score per crop type and administrative unit: Rape and turnip rape achieve F-scores above 0.9 in several countries, while crop-specific performance differs between northern and southern strata.Specific-crop overall accuracy is 78% in the northern stratum and 70.8% in the southern stratum.
- 3.3.1. LUCAS core points: The LUCAS assessment reports 76.1% overall accuracy for 19 crop classes and 80.3% when crops are grouped into main crop-type classes.Woodland and shrubland are well discriminated, whereas grasslands show confusion with that class.
- 3.3.2. GSAA: The GSAA comparison shows strong detection of common wheat, barley, maize, sunflower, and rape and turnip rape across analysed regions.Rape and turnip rape reach producer accuracy above 98% and user accuracy above 97% in the three regions where present.
4. Discussion
The discussion links mapping timeliness and accuracy to crop phenology and geography, while identifying validation, data-access, and crop-coverage constraints. It also outlines complementary data sources and operational trade-offs.
- Timeliness and accuracy: 76% overall accuracy was achieved across all crops using Sentinel-1 observations from January through July.Accuracy and parcel discrimination continued improving as the season progressed.
- Timeliness and accuracy: F-score plateaus occurred earlier in the Mediterranean stratum than in the northern stratum, consistent with earlier southern growing seasons.In the north, several crops reached plateaus from late June through September, whereas southern crops generally plateaued by June through August.
- Accuracy assessment: The LUCAS point assessment covers the continent but uses a filtered subset whose protocol was not designed for Earth Observation applications.Small-area heterogeneity can complicate reconciliation with crop-type classes, especially where few acquisitions contribute to averages.
- Accuracy assessment: Official-statistics comparisons cover the study area but evaluate classification only at aggregate administrative spatial levels.The GSAA assessment was limited to specific northern areas because openly accessible 2018 datasets were unavailable elsewhere.
- Accuracy assessment: Common wheat and maize showed regional overestimation, while barley was overestimated in the south and underestimated in the north.These patterns corresponded in part with crop-specific commission and omission errors across LUCAS, GSAA, and official-statistics assessments.
- Operational service: The framework’s operational utility is constrained by the lack of a full, free, open Sentinel-1 Level-2 application-ready archive and by limited GSAA accessibility.GSAA can provide 90–95% accuracy for more crops when available in July, whereas many member-state datasets arrive later or are not openly accessible.
5. Conclusions
The study designs, implements, and rigorously assesses a continental 10-m crop-type mapping framework from Sentinel-1. It opens avenues for consistent fine-scale agricultural monitoring over large areas.
- Conclusions: A 10-m Sentinel-1 crop-type mapping framework was designed, implemented, and rigorously assessed at continental scale.The framework is intended for arable-land monitoring relevant to environmental, agricultural, and climate-policy implementation.
- Conclusions: The framework opens avenues for a robust information system that monitors agriculture consistently at fine spatial and temporal scales over large areas.
6. Author contributions
The authors jointly conceptualized, developed, processed, analyzed, and documented the study.
- Author contributions: R.D. and A.V. conceptualized the study and designed its methodology.
- Author contributions: R.D., A.V., G.L., and P.K. processed the data, while all listed authors analyzed the data and wrote the paper.
Data dissemination
The paper provides access to the EU crop map and documents supplementary figures and tables covering training data, validation, accuracy, area estimates, and classification features.
- Map dissemination: The EU crop map, masked for non-vegetation classes and steep slopes, is available for download and visualization.
- Map dissemination: The map is downloadable as a 6.2-Gb GeoTIFF and accessible through a WMS service.
- Supplementary data: The training data contain 58,178 LUCAS Copernicus polygons, 2,956,889 Sentinel-1 10-m pixels, 21 thematic classes, and two geographical strata.
- Supplementary figures: Additional supplementary figures show temporal F-scores, declared crop areas, validation points, Eurostat differences, and Sentinel-1 backscatter groupings.
- Supplementary tables: Supplementary materials include weighted area accuracy, RF hyperparameters, crop-class areas, confusion matrices, GSAA matching data, and Eurostat legend convergence.