Nationwide annual agricultural land-use maps of Germany from 1990 to 2023 derived from satellite imagery

Nationwide annual agricultural land-use maps of Germany from 1990 to 2023 derived from satellite imagery | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Data Note Nationwide annual agricultural land-use maps of Germany from 1990 to 2023 derived from satellite imagery Gideon Okpoti Tetteh, Vu-Dong Pham, Marcel Schwieder, Lukas Blickensdörfer, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9074257/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract We present nationwide annual agricultural land-use maps of Germany from 1990 to 2023, created from Landsat and Sentinel-2 images using a deep learning approach. Based on farmers’ parcel-level declarations from 2006 to 2022, we extracted annual training samples for 13 crop classes and one grassland class. These samples were used to train a multi-year one-dimensional convolutional neural network, which was subsequently applied to generate the annual land-use maps. Overall map accuracies ranged between 85% and 93%. Dominant classes such as grassland, rapeseed, winter cereals, sugar beet, and maize were detected with high accuracy (≥ 90%). Conversely, minor classes such as fallow land and plantation were predicted with low accuracy (≤ 52%). Comparison of map areas with agricultural statistics over the entire study period revealed high correlations for most classes, particularly maize ( r = 0.978). The presented maps provide an essential basis for analyzing long-term trends in agricultural land-use. They can be used to fill temporal gaps in national agricultural statistics and to disaggregate those statistics to higher spatial units. Artificial Intelligence and Machine Learning Geographic Information Systems Agricultural Economics & Policy Environmental Policy Agricultural lands Long-term monitoring Remote sensing Time series analysis Deep Learning Land-use classification Long-term dataset Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 5 Figure 6 1. Background & Summary In Germany, the land-use (LU) in agricultural landscapes has undergone significant changes over the past decades. The drivers behind those changes are multifaceted, reflecting a complex interplay of socio-economic, policy, and environmental factors 1 – 5 . These drivers have collectively shaped the agricultural landscape, influencing the extent, type, and intensity of agricultural LU. To design effective policies that foster sustainable agriculture, we need a comprehensive understanding of those drivers and their impact on agricultural LU. This requires long-term, spatially explicit information on agricultural LU. Primary data sources such as the Authoritative Topographic-Cartographic Information System (ATKIS), the Geo-spatial Application (GSA) of the Common Agricultural Policy (CAP) framework, and official agricultural statistics are available in Germany for monitoring agricultural lands. However, those data sources have certain limitations. In ATKIS, only broad land-cover (LC) classes, such as cropland and grassland, are mapped, without detailed information on agricultural LU types. While detailed information on agricultural LU types can be found in the GSA, the data has limited spatial and temporal coverage, and it is not publicly available for all federal states. Although the official agricultural statistics contain detailed information on agricultural LU and productivity, a comprehensive farm survey is conducted only every 3–4 years at sample farm locations. Additionally, the survey data are aggregated to the district level, thereby preventing spatially explicit analysis of agricultural LU changes. Further, the LU definitions used in the statistics change over time, thereby requiring considerable effort to consolidate the LU classes. Remote sensing data, with their high spatial and temporal coverage, are suitable for continuous and long-term area-wide mapping of agricultural lands 6 , 7 , thereby providing a unique means of overcoming the limitations of the abovementioned data sources. The literature is replete with the use of remote sensing data for generating agricultural LU maps at local, national, and continental scales 8 , 9 . The proliferation of agricultural LU maps has been driven by key developments, including the increasing volume of data from various satellite missions, advances in artificial intelligence methods for agricultural LU mapping, and greater access to all-in-one cloud computing platforms 7 , 9 , 10 . The launch of Sentinel-1 in 2014 and Sentinel-2 in 2015, as part of the Copernicus satellite program, has accelerated the creation of national-scale agricultural LU maps across Europe and beyond. This is particularly true for Germany, where numerous nationwide products exist 11 – 17 . However, detailed national-scale agricultural LU maps for years before the Copernicus era, dating back to 1990, are rarely available, making effective long-term monitoring difficult. The year 1990 is pivotal because it marks the baseline year for mandatory national greenhouse gas (GHG) inventory reporting by countries, including Germany, that are parties to the United Nations Framework Convention on Climate Change (UNFCCC) 18 . Therefore, high-resolution, area-wide, and long-term LU maps in Germany will not only aid in understanding the drivers behind agricultural LU changes but can also improve national GHG emissions reporting, because the maps can be used for Tier-3 (process-based) modelling 19 , 20 as defined in the Intergovernmental Panel on Climate Change (IPCC) guidelines. Other potential uses of such long-term maps include the spatial and temporal disaggregation of agricultural statistics 21 , 22 , assessing agricultural landscape heterogeneity within the context of biodiversity monitoring 23 , 24 , analyzing crop sequences and rotations 25 , 26 , evaluating agricultural policies 27 , 28 , and estimating crop supply response 29 . This study presents a historical dataset of agricultural LU maps for Germany, containing 13 main crops and one grassland class from 1990 to 2023, generated from the entire Landsat (1990–2023) and Sentinel-2 (2015–2023) archives using a deep learning (DL) approach. The combination of Landsat and Sentinel-2 from 2015 onwards yields a temporally dense time series 30 , 31 , which enables the differentiation of agricultural LU types based on their phenological profiles 32 , 33 . However, the time series before 2015 consists only of Landsat data and thus has lower spatial resolution and larger temporal gaps, making detailed mapping of agricultural LU more challenging. Further, LU mapping relies on reference data, which is needed for model training and validation. In more recent years, reference data such as the agricultural parcels declared by farmers in the GSA of CAP have become available in the European Union (EU) 34 , 35 . However, due to the limited spatial and temporal availability of the GSA data in Germany, long-term agricultural LU mapping has been hampered. For model training and validation in this study, we had access to GSA data covering different parts of Germany from 2006 to 2022. To overcome the lack of reference data before 2006 and the temporarily sparse satellite data before 2015, we adapted the DL approach of Pham et al. 36 , which was designed for LU mapping in data-sparse conditions. 2. Methods Fig. 1 shows the general workflow we used in this study to create the agricultural LU maps. The main components of the workflow are satellite data preprocessing, agricultural LU mapping, agricultural masks creation, postprocessing of agricultural LU maps, and technical validation. Each component is explained in subsequent sections. 2.1. Preprocessing of satellite data In this study, all Level-1 (top-of-atmosphere) Landsat images from 1990 to 2023 and Sentinel-2 images from 2015 to 2023 with cloud cover less than 75% were downloaded and preprocessed to generate analysis-ready data (ARD) of Germany using the Framework for Operational Radiometric Correction for Environmental Monitoring (FORCE) 37 . Generating the ARD images involved converting the top-of-atmosphere images to Level-2 (surface reflectance) images by correcting for radiometric, atmospheric, and geometric distortions 38 – 40 , masking of clouds and cloud shadows using the Fmask algorithm 41 – 43 , and resampling of images to 30 m. Following the approach of Pham et al. 36 , we extracted the bands present in both sensors (red, green, blue, near-infrared, shortwave-infrared 1, shortwave-infrared 2) and generated three spectral indices, namely Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Soil-Adjusted Vegetation Index (SAVI). Consequently, each ARD image contained nine bands (six spectral bands and three spectral indices). An overview table showing the number of clear-sky observations (CSOs) generated from the ARD images per year, alongside the corresponding sensors, can be found in Table S1 of the Supplementary File. 2.2. Agricultural land-use reference data As reference data for agricultural LU mapping, we used the agricultural parcels declared by the EU farmers in the GSA as part of the Integrated Administration and Control System (IACS) to receive CAP payments. The GSA datasets include the spatial boundaries and crop types of all parcels for which subsidies were requested. The federal states in Germany are responsible for reviewing all GSA applications and subsequently providing the data for research and development based on individual data use agreements. Consequently, data availability varies between the federal states. The spatial and temporal coverage of the GSA data we had access to in Germany is shown in Fig. 2 . Before 2006, GSA datasets were not available. The federal state of Brandenburg had the highest number of available GSA datasets from 2006 to 2022, while the federal state of Saarland had the lowest (2014–2016). No GSA data was available for Schleswig-Holstein and Hamburg. The GSA for Bremen and Berlin are included in the datasets of Lower Saxony and Brandenburg, respectively. In 2014, 2015, and 2016, the GSA had the highest spatial coverage (12 federal states). To ensure consistency in class definitions over time and with groups of classes used in the agricultural statistics of Germany, we extracted the main GSA classes and translated them into a condensed set of 14 classes: winter cereals, summer cereals, maize, grassland, potato, sugar beet, rapeseed, sunflower, legumes, horticultural crops, fallow land, vineyards, hops, and plantations. The three perennial classes (vineyards, hops, and plantations) were included to potentially identify the year of establishment. The translation table is in Table S2 of the Supplementary File. 2.3. Agricultural land-use mapping Given the non-availability of GSA before 2006 and the relatively low temporal density of ARD images before 2015, we opted for the mapping approach of Pham et al. 36 , which has been successfully used to map annual land cover and land use in the Baltic Sea region from 2000 to 2022 44 . This approach has three components: (1) the use of temporal encoding for generating annual time series data per pixel, (2) the application of two data augmentation techniques, namely random observation selection (ROS) and random day shifting (RDS), to simulate temporal patterns of data-sparse conditions and phenological variations, (3) the use of a 1D-CNN model for training and prediction. The temporal encoding involves generating arrays of 365 values per pixel and band, where each value represents the CSO for the day of year (DOY). Where no CSO is available for a particular DOY, zero is inserted into the array. The ROS involves randomly selecting proportions of the temporally encoded data for each reference sample during training, with the proportions ranging between 5% and 100%, to simulate data-sparse situations during model training. This is followed by RDS, which involves the random shifting of observations forward or backward within sixteen days, which corresponds to the temporal resolution of Landsat. This augmented time series, together with the class-wise reference labels, is finally passed to the 1D-CNN model for training. The 1D-CNN model comprises four convolution blocks. Each convolution block consists of two convolution layers, batch normalization, and rectified linear unit activation. Max pooling is applied after the first, second, and third convolution blocks, while the output of the fourth convolution block is flattened and attached to a dense layer for prediction. For more details regarding the approach, readers are referred to Pham et al. 36 . To train the model, we randomly sampled 10,000 pixels per class for each year with GSA data. Per year and class, the samples were split into 75% for training and 25% for validation. The respective training and validation samples per year were then merged to create a multi-year training and validation pool. The multi-year training samples were used to train the 1D-CNN model, while the multi-year validation samples were used to validate model performance during training. The model was trained for 100 epochs with a batch size of 512 and a learning rate of 0.001 on a computer with 48 GB of graphics processing unit (GPU) memory. After training, the model was applied to each ARD pixel per year to generate the initial agricultural LU maps of Germany from 1990 to 2023. 2.4. Agricultural masks creation The initial agricultural LU maps created in the previous section contain information for all pixel locations in Germany. To remove the non-agricultural pixels from those initial LU maps, we created annual agricultural masks. To create the annual agricultural masks, we first generated annual LC maps for Germany. As reference data for LC mapping, we used the German-wide reference database used for the calculation of GHG emissions in Germany by the Thünen Institute 45 . This reference database, which contains the main LC types of Germany, was created from a combination of LC datasets from the Coordination of Information on the Environment (CORINE) before 2000 and ATKIS thereafter. For our research, we had access to the LC samples in the reference database from 1990 to 2022. Per year, we randomly sampled 1,500 pixels from the following LC classes: forest, cropland, perennial cropland, grassland, woody plants and hedges, wetlands, water, peat mining, built-up areas, and others. Those ten classes cover the entire surface area of Germany. The LC samples were split into 75% training and 25% validation per year and class, and subsequently merged to create a multi-year training and validation pool. We followed the same training procedure used for LU mapping, and trained another 1D-CNN model using the multi-year pool of LC samples. The trained model was then applied annually to the ARD images to generate the annual LC maps. A spatiotemporal majority filter with a moving window of 7×3×3 (time, height, width), established after testing different window sizes, was applied to the stack of LC maps from 1990 to 2023 to generate smoothed LC maps. The annual agricultural masks were finally created by merging the cropland, perennial cropland, and grassland pixels into a mask layer per year. Compared with all agricultural pixels in the reference LC database that were neither used for training nor validation, the accuracies of the annual agricultural masks ranged between 92% and 94%. 2.5 Postprocessing of land-use maps The initial agricultural LU maps underwent several postprocessing steps to derive the final maps. All pixels per LU map outside the agricultural mask of the corresponding year were removed. After the removal, a moving spatial majority filter (1×3×3) was applied to each LU map to remove isolated pixels. Given the relative stability of perennial crops (vineyards, hops, plantations) in Germany over time, a moving temporal majority filter (3×1×1) was applied at any perennial pixel location within 3-year windows to reinforce persistence. The final agricultural LU map per year was then created by applying a sieve filter to remove objects smaller than 5 pixels (0.45 ha). 3. Data Records The agricultural LU maps are available in this data repository ( https://cloud.thuenen.de/index.php/s/yLGJNi6MoePycqf ). The dataset consists of 34 cloud-optimized GeoTIFF (COG) images, corresponding to the LU maps from 1990 to 2023. Each image is named according to the format “HCTM_GER_[year]_rst_v101_COG.tif”. All data are projected to the EPSG:3035 (ETRS89-extended / LAEA Europe) coordinate system, with a spatial resolution of 30 m. Each image contains a single band, and each pixel per band contains a positive integer value representing the agricultural LU type. Interpretation of the pixel values can be found in the color table files “HCTM_GER_LegendEN_rst_v101.clr” (English version) and “HCTM_GER_LegendDE_rst_v101.clr” (German version), which are also shared in the same data repository. In the color table, the first column represents the pixel value, the second to the fifth columns represent the RGBA color format, and the last column is the agricultural LU name. 4. Data Overview Fig. 3 shows zoomed-in views of agricultural LU dynamics at the locations of the red boxes (a, b, c) previously shown in Fig. 2 . Each box location corresponds to a 10 km 2 grid. The grid in Fig. 3 a is situated in the old drift morainic landscape of Lower Saxony, where the soil is sandy and less fertile. In both 1990 and 2000, grassland was the most dominant class. However, in both 2010 and 2023, many transitions from grassland to maize occurred, making maize the most dominant class. Similar results in Lower Saxony were observed by other researchers 25 , 26 . Those researchers assigned the German Renewable Energy Act (EEG), first introduced in 2000, as the main driver behind the transitions to maize. In Fig. 3 b, the grid borders Lower Saxony and Saxony-Anhalt, and falls within the loess and sandy loess landscape of Germany, where the soils are the most fertile, hence used more intensively for high-yield crops. Here, the LU types remained relatively stable over time, with winter cereals being the most dominant throughout the period. In Fig. 3 c, the grid is located in the sheet gravel plains and tertiary hills of the Alpine Foreland in Bavaria, where the soils are moderately fertile. In 1990 and 2000, grassland had the highest share. Both 2010 and 2023 saw many grasslands transitioning to winter cereals, which then became the most dominant in both years. In summary, the data overview in Fig. 3 shows the usefulness of our LU maps in capturing long-term trends in agricultural LU and tracking LU changes over time. 5. Technical validation 5.1. Overall accuracies To assess the overall accuracy (OA) of each final agricultural LU map, we used all GSA pixels that were neither used for training nor validation to generate a confusion matrix per year (Figs. S1 – S13 in the Supplementary File). We excluded the years from 2006 to 2010 in the accuracy assessment because the GSA was only available for Brandenburg, which does not cover all 14 classes. Based on each confusion matrix, we computed OA: where TP are the true positives, N is the total number of samples, and K is the number of classes. Table 1 presents the OA values and the corresponding median number of CSOs. Generally, the accuracies were high, ranging between 85% and 93%. The highest accuracy of 92.3% was achieved in 2018, and the lowest (85.3%) was in 2012. Accuracies above 90% were mostly achieved after 2015, when both Landsat and Sentinel images were available. The lowest median number of CSOs and the lowest OA were both recorded in 2012. The Pearson correlation coefficient ( r ) of 0.712 between the OA values and the median number of CSOs was statistically significant (p-value = 0.006), highlighting the impact of the number of observations on classification accuracy. Table 1 Overall accuracy per year. The median CSOs were computed based on the reference areas per year. Year Overall accuracy Median number of CSOs 2010 89.6% 11 2011 89.1% 17 2012 85.3% 6 2013 89.0% 11 2014 91.9% 15 2015 89.3% 23 2016 91.6% 29 2017 91.7% 28 2018 92.3% 54 2019 92.2% 46 2020 91.7% 53 2021 91.8% 37 2022 91.8% 61 5.2. Class accuracies To assess the accuracy of each class per year, we used each confusion matrix from 2010 to 2022 as the basis to calculate the F1-score, which is the harmonic mean of the user’s accuracy (UA) and producer’s accuracy (PA): where TP, FP, and FN are the true positives, false positives, and false negatives, respectively. The F1-scores of each LU type corresponding to the years shown in Table 1 are depicted in Fig. 4 . When averaged over time, the highest accuracies (≥ 90%) were obtained by classes occupying high shares of agricultural lands, like grassland, rapeseed, winter cereals, sugar beet, and maize. In comparison, classes such as fallow land and plantation, which occupy small shares of agricultural land, were predicted with the lowest accuracies (≤ 52%). Similar to the OA results, the lowest F1-scores were mostly recorded in 2012, while the highest were mostly obtained after 2015. The lowest accuracy of grassland was 91.9% (2015), and the highest was 94.4% (2011). Rapeseed, winter cereals, sugar beet, and maize had accuracies greater than 85% in all years, with their lowest accuracies occurring in 2012. The highest accuracies of rapeseed (95.1%) and maize (92.3%) both occurred in 2016, while those of winter cereals (94.9%) and sugar beet (93.9%) both occurred in 2019. Hops had accuracies ranging from 74.8% (2011) to 92.4% (2020). Similar to hops, vineyards were at their lowest (57.5%) in 2011, but the highest (87.9%) was in 2019. In 2012, the potato class registered its lowest accuracy (42.0%), with its highest (92.2%) in 2020. Indeed, the potato class had the highest disparity between its lowest and highest accuracies. Summer cereals were detected at the lowest accuracy (65.0%) in 2011 and the highest (83.3%) in 2018. Legumes were classified with accuracies ranging from 48.7% (2012) and 86.7% (2021). Sunflower accuracies were between 65.7% (2012) and 83.3% (2017). The lowest accuracy of horticultural crops was 43.2% in 2011, and the highest was 71.6% in 2018. Plantation had accuracies ranging from 35.7% (2010) to 51.3% (2021). Overall, the lowest accuracies were recorded for fallow land, ranging from 12.9% (2012) to 39.5% (2022), making it the most difficult class to predict. For a one-shot overview of the confusion between the classes over time, we used the yearly confusion matrices to create a composite confusion matrix (Fig. 5 ). Among the best-performing classes, rapeseed and sugar beet achieved very high producers’ accuracies (PA = 96.4% and 94.2%, respectively), indicating low omission errors and strong separability from other crops. Winter cereals (PA = 93.0%) and grassland (PA = 92.4%) also exhibited high detection rates. Users’ accuracies for these classes were similarly high (UA > 91% for rapeseed, sugar beet, winter cereals, maize, potato, and grassland), demonstrating low commission errors and high thematic reliability of the mapped categories. Maize, potato, sunflower, horticultural crops, and hops showed good to moderate producers’ accuracies (PA between 86% and 89%), although sunflower exhibited a comparatively low user’s accuracy (UA = 68.0%), indicating notable commission errors. Legumes, vineyards, summer cereals, and plantations reached moderate producers’ accuracies (PA between 76% and 81%), suggesting increased omission errors, likely due to spectral and phenological similarities with other crop types. Fallow land was the weakest-performing class, with a low producer’s accuracy (PA = 63.1%) and particularly low user’s accuracy (UA = 19.5%). This indicates both substantial omission and commission errors and highlights the difficulty in distinguishing fallow land from grassland and certain arable crops. Similarly, plantation (UA = 34.0%) and horticultural crops (UA = 46.9%) exhibited low users’ accuracies, suggesting that these mapped classes contain considerable proportions of misclassified pixels. Generally, the dominant misclassification patterns revealed three main confusion groups. First, mutual confusion within the broad cereal group (winter cereals, summer cereals, and maize) was evident. Second, a confusion within the semi-natural vegetation group was noticed, characterized by strong confusion between grassland and fallow land, and to a lesser extent, plantation. Third, confusion among intensively managed specialty crops was observed, particularly between potato, horticultural crops, and sugar beet. These patterns likely reflect overlapping phenological trajectories, similar canopy structures, and management practices that are difficult to disentangle using spectral time series alone. In summary, the classification performs robustly for major crop types with distinct phenological signatures. Classes characterized by heterogeneous management or transitional vegetation states (e.g., fallow land and plantation) remain challenging. However, those classes cover very small shares of agricultural land in Germany. 5.3. Comparison with official statistics To further assess the plausibility of our maps, we compared our map areas with the corresponding areas in the official statistics 46 . For the comparison, we calculated the Pearson correlation coefficient ( r ) and mean deviation ( MD ) between our map areas and the official statistics. With the Pearson correlation coefficient, we measured the strength and direction of the linear relationship between our map areas and those of the official statistics. With MD , we assessed the level of overestimation or underestimation in our map areas over time. Fig. 6 shows the comparison between the map areas (orange lines) and official statistics (blue lines) for the total agricultural area, followed by the most dominant LU classes. The classes have been ordered by their respective Pearson correlation coefficients. Regarding the total agricultural area, a high positive correlation ( r = 0.757) was observed, with a tendency towards overestimation in the map areas for all years. This overestimation can be explained by the fact that in our agricultural LU maps, we map all agricultural areas in the open landscape, whereas the official statistics only account for land that is officially used for agricultural production. Eight classes (maize, sugar beet, summer cereals, rapeseed, grassland, winter cereals, sunflower, potato) showed positive correlations, while mapped legume areas did not correlate with the official statistics. The remaining classes (fallow land, horticultural crops) exhibited negative correlations, confirming the difficulty in accurately predicting them. Fallow lands were often mistaken for grasslands due to similar spectral profiles, thereby increasing uncertainty in their prediction. Horticultural crops, which are usually grown on fields barely larger than a Sentinel-2 pixel 15 , often spectrally mix with adjacent LU classes, making their prediction difficult. Maize and sugar beet had very high correlations ( r > 0.9), with maize being underestimated and sugar beet being overestimated. Summer cereals, rapeseed, grassland, winter cereals, and sunflower had high correlations (0.6 < r ≤ 0.83), and all of them were overestimated. Potatoes, which were underestimated, had a moderate correlation (0.33). Legumes were overestimated and exhibited a weak relationship ( r = 0.001). Although both fallow land and horticultural crops obtained negative correlations, fallow land tended to be underestimated, while horticultural crops were overestimated. Overall, sugar beet obtained the smallest deviation from the official statistics over time. Usage Notes The provided LU maps in the COG format enable data hosting in Hypertext Transfer Protocol (HTTP) servers and efficient data handling in cloud computing environments. Together with the provided color table, the maps can be visualized in various geospatial software, such as QGIS and ArcGIS. The distinct advantage of our LU map time series is the ability to facilitate spatially explicit analysis of agricultural LU changes that cannot be achieved with aggregated data, such as agricultural statistics. However, the statistics extracted from our maps are not meant to replace the official agricultural statistics. The LU maps could be used to spatially and temporally disaggregate the agricultural statistics. Given the significant correlation between the number of CSOs and OAs, the main uncertainty in our products could be tied to the number of CSOs per pixel. Therefore, when using our maps for further analysis, uncertainty modelling ought to be considered. For each LU map, an uncertainty map could be derived by calculating the inverse of the number of CSOs per pixel. While the maps provided in this study are from 2017 to 2023, work is ongoing to update the maps up to the present. This will be done in a rolling manner as follows: existing reference samples will be refined, new samples with a longer time series will be generated, the models will be updated, and then applied to all years from 1990 up to the present. Based on the translation table provided in Table S3, the maps provided in this study could be integrated with the more current and detailed maps of Tetteh et al. 47,48 . Declarations Competing Interests The authors declare no competing interests. Additional Information A supplementary file is attached to this data descriptor and also available at this weblink ( https://1drv.ms/b/c/2ce41295938fadfa/IQCkK9EqF2ndRZRSYgSeNF1KAchORNChpfwPA_91y71SgYw?e=krV3zG ). Author Contributions All authors contributed to the final version of the manuscript through discussion and editing. G.O.T. collected training and validation samples, trained and validated the prediction model, applied the trained model to generate the maps, postprocessed the maps, analysed the results, and wrote the original draft; M.S. and S.E. conceptualized the study and supervised the research. M.S. preprocessed the Landsat and Sentinel images. V.-D.P. implemented the foundational code for model training and prediction. Acknowledgments This research was conducted within the framework of the Nationwide Monitoring Program of Biodiversity in Agricultural Landscapes (MonViA) project, which is under the auspices of the German Federal Ministry of Agriculture, Food, and Regional Identity (BMELH). Additionally, part of the work shown here was financially supported by funds for climate reporting and projection at the Thünen Institute. The BMELH has transferred climate reporting as a permanent, sovereign task to the Thünen Institute, as an official ongoing mandate, to fulfil the legal requirements of the Federal Climate Protection Act and international treaties. Vu-Dong Pham acknowledges support by the research project Fragmented Transformations (German Federal Ministry of Research, Technology, and Space; FKZ 01UC2102). Code Availability The code used in this study to create the LU maps is proprietary and not publicly available. Generating and analyzing the maps was done in Python 3.10.12 using several libraries, including NumPy 1.26.4, GDAL 3.6.2, Rasterio 1.4.3, and Tensorflow 2.20.0. The preprocessing of the Landsat and Sentinel images to create the ARD images was done with FORCE ( https://github.com/davidfrantz/force ), and the visualization of the preprocessed images and predicted LU maps was done in QGIS 3.4.0 LTR. References Kirschke D, Häger A, Schmid JC (2021) New Trends and Drivers for Agricultural Land Use in Germany. In: Weith T et al (eds) Sustainable Land Management in a European Context: A Co-Design Approach. Springer International Publishing, Cham, pp 39–61. doi: 10.1007/978-3-030-50841-8_3 . Lüker-Jans N, Simmering D, Otte A (2016) Analysing Data of the Integrated Administration and Control System (IACS) to Detect Patterns of Agricultural Land-Use Change at Municipality Level. Landsc Online 48–48. 10.3097/LO.201648 Meyer MA, Früh-Müller A (2020) Patterns and drivers of recent agricultural land-use change in Southern Germany. Land Use Policy 99:104959 Niedertscheider M, Kuemmerle T, Müller D, Erb K-H (2014) Exploring the effects of drastic institutional and socio-economic changes on land system dynamics in Germany between 1883 and 2007. Glob Environ Change 28:98–108 Stein S, Steinmann H-H, Isselstein J (2019) Linking Arable Crop Occurrence with Site Conditions by the Use of Highly Resolved Spatial Data. Land 8:65 Atzberger C (2013) Advances in Remote Sensing of Agriculture: Context Description, Existing Operational Monitoring Systems and Major Information Needs. Remote Sens 5:949–981 Weiss M, Jacob F, Duveiller G (2020) Remote sensing for agricultural applications: A meta-review. Remote Sens Environ 236:111402 Donmez E, Heckelei T, Storm H (2025) Satellite remote sensing-based crop cover classification over Europe: accuracy of different methodological approaches. Int J Remote Sens 46:8251–8294 Zhang C et al (2025) Remote sensing for crop mapping: A perspective on current and future crop-specific land cover data products. Remote Sens Environ 330:114995 Huang J et al (2026) Progress and Perspectives of Crop Type Mapping With Remote Sensing: A review. IEEE Geosci Remote Sens Mag 2–35. 10.1109/MGRS.2025.3648119 Asam S et al (2022) Mapping Crop Types of Germany by Combining Temporal Statistical Metrics of Sentinel-1 and Sentinel-2 Time Series with LPIS Data. Remote Sens 14:2981 Blickensdörfer L et al (2022) Mapping of crop types and crop sequences with combined time series of Sentinel-1, Sentinel-2 and Landsat 8 data for Germany. Remote Sens Environ 269:112831 Gessner U, Hirner A, Asam S, Wenzl M, Kuenzer C (2025) Combining Machine Learning and Spatiotemporal Filtering to Map Crop Types of Germany for Seven Years. IEEE Geosci Remote Sens Lett 22:1–5 Griffiths P, Nendel C, Hostert P (2019) Intra-annual reflectance composites from Sentinel-2 and Landsat for national-scale crop and land cover mapping. Remote Sens Environ 220:135–151 Preidl S, Lange M, Doktor D (2020) Introducing APiC for regionalised land cover mapping on the national scale using Sentinel-2A imagery. Remote Sens Environ 240:111673 Schwieder M, Tetteh GO, Blickensdörfer L, Gocht A, Erasmi S (2024) Agricultural land use (raster): National-scale crop type maps for Germany from combined time series of Sentinel-1, Sentinel-2 and Landsat data (2017 to 2021). Zenodo https://doi.org/10.5281/zenodo.10640528 Tetteh GO, Schwieder M, Blickensdörfer L, Gocht A, Erasmi S (2024) Agricultural land use (vector): National-scale crop type maps for Germany from combined time series of Sentinel-1, Sentinel-2 and Landsat data (2017 to 2021). Zenodo https://doi.org/10.5281/zenodo.10619783 German Environment Agency. National Inventory Report for the German Greenhouse Gas Inventory 1990–2004 (2009) Laamrani A, Voroney PR, Gillespie AW, Chehbouni A (2021) Development of a Land Use Carbon Inventory for Agricultural Soils in the Canadian Province of Ontario. Land 10:765 Lokupitiya E, Paustian K (2006) Agricultural Soil Greenhouse Gas Emissions. J Environ Qual 35:1413–1427 Li X, Yu L, Du Z, Liu X (2025) Crop Statistic to Annual Map: Tracking spatiotemporal dynamics of crop-specific areas through machine learning and statistics disaggregating. Sci Data 12:1249 Lamsal G, Marston LT (2024) HarvestGRID: high-resolution harvested crop areas of the United States from 1981 to 2019. Environ Res Food Syst 2:015006 Frank C et al (2024) Woody semi-natural habitats modulate the effects of field size and functional crop diversity on farmland birds. J Appl Ecol 61:987–999 Kasiske T et al (2025) Identifying typical patterns of land-use and landscape structure in citizen science butterfly monitoring. Ecol Indic 180:114317 Jänicke C et al (2022) Field-level land-use data reveal heterogeneous crop sequences with distinct regional differences in Germany. Eur J Agron 141:126632 Stein S, Steinmann H-H (2018) Identifying crop rotation practice by the typification of crop sequence patterns for arable farming systems – A case study from Central Europe. Eur J Agron 92:30–40 der Velde M et al (2024) Policy relevant crop diversity monitoring based on earth observation and farmers’ declarations. Environ Res Lett 20:014054 Rousi M et al (2021) Semantically Enriched Crop Type Classification and Linked Earth Observation Data to Support the Common Agricultural Policy Monitoring. IEEE J Sel Top Appl Earth Obs Remote Sens 14:529–552 Hendricks NP, Smith A, Sumner DA (2014) Crop Supply Dynamics and the Illusion of Partial Adjustment. Am J Agric Econ 96:1469–1491 Li J, Roy DP (2017) A Global Analysis of Sentinel-2A, Sentinel-2B and Landsat-8 Data Revisit Intervals and Implications for Terrestrial Monitoring. Remote Sens 9:902 Claverie M et al (2018) The Harmonized Landsat and Sentinel-2 surface reflectance data set. Remote Sens Environ 219:145–161 Foerster S, Kaden K, Foerster M, Itzerott S (2012) Crop type mapping using spectral–temporal profiles and phenological information. Comput Electron Agric 89:30–40 Azar R et al (2016) Assessing in-season crop classification performance using satellite data: a test case in Northern Italy. Eur J Remote Sens 49:361–380 Jänicke C, Petersen KA, Schmidts P, Müller D, Jepsen MR (2025) Field and farm-level data on agricultural land use for the European Union. Sci Data 12:1050 Schneider M, Schelte T, Schmitz F, Körner M (2023) EuroCrops: The Largest Harmonized Open Crop Dataset Across the European Union. Sci Data 10:612 Pham V-D et al (2024) Temporally transferable crop mapping with temporal encoding and deep learning augmentations. Int J Appl Earth Obs Geoinf 129:103867 Frantz D (2019) FORCE—Landsat + Sentinel-2 Analysis Ready Data and Beyond. Remote Sens 11:1124 Buchner J et al (2020) Land-cover change in the Caucasus Mountains since 1987 based on the topographic correction of multi-temporal Landsat composites. Remote Sens Environ 248:111967 Frantz D, Roder A, Stellmes M, Hill J (2016) An Operational Radiometric Landsat Preprocessing Framework for Large-Area Time Series Applications. IEEE Trans Geosci Remote Sens 54:3928–3943 Roy DP, Li Z, Zhang HK (2017) Adjustment of Sentinel-2 Multi-Spectral Instrument (MSI) Red-Edge Band Reflectance to Nadir BRDF Adjusted Reflectance (NBAR) and Quantification of Red-Edge Band BRDF Effects. Remote Sens 9:1325 Frantz D, Haß E, Uhl A, Stoffels J, Hill J (2018) Improvement of the Fmask algorithm for Sentinel-2 images: Separating clouds from bright surfaces based on parallax effects. Remote Sens Environ 215:471–481 Zhu Z, Wang S, Woodcock CE (2015) Improvement and expansion of the Fmask algorithm: cloud, cloud shadow, and snow detection for Landsats 4–7, 8, and Sentinel 2 images. Remote Sens Environ 159:269–277 Zhu Z, Woodcock CE (2012) Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens Environ 118:83–94 Pham V-D et al (2024) An annual land cover dataset for the Baltic Sea Region with crop types and peat bogs at 30 m from 2000 to 2022. Sci Data 11:1242 Gensior A, Drexler S, Fuß R, Stümer W, Rüter S (2025) Emissions of greenhouse gases from land use, land-use change and forestry (LULUCF). https://www.thuenen.de/en/thuenen-topics/climate-and-air/emission-inventories-accounting-for-climate-protection/treibhausgas-emissionen-durch-landnutzung-landnutzungsaenderung-und-forstwirtschaft-lulucf Federal Statistical Office (2025) The database of the Federal Statistical Office. https://www-genesis.destatis.de/datenbank/online Tetteh GO, Schwieder M, Blickensdörfer L, Gocht A, Erasmi S (2024) Agricultural land use (raster): National-scale crop type maps for Germany from combined time series of Sentinel-2 and Landsat data Zenodo https://doi.org/10.5281/zenodo.17197830 (2025) Tetteh GO, Schwieder M, Blickensdörfer L, Gocht A, Erasmi S (2025) Agricultural land use (raster): National-scale crop type maps for Germany from combined time series of Sentinel-2 and Landsat data Zenodo https://doi.org/10.5281/zenodo.17197871 (2025) Additional Declarations The authors declare no competing interests. Supplementary Files SupplementAnnualCropTypeMapsFinal.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9074257","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Data Note","associatedPublications":[],"authors":[{"id":603204436,"identity":"6f5ecaa9-c718-425e-9308-d1aae04ac944","order_by":0,"name":"Gideon Okpoti Tetteh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABB0lEQVRIiWNgGAWjYDACZgYDJHYFQwKcTUiLBAMbM5B9xoAILQzIWhjbiNAi78688cHHHQx1/PL9Bx8XzvuTZ3C89+DjAgZrOVxaDA+zFRvOPMMgIdnGzGw8c5tBscGZc8nGMxjSjXFqaeYxk+ZtY5AwOMbMJs27zSBxw40cM2kehsOJDbi1mP8GabE/xsz+m3cOUMv9N+a/gVrqcWmRZ+YxYwbbwsbMxszbALIFKALUkoBDB4MBM1ux5Mw2CckZx5KNpXmOGSfOPJMDZBikG+K0pf/wxg8f22z4+ZsPPvzMUyOX2Hf8jOFnngpreZy2HABTEhjiuDQAbcFl/SgYBaNgFIwCOAAAyCFLezjZ90cAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0001-5430-5967","institution":"Thünen Institute of Farm Economics, Bundesallee 63, 38116 Braunschweig, Germany","correspondingAuthor":true,"prefix":"","firstName":"Gideon","middleName":"Okpoti","lastName":"Tetteh","suffix":""},{"id":603211360,"identity":"56e3eebc-4feb-4ddb-9436-35779adbca51","order_by":1,"name":"Vu-Dong Pham","email":"","orcid":"https://orcid.org/0000-0001-5005-6510","institution":"Earth Observation and Geoinformation Science Lab, Institute of Geography and Geology, University of Greifswald, Friedrich-Ludwig-Jahn-Str. 16, 17489 Greifswald, Germany | Interdisciplinary Centre for Baltic Sea Region Research (IFZO), University of Greifswald, 17489 Greifswald, Germany","correspondingAuthor":false,"prefix":"","firstName":"Vu-Dong","middleName":"","lastName":"Pham","suffix":""},{"id":603211361,"identity":"531ad179-bb2b-4e01-9fa5-94ff28953eb9","order_by":2,"name":"Marcel Schwieder","email":"","orcid":"https://orcid.org/0000-0003-2103-8828","institution":"Thünen Institute of Farm Economics, Bundesallee 63, 38116 Braunschweig, Germany | Geography Department, Humboldt University of Berlin, Unter den Linden 6, 10099 Berlin, Germany","correspondingAuthor":false,"prefix":"","firstName":"Marcel","middleName":"","lastName":"Schwieder","suffix":""},{"id":603211362,"identity":"dc75c87b-d2bc-4b34-9730-9761f45718c3","order_by":3,"name":"Lukas Blickensdörfer","email":"","orcid":"https://orcid.org/0000-0001-7255-1416","institution":"Thünen Institute of Farm Economics, Bundesallee 63, 38116 Braunschweig, Germany | Geography Department, Humboldt University of Berlin, Unter den Linden 6, 10099 Berlin, Germany","correspondingAuthor":false,"prefix":"","firstName":"Lukas","middleName":"","lastName":"Blickensdörfer","suffix":""},{"id":603211363,"identity":"790fc190-695b-4c39-9d94-dec769fb6c5d","order_by":4,"name":"Alexander Gocht","email":"","orcid":"https://orcid.org/0000-0002-8913-1538","institution":"Thünen Institute of Farm Economics, Bundesallee 63, 38116 Braunschweig, Germany | Agricultural Development and Trade Group, Department of Agricultural Economics, Humboldt University of Berlin, Unter den Linden 6, 10099 Berlin, Germany","correspondingAuthor":false,"prefix":"","firstName":"Alexander","middleName":"","lastName":"Gocht","suffix":""},{"id":603211364,"identity":"8ace148b-1e03-41af-9d91-b188f60b1f94","order_by":5,"name":"Sebastian van der Linden","email":"","orcid":"https://orcid.org/0000-0001-6576-8377","institution":"Earth Observation and Geoinformation Science Lab, Institute of Geography and Geology, University of Greifswald, Friedrich-Ludwig-Jahn-Str. 16, 17489 Greifswald, Germany | Interdisciplinary Centre for Baltic Sea Region Research (IFZO), University of Greifswald, 17489 Greifswald, Germany","correspondingAuthor":false,"prefix":"","firstName":"Sebastian","middleName":"van der","lastName":"Linden","suffix":""},{"id":603211365,"identity":"1474459c-0e1f-432b-977d-97cf8ef831ac","order_by":6,"name":"Stefan Erasmi","email":"","orcid":"https://orcid.org/0000-0002-6393-6071","institution":"Thünen Institute of Farm Economics, Bundesallee 63, 38116 Braunschweig, Germany","correspondingAuthor":false,"prefix":"","firstName":"Stefan","middleName":"","lastName":"Erasmi","suffix":""}],"badges":[],"createdAt":"2026-03-09 14:35:22","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9074257/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9074257/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104342366,"identity":"d1377a5a-4c82-4192-b813-4b6ddcb906fb","added_by":"auto","created_at":"2026-03-10 17:02:04","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":86242,"visible":true,"origin":"","legend":"\u003cp\u003eThe general workflow we followed to create the agricultural land-use maps of Germany.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/be73abfde59fd61ce4d30fdb.png"},{"id":104405925,"identity":"53c8c7fc-52a3-4ca0-bba5-d4445668d58c","added_by":"auto","created_at":"2026-03-11 12:24:13","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":327589,"visible":true,"origin":"","legend":"\u003cp\u003eThe number of available GSA datasets in Germany between 1900 and 2023. The years in which GSA datasets were available per federal state are in parentheses. Red boxes (a, b, c): see detailed views in Fig. 3. BW: Baden-Württemberg, BY: Bavaria, BE: Berlin, BB: Brandenburg, HB: Bremen, HH: Hamburg, HE: Hesse, NI: Lower Saxony, MV: Mecklenburg-Vorpommern, NW: North Rhine-Westphalia, RP: Rhineland-Palatinate, SL: Saarland, SN: Saxony, ST: Saxony-Anhalt, SH: Schleswig-Holstein, TH: Thuringia.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/11ad0bc52324eb9200d045ab.png"},{"id":104342368,"identity":"99d0c7b5-aee1-4c4b-a698-8b5c7e9de417","added_by":"auto","created_at":"2026-03-10 17:02:04","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1483372,"visible":true,"origin":"","legend":"\u003cp\u003eAgricultural land-use maps of 1990, 2000, 2010, and 2023 for the red boxes (a, b, c) shown in Fig. 2.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/14cd01c964f8da6253d9e954.png"},{"id":104342372,"identity":"8f68c543-73ad-4ec4-94f1-c81b3ec13668","added_by":"auto","created_at":"2026-03-10 17:02:04","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":293114,"visible":true,"origin":"","legend":"\u003cp\u003eF1-scores for each agricultural LU type from 2010 to 2022.\u003c/p\u003e","description":"","filename":"image4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/bc387b312baf6f43b31f5115.jpeg"},{"id":104780107,"identity":"f5f9688f-278a-45e3-a104-3d4a6ea00be2","added_by":"auto","created_at":"2026-03-17 07:50:29","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":421641,"visible":true,"origin":"","legend":"\u003cp\u003eThe composite confusion matrix generated from all yearly confusion matrices from 2010 to 2022 (Figs. S1 – S13 in the Supplementary File). The matrix entries are areas (km\u003csup\u003e2\u003c/sup\u003e).\u003c/p\u003e","description":"","filename":"image5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/2a96b792e8bca3e4af6945f5.jpeg"},{"id":104445051,"identity":"3cb20651-7ec0-4828-834c-baa5e8cedcd7","added_by":"auto","created_at":"2026-03-11 19:42:42","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":421641,"visible":true,"origin":"","legend":"\u003cp\u003eThe composite confusion matrix generated from all yearly confusion matrices from 2010 to 2022 (Figs. S1 – S13 in the Supplementary File). The matrix entries are areas (km2).\u003c/p\u003e","description":"","filename":"image5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/58778bf82a1305ec26b9061b.jpeg"},{"id":104342373,"identity":"60c8a304-b9da-4fe8-940b-4e0e4c176876","added_by":"auto","created_at":"2026-03-10 17:02:04","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":525496,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of map areas with official statistics. In all plots, \u003cem\u003er\u003c/em\u003e is the Pearson correlation coefficient, and \u003cem\u003eMD\u003c/em\u003e is the mean deviation between the map areas and the official statistics.\u003c/p\u003e","description":"","filename":"image6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/1e88ebc0fb3696dc64565ee2.jpeg"},{"id":104835489,"identity":"49b2acdd-5b57-443d-89e8-c4a98c74b4a3","added_by":"auto","created_at":"2026-03-17 17:45:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4123334,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/7f33116b-77fa-48c2-a342-b08a70830f99.pdf"},{"id":104342371,"identity":"29d46c3d-8484-4f2c-b6ca-c615b9a41a22","added_by":"auto","created_at":"2026-03-10 17:02:04","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":5264571,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementAnnualCropTypeMapsFinal.docx","url":"https://assets-eu.researchsquare.com/files/rs-9074257/v1/cae2ce77632fa7c02211d93e.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eNationwide annual agricultural land-use maps of Germany from 1990 to 2023 derived from satellite imagery\u003c/p\u003e","fulltext":[{"header":"1. Background \u0026 Summary","content":"\u003cp\u003eIn Germany, the land-use (LU) in agricultural landscapes has undergone significant changes over the past decades. The drivers behind those changes are multifaceted, reflecting a complex interplay of socio-economic, policy, and environmental factors \u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. These drivers have collectively shaped the agricultural landscape, influencing the extent, type, and intensity of agricultural LU. To design effective policies that foster sustainable agriculture, we need a comprehensive understanding of those drivers and their impact on agricultural LU. This requires long-term, spatially explicit information on agricultural LU.\u003c/p\u003e \u003cp\u003ePrimary data sources such as the Authoritative Topographic-Cartographic Information System (ATKIS), the Geo-spatial Application (GSA) of the Common Agricultural Policy (CAP) framework, and official agricultural statistics are available in Germany for monitoring agricultural lands. However, those data sources have certain limitations. In ATKIS, only broad land-cover (LC) classes, such as cropland and grassland, are mapped, without detailed information on agricultural LU types. While detailed information on agricultural LU types can be found in the GSA, the data has limited spatial and temporal coverage, and it is not publicly available for all federal states. Although the official agricultural statistics contain detailed information on agricultural LU and productivity, a comprehensive farm survey is conducted only every 3\u0026ndash;4 years at sample farm locations. Additionally, the survey data are aggregated to the district level, thereby preventing spatially explicit analysis of agricultural LU changes. Further, the LU definitions used in the statistics change over time, thereby requiring considerable effort to consolidate the LU classes.\u003c/p\u003e \u003cp\u003eRemote sensing data, with their high spatial and temporal coverage, are suitable for continuous and long-term area-wide mapping of agricultural lands \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e, thereby providing a unique means of overcoming the limitations of the abovementioned data sources. The literature is replete with the use of remote sensing data for generating agricultural LU maps at local, national, and continental scales \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. The proliferation of agricultural LU maps has been driven by key developments, including the increasing volume of data from various satellite missions, advances in artificial intelligence methods for agricultural LU mapping, and greater access to all-in-one cloud computing platforms \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe launch of Sentinel-1 in 2014 and Sentinel-2 in 2015, as part of the Copernicus satellite program, has accelerated the creation of national-scale agricultural LU maps across Europe and beyond. This is particularly true for Germany, where numerous nationwide products exist \u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13 CR14 CR15 CR16\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. However, detailed national-scale agricultural LU maps for years before the Copernicus era, dating back to 1990, are rarely available, making effective long-term monitoring difficult. The year 1990 is pivotal because it marks the baseline year for mandatory national greenhouse gas (GHG) inventory reporting by countries, including Germany, that are parties to the United Nations Framework Convention on Climate Change (UNFCCC) \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Therefore, high-resolution, area-wide, and long-term LU maps in Germany will not only aid in understanding the drivers behind agricultural LU changes but can also improve national GHG emissions reporting, because the maps can be used for Tier-3 (process-based) modelling \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e as defined in the Intergovernmental Panel on Climate Change (IPCC) guidelines. Other potential uses of such long-term maps include the spatial and temporal disaggregation of agricultural statistics \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, assessing agricultural landscape heterogeneity within the context of biodiversity monitoring \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, analyzing crop sequences and rotations \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, evaluating agricultural policies \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e, and estimating crop supply response \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThis study presents a historical dataset of agricultural LU maps for Germany, containing 13 main crops and one grassland class from 1990 to 2023, generated from the entire Landsat (1990\u0026ndash;2023) and Sentinel-2 (2015\u0026ndash;2023) archives using a deep learning (DL) approach. The combination of Landsat and Sentinel-2 from 2015 onwards yields a temporally dense time series \u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, which enables the differentiation of agricultural LU types based on their phenological profiles \u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e,\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. However, the time series before 2015 consists only of Landsat data and thus has lower spatial resolution and larger temporal gaps, making detailed mapping of agricultural LU more challenging. Further, LU mapping relies on reference data, which is needed for model training and validation. In more recent years, reference data such as the agricultural parcels declared by farmers in the GSA of CAP have become available in the European Union (EU) \u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e,\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. However, due to the limited spatial and temporal availability of the GSA data in Germany, long-term agricultural LU mapping has been hampered. For model training and validation in this study, we had access to GSA data covering different parts of Germany from 2006 to 2022. To overcome the lack of reference data before 2006 and the temporarily sparse satellite data before 2015, we adapted the DL approach of Pham et al. \u003csup\u003e36\u003c/sup\u003e, which was designed for LU mapping in data-sparse conditions.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cp\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the general workflow we used in this study to create the agricultural LU maps. The main components of the workflow are satellite data preprocessing, agricultural LU mapping, agricultural masks creation, postprocessing of agricultural LU maps, and technical validation. Each component is explained in subsequent sections.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Preprocessing of satellite data\u003c/h2\u003e \u003cp\u003eIn this study, all Level-1 (top-of-atmosphere) Landsat images from 1990 to 2023 and Sentinel-2 images from 2015 to 2023 with cloud cover less than 75% were downloaded and preprocessed to generate analysis-ready data (ARD) of Germany using the Framework for Operational Radiometric Correction for Environmental Monitoring (FORCE) \u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. Generating the ARD images involved converting the top-of-atmosphere images to Level-2 (surface reflectance) images by correcting for radiometric, atmospheric, and geometric distortions \u003csup\u003e\u003cspan additionalcitationids=\"CR39\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e, masking of clouds and cloud shadows using the Fmask algorithm \u003csup\u003e\u003cspan additionalcitationids=\"CR42\" citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e, and resampling of images to 30 m. Following the approach of Pham et al. \u003csup\u003e36\u003c/sup\u003e, we extracted the bands present in both sensors (red, green, blue, near-infrared, shortwave-infrared 1, shortwave-infrared 2) and generated three spectral indices, namely Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Soil-Adjusted Vegetation Index (SAVI). Consequently, each ARD image contained nine bands (six spectral bands and three spectral indices). An overview table showing the number of clear-sky observations (CSOs) generated from the ARD images per year, alongside the corresponding sensors, can be found in Table S1 of the Supplementary File.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Agricultural land-use reference data\u003c/h2\u003e \u003cp\u003eAs reference data for agricultural LU mapping, we used the agricultural parcels declared by the EU farmers in the GSA as part of the Integrated Administration and Control System (IACS) to receive CAP payments. The GSA datasets include the spatial boundaries and crop types of all parcels for which subsidies were requested. The federal states in Germany are responsible for reviewing all GSA applications and subsequently providing the data for research and development based on individual data use agreements. Consequently, data availability varies between the federal states. The spatial and temporal coverage of the GSA data we had access to in Germany is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Before 2006, GSA datasets were not available. The federal state of Brandenburg had the highest number of available GSA datasets from 2006 to 2022, while the federal state of Saarland had the lowest (2014\u0026ndash;2016). No GSA data was available for Schleswig-Holstein and Hamburg. The GSA for Bremen and Berlin are included in the datasets of Lower Saxony and Brandenburg, respectively.\u003c/p\u003e \u003cp\u003eIn 2014, 2015, and 2016, the GSA had the highest spatial coverage (12 federal states). To ensure consistency in class definitions over time and with groups of classes used in the agricultural statistics of Germany, we extracted the main GSA classes and translated them into a condensed set of 14 classes: winter cereals, summer cereals, maize, grassland, potato, sugar beet, rapeseed, sunflower, legumes, horticultural crops, fallow land, vineyards, hops, and plantations. The three perennial classes (vineyards, hops, and plantations) were included to potentially identify the year of establishment. The translation table is in Table S2 of the Supplementary File.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Agricultural land-use mapping\u003c/h2\u003e \u003cp\u003eGiven the non-availability of GSA before 2006 and the relatively low temporal density of ARD images before 2015, we opted for the mapping approach of Pham et al. \u003csup\u003e36\u003c/sup\u003e, which has been successfully used to map annual land cover and land use in the Baltic Sea region from 2000 to 2022 \u003csup\u003e44\u003c/sup\u003e. This approach has three components: (1) the use of temporal encoding for generating annual time series data per pixel, (2) the application of two data augmentation techniques, namely random observation selection (ROS) and random day shifting (RDS), to simulate temporal patterns of data-sparse conditions and phenological variations, (3) the use of a 1D-CNN model for training and prediction.\u003c/p\u003e \u003cp\u003eThe temporal encoding involves generating arrays of 365 values per pixel and band, where each value represents the CSO for the day of year (DOY). Where no CSO is available for a particular DOY, zero is inserted into the array. The ROS involves randomly selecting proportions of the temporally encoded data for each reference sample during training, with the proportions ranging between 5% and 100%, to simulate data-sparse situations during model training. This is followed by RDS, which involves the random shifting of observations forward or backward within sixteen days, which corresponds to the temporal resolution of Landsat. This augmented time series, together with the class-wise reference labels, is finally passed to the 1D-CNN model for training. The 1D-CNN model comprises four convolution blocks. Each convolution block consists of two convolution layers, batch normalization, and rectified linear unit activation. Max pooling is applied after the first, second, and third convolution blocks, while the output of the fourth convolution block is flattened and attached to a dense layer for prediction. For more details regarding the approach, readers are referred to Pham et al. \u003csup\u003e36\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo train the model, we randomly sampled 10,000 pixels per class for each year with GSA data. Per year and class, the samples were split into 75% for training and 25% for validation. The respective training and validation samples per year were then merged to create a multi-year training and validation pool. The multi-year training samples were used to train the 1D-CNN model, while the multi-year validation samples were used to validate model performance during training. The model was trained for 100 epochs with a batch size of 512 and a learning rate of 0.001 on a computer with 48 GB of graphics processing unit (GPU) memory. After training, the model was applied to each ARD pixel per year to generate the initial agricultural LU maps of Germany from 1990 to 2023.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Agricultural masks creation\u003c/h2\u003e \u003cp\u003eThe initial agricultural LU maps created in the previous section contain information for all pixel locations in Germany. To remove the non-agricultural pixels from those initial LU maps, we created annual agricultural masks. To create the annual agricultural masks, we first generated annual LC maps for Germany. As reference data for LC mapping, we used the German-wide reference database used for the calculation of GHG emissions in Germany by the Th\u0026uuml;nen Institute \u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e. This reference database, which contains the main LC types of Germany, was created from a combination of LC datasets from the Coordination of Information on the Environment (CORINE) before 2000 and ATKIS thereafter. For our research, we had access to the LC samples in the reference database from 1990 to 2022.\u003c/p\u003e \u003cp\u003ePer year, we randomly sampled 1,500 pixels from the following LC classes: forest, cropland, perennial cropland, grassland, woody plants and hedges, wetlands, water, peat mining, built-up areas, and others. Those ten classes cover the entire surface area of Germany. The LC samples were split into 75% training and 25% validation per year and class, and subsequently merged to create a multi-year training and validation pool. We followed the same training procedure used for LU mapping, and trained another 1D-CNN model using the multi-year pool of LC samples. The trained model was then applied annually to the ARD images to generate the annual LC maps.\u003c/p\u003e \u003cp\u003eA spatiotemporal majority filter with a moving window of 7\u0026times;3\u0026times;3 (time, height, width), established after testing different window sizes, was applied to the stack of LC maps from 1990 to 2023 to generate smoothed LC maps. The annual agricultural masks were finally created by merging the cropland, perennial cropland, and grassland pixels into a mask layer per year. Compared with all agricultural pixels in the reference LC database that were neither used for training nor validation, the accuracies of the annual agricultural masks ranged between 92% and 94%.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Postprocessing of land-use maps\u003c/h2\u003e \u003cp\u003eThe initial agricultural LU maps underwent several postprocessing steps to derive the final maps. All pixels per LU map outside the agricultural mask of the corresponding year were removed. After the removal, a moving spatial majority filter (1\u0026times;3\u0026times;3) was applied to each LU map to remove isolated pixels. Given the relative stability of perennial crops (vineyards, hops, plantations) in Germany over time, a moving temporal majority filter (3\u0026times;1\u0026times;1) was applied at any perennial pixel location within 3-year windows to reinforce persistence. The final agricultural LU map per year was then created by applying a sieve filter to remove objects smaller than 5 pixels (0.45 ha).\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Data Records","content":"\u003cp\u003eThe agricultural LU maps are available in this data repository (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cloud.thuenen.de/index.php/s/yLGJNi6MoePycqf\u003c/span\u003e\u003cspan address=\"https://cloud.thuenen.de/index.php/s/yLGJNi6MoePycqf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The dataset consists of 34 cloud-optimized GeoTIFF (COG) images, corresponding to the LU maps from 1990 to 2023. Each image is named according to the format \u0026ldquo;HCTM_GER_[year]_rst_v101_COG.tif\u0026rdquo;. All data are projected to the EPSG:3035 (ETRS89-extended / LAEA Europe) coordinate system, with a spatial resolution of 30 m. Each image contains a single band, and each pixel per band contains a positive integer value representing the agricultural LU type. Interpretation of the pixel values can be found in the color table files \u0026ldquo;HCTM_GER_LegendEN_rst_v101.clr\u0026rdquo; (English version) and \u0026ldquo;HCTM_GER_LegendDE_rst_v101.clr\u0026rdquo; (German version), which are also shared in the same data repository. In the color table, the first column represents the pixel value, the second to the fifth columns represent the RGBA color format, and the last column is the agricultural LU name.\u003c/p\u003e"},{"header":"4. Data Overview","content":"\u003cp\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows zoomed-in views of agricultural LU dynamics at the locations of the red boxes (a, b, c) previously shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Each box location corresponds to a 10 km\u003csup\u003e2\u003c/sup\u003e grid. The grid in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea is situated in the old drift morainic landscape of Lower Saxony, where the soil is sandy and less fertile. In both 1990 and 2000, grassland was the most dominant class. However, in both 2010 and 2023, many transitions from grassland to maize occurred, making maize the most dominant class. Similar results in Lower Saxony were observed by other researchers \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Those researchers assigned the German Renewable Energy Act (EEG), first introduced in 2000, as the main driver behind the transitions to maize.\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb, the grid borders Lower Saxony and Saxony-Anhalt, and falls within the loess and sandy loess landscape of Germany, where the soils are the most fertile, hence used more intensively for high-yield crops. Here, the LU types remained relatively stable over time, with winter cereals being the most dominant throughout the period. In Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec, the grid is located in the sheet gravel plains and tertiary hills of the Alpine Foreland in Bavaria, where the soils are moderately fertile. In 1990 and 2000, grassland had the highest share. Both 2010 and 2023 saw many grasslands transitioning to winter cereals, which then became the most dominant in both years. In summary, the data overview in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the usefulness of our LU maps in capturing long-term trends in agricultural LU and tracking LU changes over time.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"5. Technical validation","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e5.1. Overall accuracies\u003c/h2\u003e\n \u003cp\u003eTo assess the overall accuracy (OA) of each final agricultural LU map, we used all GSA pixels that were neither used for training nor validation to generate a confusion matrix per year (Figs. S1 \u0026ndash; S13 in the Supplementary File). We excluded the years from 2006 to 2010 in the accuracy assessment because the GSA was only available for Brandenburg, which does not cover all 14 classes. Based on each confusion matrix, we computed OA:\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAlYAAAAuCAYAAADqS1eaAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAAAiiSURBVHhe7d2/TxP/Hwfw533/AQJlQl16XZggWNCQMsDAVVaNrcFZbElINAYMlo0aqYbBmJIaTFhIrsTExASkNbEDJym0AzewcRcSiVOvNfn8Afdd6KV3bUGwSinPR3JRXu9X8cf0zPv95nWCaZomiIiIiOiP/c9ZICIiIqLzYbAiIiIiahAGKyIiIqIGYbAiIiIiahAGKyIiIqIGYbAiIiIiapArH6wMw4AgCBAEAYqiIBaLWV8TERERncWVD1YulwuhUAgLCwvw+XyQJAmSJEHTNGcrERER0YlaLlglEglrx6neE4vFbJ9Jp9OQJAmJRALZbBabm5twu922HiIiIqLTtFywunfvHkRRhCiKKBQKME3Tevb29tDe3m7r13Udmqaht7cXy8vLmJiYsK2fh6IoVWGu8pFl2er1+/1V6+FwGIZh2L4nERERNb+WC1YulwsrKyvQNA2Tk5O2tZ6eHoRCIVvt69evCAQCKBQK0DTNFnrOa3t7G16vF4VCAevr6wAA0zRRKBTQ3t6O7u5uq3d1dRUAsL6+boU/WZbx8eNHq4eIiIguh5YLVgDg8/kwOzuLZDKJjY0N21o0GsX09LT19adPn9DX1weXy4XPnz83ZLfo8PAQy8vLcLlc+P79OyRJAo5D38DAAHp6eqzeo6MjAMCtW7eA4/A3MDCATCZj9RAREdHl0JLBCgCePn0KURTx8OHDukFpY2MDqVQKr169Ao53mkqlEsbHx52tZxKPx63wlE6nMTIyYq1tbm5WdALZbBZerxcul8uqHRwcoK+vz9ZHREREzU8wTdN0FluFoigYGhqCJElVgeZfMAwDnZ2d2Nrags/ncy4DAILBIERRRDQaha7rePPmDWRZRj6f5wV6IiKiS6Zld6xQcSSYSqWgqqpz+a/b2dkBjv8e9aTTabx8+RKCIEAURRSLRWQyGYYqIiKiS6ilgxUAlEolhEIh272mf6XyflUtqqqiVCphb2/P+slFWZatv6uu6+jo6HB+jIiIiJpUSwcrWZaRy+UQj8edSzbhcBiJRMJZrsswDMRiMQSDQeeSjfN+lVM2m4UoinVDn9vtRrFYdJaJiIioSbVssFJVFZFIBMlk0lZXFAV+v99Wi8fj1vyqylfaVD7loaKGYWBxcREA8OvXL9v3qaSqKvL5vLNss7y8DI/H4ywDFbOwzhL4iIiI6GK1ZLAyDAN3797F27dvq+4q7e/v4+bNm9bXfr/fdtw2PT1tGypafsojGlwuF6LRKAYHB63P1NLb2wsAmJmZqXm/KxaLIZ/PI5VKVU2Cx/G9LEmScPv2becSERERNamWDFaTk5N49OgRxsbGbHXDMPD69Wu0tbVZtUgkgtHRUVtfI1SGslpHfZUBrnKuVqWDg4OanyUiIqLm1HLBKpFIIJlMYmZmpuo4r7Ozs+rlyvv7+xgeHra+Pu0o8F9RVRUejweKokDXdecyERERNaE/ClaKoiAYDFrh43emlns8nr8aUiYmJqqO8ZxP5Q5RJpPBjRs3rAntpx0Flu3v75/6b/1TqVQKX758qTrOJCIiouZ07mAlyzKGhoYwPDwM0zShaRrS6TTm5uacrZZyoDo8PHQuXaipqSlcu3bNWa5LEAQ8fvwY+XwegiBAURRnyx/r6emBaZqIRqPOJSIiImpS55q8rus6vF4vnj9/btvJicVieP/+PQ4ODmz9OL7fND4+jpGREXz79u1CJqETERER/U3n2rH68OEDOjo6qo7HANSduzQ3N4dIJOIsExEREV0oRVEQDoed5Sp+v//Ue8/nClbJZBKBQMBZxuHhIQYGBpxl66jspFe7VKp3gfwiL5MTERFR65FlGfPz81XDxD0eT1XY2tzchCiKJ14BOnOw0nUdmqbhzp07ziWk02nbjKiyd+/e4dmzZwCAtrY27O7uOlts6l0gP+kyeZkzgLXSQ0RERI2j6zoikYjtepKu6xAEoWqKQJlpmhgaGnKWLWcOVj9//gQAdHV12eqKokDTNNy/f99Wl2UZyWQSoihCOL70XSqVbD2N5AxgrfQQERFR44TDYczPz9tqbrcbpmme+K7fhYWFqt2ssjMHq7L//vvP+r1hGHjy5AkkSbINtDQMA+FwGJqmWeFga2vLWq+HR4FERET0N+m6jlQqdep7f2sZHBzE0tKSswycJ1j5fD60t7db77DTdR2Tk5MolUpYXV219S4uLiIYDNrmMB0dHQHHoauePzkKJCIiIjrN7u4uRFF0ln9L+c54rYvsZw5WOB6qmcvlIAgCvF4vOjo6sLOzA5fLZfUkEgm8fPkSuq5bf7CiKHjw4AFwHLqIiIiILqvy9ahK5wpWPT09yOVyME0TxWIR8XjcFqpQMQF9c3PT2rHy+XzWjtNlHXxZPop88eKFVVMUxaqf9JMCRERE1NrOFayusvX1dUiShGQyadV8Ph9CoRBmZ2d/e6QEERERXZzr1687Sw3BYHVGP378QCQSQbFYhKqqVl3X9ZojKIiIiKj5dHV1QdO0mvekTnPSfE4GqzPKZDLo7u5GMBjE2tqaVd/d3a35H0xERETNx+12Q5KkmrM1BUFAKpXC0tISPB6Pcxnb29sIhULOMnDedwVeZf39/cjlctjY2MDU1BQODg6gqipmZmb4/kMiIqJLRNd1jI6O1nzH8UkEQag7X5I7Vmegqir6+/sBAGNjY9A0DaqqIpvNYmRkxNlORERETcztdmN+fr7mrlQ9Ho/nxJmcDFZnkM1mbQNQA4EA1tbWkMlkMDg4aOslIiKi5hcMBrGyslJ3knolv9+PdDp94tUfHgWegd/vx8LCghWuZFlGJBIBgKo5XkRERHT1MFj9JsMw0NnZiUKhYAWocs3r9SKXyzk/QkRERFcMjwJ/0/j4uO1XAHC5XAgEAhgdHa3oJCIioquKO1ZEREREDcIdKyIiIqIGYbAiIiIiahAGKyIiIqIG+T/rb1m2elYQQwAAAABJRU5ErkJggg==\" width=\"598\" height=\"46\"\u003e\u003c/p\u003e\n \u003cp\u003ewhere TP are the true positives, N is the total number of samples, and K is the number of classes. Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e presents the OA values and the corresponding median number of CSOs. Generally, the accuracies were high, ranging between 85% and 93%. The highest accuracy of 92.3% was achieved in 2018, and the lowest (85.3%) was in 2012. Accuracies above 90% were mostly achieved after 2015, when both Landsat and Sentinel images were available. The lowest median number of CSOs and the lowest OA were both recorded in 2012. The Pearson correlation coefficient (\u003cem\u003er\u003c/em\u003e) of 0.712 between the OA values and the median number of CSOs was statistically significant (p-value\u0026thinsp;=\u0026thinsp;0.006), highlighting the impact of the number of observations on classification accuracy.\u0026nbsp;\u003c/p\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eOverall accuracy per year. The median CSOs were computed based on the reference areas per year.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eYear\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOverall accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMedian number of CSOs\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2011\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e17\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2014\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2016\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2018\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e54\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e46\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e53\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2022\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e61\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003e5.2. Class accuracies\u003c/h2\u003e\n \u003cp\u003eTo assess the accuracy of each class per year, we used each confusion matrix from 2010 to 2022 as the basis to calculate the F1-score, which is the harmonic mean of the user\u0026rsquo;s accuracy (UA) and producer\u0026rsquo;s accuracy (PA):\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"609\" height=\"147\"\u003e\u003c/p\u003e\n \u003cp\u003ewhere TP, FP, and FN are the true positives, false positives, and false negatives, respectively. The F1-scores of each LU type corresponding to the years shown in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e are depicted in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eWhen averaged over time, the highest accuracies (\u0026ge;\u0026thinsp;90%) were obtained by classes occupying high shares of agricultural lands, like grassland, rapeseed, winter cereals, sugar beet, and maize. In comparison, classes such as fallow land and plantation, which occupy small shares of agricultural land, were predicted with the lowest accuracies (\u0026le;\u0026thinsp;52%). Similar to the OA results, the lowest F1-scores were mostly recorded in 2012, while the highest were mostly obtained after 2015.\u003c/p\u003e\n \u003cp\u003eThe lowest accuracy of grassland was 91.9% (2015), and the highest was 94.4% (2011). Rapeseed, winter cereals, sugar beet, and maize had accuracies greater than 85% in all years, with their lowest accuracies occurring in 2012. The highest accuracies of rapeseed (95.1%) and maize (92.3%) both occurred in 2016, while those of winter cereals (94.9%) and sugar beet (93.9%) both occurred in 2019. Hops had accuracies ranging from 74.8% (2011) to 92.4% (2020). Similar to hops, vineyards were at their lowest (57.5%) in 2011, but the highest (87.9%) was in 2019. In 2012, the potato class registered its lowest accuracy (42.0%), with its highest (92.2%) in 2020. Indeed, the potato class had the highest disparity between its lowest and highest accuracies. Summer cereals were detected at the lowest accuracy (65.0%) in 2011 and the highest (83.3%) in 2018. Legumes were classified with accuracies ranging from 48.7% (2012) and 86.7% (2021). Sunflower accuracies were between 65.7% (2012) and 83.3% (2017). The lowest accuracy of horticultural crops was 43.2% in 2011, and the highest was 71.6% in 2018. Plantation had accuracies ranging from 35.7% (2010) to 51.3% (2021). Overall, the lowest accuracies were recorded for fallow land, ranging from 12.9% (2012) to 39.5% (2022), making it the most difficult class to predict.\u003c/p\u003e\n \u003cp\u003eFor a one-shot overview of the confusion between the classes over time, we used the yearly confusion matrices to create a composite confusion matrix (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e). Among the best-performing classes, rapeseed and sugar beet achieved very high producers\u0026rsquo; accuracies (PA\u0026thinsp;=\u0026thinsp;96.4% and 94.2%, respectively), indicating low omission errors and strong separability from other crops. Winter cereals (PA\u0026thinsp;=\u0026thinsp;93.0%) and grassland (PA\u0026thinsp;=\u0026thinsp;92.4%) also exhibited high detection rates. Users\u0026rsquo; accuracies for these classes were similarly high (UA\u0026thinsp;\u0026gt;\u0026thinsp;91% for rapeseed, sugar beet, winter cereals, maize, potato, and grassland), demonstrating low commission errors and high thematic reliability of the mapped categories.\u003c/p\u003e\n \u003cp\u003eMaize, potato, sunflower, horticultural crops, and hops showed good to moderate producers\u0026rsquo; accuracies (PA between 86% and 89%), although sunflower exhibited a comparatively low user\u0026rsquo;s accuracy (UA\u0026thinsp;=\u0026thinsp;68.0%), indicating notable commission errors. Legumes, vineyards, summer cereals, and plantations reached moderate producers\u0026rsquo; accuracies (PA between 76% and 81%), suggesting increased omission errors, likely due to spectral and phenological similarities with other crop types.\u003c/p\u003e\n \u003cp\u003eFallow land was the weakest-performing class, with a low producer\u0026rsquo;s accuracy (PA\u0026thinsp;=\u0026thinsp;63.1%) and particularly low user\u0026rsquo;s accuracy (UA\u0026thinsp;=\u0026thinsp;19.5%). This indicates both substantial omission and commission errors and highlights the difficulty in distinguishing fallow land from grassland and certain arable crops. Similarly, plantation (UA\u0026thinsp;=\u0026thinsp;34.0%) and horticultural crops (UA\u0026thinsp;=\u0026thinsp;46.9%) exhibited low users\u0026rsquo; accuracies, suggesting that these mapped classes contain considerable proportions of misclassified pixels.\u003c/p\u003e\n \u003cp\u003eGenerally, the dominant misclassification patterns revealed three main confusion groups. First, mutual confusion within the broad cereal group (winter cereals, summer cereals, and maize) was evident. Second, a confusion within the semi-natural vegetation group was noticed, characterized by strong confusion between grassland and fallow land, and to a lesser extent, plantation. Third, confusion among intensively managed specialty crops was observed, particularly between potato, horticultural crops, and sugar beet. These patterns likely reflect overlapping phenological trajectories, similar canopy structures, and management practices that are difficult to disentangle using spectral time series alone.\u003c/p\u003e\n \u003cp\u003eIn summary, the classification performs robustly for major crop types with distinct phenological signatures. Classes characterized by heterogeneous management or transitional vegetation states (e.g., fallow land and plantation) remain challenging. However, those classes cover very small shares of agricultural land in Germany.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003e5.3. Comparison with official statistics\u003c/h2\u003e\n \u003cp\u003eTo further assess the plausibility of our maps, we compared our map areas with the corresponding areas in the official statistics \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. For the comparison, we calculated the Pearson correlation coefficient (\u003cem\u003er\u003c/em\u003e) and mean deviation (\u003cem\u003eMD\u003c/em\u003e) between our map areas and the official statistics. With the Pearson correlation coefficient, we measured the strength and direction of the linear relationship between our map areas and those of the official statistics. With \u003cem\u003eMD\u003c/em\u003e, we assessed the level of overestimation or underestimation in our map areas over time. Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e shows the comparison between the map areas (orange lines) and official statistics (blue lines) for the total agricultural area, followed by the most dominant LU classes. The classes have been ordered by their respective Pearson correlation coefficients.\u003c/p\u003e\n \u003cp\u003eRegarding the total agricultural area, a high positive correlation (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.757) was observed, with a tendency towards overestimation in the map areas for all years. This overestimation can be explained by the fact that in our agricultural LU maps, we map all agricultural areas in the open landscape, whereas the official statistics only account for land that is officially used for agricultural production.\u003c/p\u003e\n \u003cp\u003eEight classes (maize, sugar beet, summer cereals, rapeseed, grassland, winter cereals, sunflower, potato) showed positive correlations, while mapped legume areas did not correlate with the official statistics. The remaining classes (fallow land, horticultural crops) exhibited negative correlations, confirming the difficulty in accurately predicting them. Fallow lands were often mistaken for grasslands due to similar spectral profiles, thereby increasing uncertainty in their prediction. Horticultural crops, which are usually grown on fields barely larger than a Sentinel-2 pixel \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, often spectrally mix with adjacent LU classes, making their prediction difficult.\u003c/p\u003e\n \u003cp\u003eMaize and sugar beet had very high correlations (\u003cem\u003er\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.9), with maize being underestimated and sugar beet being overestimated. Summer cereals, rapeseed, grassland, winter cereals, and sunflower had high correlations (0.6\u0026thinsp;\u0026lt;\u0026thinsp;\u003cem\u003er\u003c/em\u003e\u0026thinsp;\u0026le;\u0026thinsp;0.83), and all of them were overestimated. Potatoes, which were underestimated, had a moderate correlation (0.33). Legumes were overestimated and exhibited a weak relationship (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.001). Although both fallow land and horticultural crops obtained negative correlations, fallow land tended to be underestimated, while horticultural crops were overestimated. Overall, sugar beet obtained the smallest deviation from the official statistics over time.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eUsage Notes\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eThe provided LU maps in the COG format enable data hosting in Hypertext Transfer Protocol (HTTP) servers and efficient data handling in cloud computing environments. Together with the provided color table, the maps can be visualized in various geospatial software, such as QGIS and ArcGIS.\u003c/p\u003e\n \u003cp\u003eThe distinct advantage of our LU map time series is the ability to facilitate spatially explicit analysis of agricultural LU changes that cannot be achieved with aggregated data, such as agricultural statistics. However, the statistics extracted from our maps are not meant to replace the official agricultural statistics. The LU maps could be used to spatially and temporally disaggregate the agricultural statistics.\u003c/p\u003e\n \u003cp\u003eGiven the significant correlation between the number of CSOs and OAs, the main uncertainty in our products could be tied to the number of CSOs per pixel. Therefore, when using our maps for further analysis, uncertainty modelling ought to be considered. For each LU map, an uncertainty map could be derived by calculating the inverse of the number of CSOs per pixel.\u003c/p\u003e\n \u003cp\u003eWhile the maps provided in this study are from 2017 to 2023, work is ongoing to update the maps up to the present. This will be done in a rolling manner as follows: existing reference samples will be refined, new samples with a longer time series will be generated, the models will be updated, and then applied to all years from 1990 up to the present. Based on the translation table provided in Table S3, the maps provided in this study could be integrated with the more current and detailed maps of Tetteh et al.\u003csup\u003e47,48\u003c/sup\u003e.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eCompeting Interests\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\u003ch2\u003eAdditional Information\u003c/h2\u003e \u003cp\u003eA supplementary file is attached to this data descriptor and also available at this weblink (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://1drv.ms/b/c/2ce41295938fadfa/IQCkK9EqF2ndRZRSYgSeNF1KAchORNChpfwPA_91y71SgYw?e=krV3zG\u003c/span\u003e\u003cspan address=\"https://1drv.ms/b/c/2ce41295938fadfa/IQCkK9EqF2ndRZRSYgSeNF1KAchORNChpfwPA_91y71SgYw?e=krV3zG\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e\u003ch2\u003eAuthor Contributions\u003c/h2\u003e \u003cp\u003eAll authors contributed to the final version of the manuscript through discussion and editing. G.O.T. collected training and validation samples, trained and validated the prediction model, applied the trained model to generate the maps, postprocessed the maps, analysed the results, and wrote the original draft; M.S. and S.E. conceptualized the study and supervised the research. M.S. preprocessed the Landsat and Sentinel images. V.-D.P. implemented the foundational code for model training and prediction.\u003c/p\u003e\u003ch2\u003eAcknowledgments\u003c/h2\u003e \u003cp\u003e This research was conducted within the framework of the Nationwide Monitoring Program of Biodiversity in Agricultural Landscapes (MonViA) project, which is under the auspices of the German Federal Ministry of Agriculture, Food, and Regional Identity (BMELH). Additionally, part of the work shown here was financially supported by funds for climate reporting and projection at the Th\u0026uuml;nen Institute. The BMELH has transferred climate reporting as a permanent, sovereign task to the Th\u0026uuml;nen Institute, as an official ongoing mandate, to fulfil the legal requirements of the Federal Climate Protection Act and international treaties. Vu-Dong Pham acknowledges support by the research project Fragmented Transformations (German Federal Ministry of Research, Technology, and Space; FKZ 01UC2102).\u003c/p\u003e\u003ch2\u003eCode Availability\u003c/h2\u003e \u003cp\u003eThe code used in this study to create the LU maps is proprietary and not publicly available. Generating and analyzing the maps was done in Python 3.10.12 using several libraries, including NumPy 1.26.4, GDAL 3.6.2, Rasterio 1.4.3, and Tensorflow 2.20.0. The preprocessing of the Landsat and Sentinel images to create the ARD images was done with FORCE (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/davidfrantz/force\u003c/span\u003e\u003cspan address=\"https://github.com/davidfrantz/force\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), and the visualization of the preprocessed images and predicted LU maps was done in QGIS 3.4.0 LTR.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKirschke D, H\u0026auml;ger A, Schmid JC (2021) New Trends and Drivers for Agricultural Land Use in Germany. In: Weith T et al (eds) Sustainable Land Management in a European Context: A Co-Design Approach. Springer International Publishing, Cham, pp 39\u0026ndash;61. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-030-50841-8_3\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-50841-8_3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL\u0026uuml;ker-Jans N, Simmering D, Otte A (2016) Analysing Data of the Integrated Administration and Control System (IACS) to Detect Patterns of Agricultural Land-Use Change at Municipality Level. Landsc Online 48\u0026ndash;48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3097/LO.201648\u003c/span\u003e\u003cspan address=\"10.3097/LO.201648\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeyer MA, Fr\u0026uuml;h-M\u0026uuml;ller A (2020) Patterns and drivers of recent agricultural land-use change in Southern Germany. Land Use Policy 99:104959\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNiedertscheider M, Kuemmerle T, M\u0026uuml;ller D, Erb K-H (2014) Exploring the effects of drastic institutional and socio-economic changes on land system dynamics in Germany between 1883 and 2007. Glob Environ Change 28:98\u0026ndash;108\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStein S, Steinmann H-H, Isselstein J (2019) Linking Arable Crop Occurrence with Site Conditions by the Use of Highly Resolved Spatial Data. Land 8:65\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAtzberger C (2013) Advances in Remote Sensing of Agriculture: Context Description, Existing Operational Monitoring Systems and Major Information Needs. Remote Sens 5:949\u0026ndash;981\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeiss M, Jacob F, Duveiller G (2020) Remote sensing for agricultural applications: A meta-review. Remote Sens Environ 236:111402\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDonmez E, Heckelei T, Storm H (2025) Satellite remote sensing-based crop cover classification over Europe: accuracy of different methodological approaches. Int J Remote Sens 46:8251\u0026ndash;8294\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang C et al (2025) Remote sensing for crop mapping: A perspective on current and future crop-specific land cover data products. Remote Sens Environ 330:114995\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang J et al (2026) Progress and Perspectives of Crop Type Mapping With Remote Sensing: A review. IEEE Geosci Remote Sens Mag 2\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/MGRS.2025.3648119\u003c/span\u003e\u003cspan address=\"10.1109/MGRS.2025.3648119\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsam S et al (2022) Mapping Crop Types of Germany by Combining Temporal Statistical Metrics of Sentinel-1 and Sentinel-2 Time Series with LPIS Data. Remote Sens 14:2981\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBlickensd\u0026ouml;rfer L et al (2022) Mapping of crop types and crop sequences with combined time series of Sentinel-1, Sentinel-2 and Landsat 8 data for Germany. Remote Sens Environ 269:112831\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGessner U, Hirner A, Asam S, Wenzl M, Kuenzer C (2025) Combining Machine Learning and Spatiotemporal Filtering to Map Crop Types of Germany for Seven Years. IEEE Geosci Remote Sens Lett 22:1\u0026ndash;5\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGriffiths P, Nendel C, Hostert P (2019) Intra-annual reflectance composites from Sentinel-2 and Landsat for national-scale crop and land cover mapping. Remote Sens Environ 220:135\u0026ndash;151\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePreidl S, Lange M, Doktor D (2020) Introducing APiC for regionalised land cover mapping on the national scale using Sentinel-2A imagery. Remote Sens Environ 240:111673\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchwieder M, Tetteh GO, Blickensd\u0026ouml;rfer L, Gocht A, Erasmi S (2024) Agricultural land use (raster): National-scale crop type maps for Germany from combined time series of Sentinel-1, Sentinel-2 and Landsat data (2017 to 2021). Zenodo \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.10640528\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.10640528\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTetteh GO, Schwieder M, Blickensd\u0026ouml;rfer L, Gocht A, Erasmi S (2024) Agricultural land use (vector): National-scale crop type maps for Germany from combined time series of Sentinel-1, Sentinel-2 and Landsat data (2017 to 2021). Zenodo \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.10619783\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.10619783\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGerman Environment Agency. National Inventory Report for the German Greenhouse Gas Inventory 1990\u0026ndash;2004 (2009)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLaamrani A, Voroney PR, Gillespie AW, Chehbouni A (2021) Development of a Land Use Carbon Inventory for Agricultural Soils in the Canadian Province of Ontario. Land 10:765\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLokupitiya E, Paustian K (2006) Agricultural Soil Greenhouse Gas Emissions. J Environ Qual 35:1413\u0026ndash;1427\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X, Yu L, Du Z, Liu X (2025) Crop Statistic to Annual Map: Tracking spatiotemporal dynamics of crop-specific areas through machine learning and statistics disaggregating. Sci Data 12:1249\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLamsal G, Marston LT (2024) HarvestGRID: high-resolution harvested crop areas of the United States from 1981 to 2019. Environ Res Food Syst 2:015006\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrank C et al (2024) Woody semi-natural habitats modulate the effects of field size and functional crop diversity on farmland birds. J Appl Ecol 61:987\u0026ndash;999\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKasiske T et al (2025) Identifying typical patterns of land-use and landscape structure in citizen science butterfly monitoring. Ecol Indic 180:114317\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ\u0026auml;nicke C et al (2022) Field-level land-use data reveal heterogeneous crop sequences with distinct regional differences in Germany. Eur J Agron 141:126632\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStein S, Steinmann H-H (2018) Identifying crop rotation practice by the typification of crop sequence patterns for arable farming systems \u0026ndash; A case study from Central Europe. Eur J Agron 92:30\u0026ndash;40\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eder Velde M et al (2024) Policy relevant crop diversity monitoring based on earth observation and farmers\u0026rsquo; declarations. Environ Res Lett 20:014054\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRousi M et al (2021) Semantically Enriched Crop Type Classification and Linked Earth Observation Data to Support the Common Agricultural Policy Monitoring. IEEE J Sel Top Appl Earth Obs Remote Sens 14:529\u0026ndash;552\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHendricks NP, Smith A, Sumner DA (2014) Crop Supply Dynamics and the Illusion of Partial Adjustment. Am J Agric Econ 96:1469\u0026ndash;1491\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, Roy DP (2017) A Global Analysis of Sentinel-2A, Sentinel-2B and Landsat-8 Data Revisit Intervals and Implications for Terrestrial Monitoring. Remote Sens 9:902\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClaverie M et al (2018) The Harmonized Landsat and Sentinel-2 surface reflectance data set. Remote Sens Environ 219:145\u0026ndash;161\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFoerster S, Kaden K, Foerster M, Itzerott S (2012) Crop type mapping using spectral\u0026ndash;temporal profiles and phenological information. Comput Electron Agric 89:30\u0026ndash;40\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAzar R et al (2016) Assessing in-season crop classification performance using satellite data: a test case in Northern Italy. Eur J Remote Sens 49:361\u0026ndash;380\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ\u0026auml;nicke C, Petersen KA, Schmidts P, M\u0026uuml;ller D, Jepsen MR (2025) Field and farm-level data on agricultural land use for the European Union. Sci Data 12:1050\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchneider M, Schelte T, Schmitz F, K\u0026ouml;rner M (2023) EuroCrops: The Largest Harmonized Open Crop Dataset Across the European Union. Sci Data 10:612\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePham V-D et al (2024) Temporally transferable crop mapping with temporal encoding and deep learning augmentations. Int J Appl Earth Obs Geoinf 129:103867\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrantz D (2019) FORCE\u0026mdash;Landsat\u0026thinsp;+\u0026thinsp;Sentinel-2 Analysis Ready Data and Beyond. Remote Sens 11:1124\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBuchner J et al (2020) Land-cover change in the Caucasus Mountains since 1987 based on the topographic correction of multi-temporal Landsat composites. Remote Sens Environ 248:111967\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrantz D, Roder A, Stellmes M, Hill J (2016) An Operational Radiometric Landsat Preprocessing Framework for Large-Area Time Series Applications. IEEE Trans Geosci Remote Sens 54:3928\u0026ndash;3943\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoy DP, Li Z, Zhang HK (2017) Adjustment of Sentinel-2 Multi-Spectral Instrument (MSI) Red-Edge Band Reflectance to Nadir BRDF Adjusted Reflectance (NBAR) and Quantification of Red-Edge Band BRDF Effects. Remote Sens 9:1325\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrantz D, Ha\u0026szlig; E, Uhl A, Stoffels J, Hill J (2018) Improvement of the Fmask algorithm for Sentinel-2 images: Separating clouds from bright surfaces based on parallax effects. Remote Sens Environ 215:471\u0026ndash;481\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu Z, Wang S, Woodcock CE (2015) Improvement and expansion of the Fmask algorithm: cloud, cloud shadow, and snow detection for Landsats 4\u0026ndash;7, 8, and Sentinel 2 images. Remote Sens Environ 159:269\u0026ndash;277\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu Z, Woodcock CE (2012) Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens Environ 118:83\u0026ndash;94\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePham V-D et al (2024) An annual land cover dataset for the Baltic Sea Region with crop types and peat bogs at 30 m from 2000 to 2022. Sci Data 11:1242\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGensior A, Drexler S, Fu\u0026szlig; R, St\u0026uuml;mer W, R\u0026uuml;ter S (2025) Emissions of greenhouse gases from land use, land-use change and forestry (LULUCF). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.thuenen.de/en/thuenen-topics/climate-and-air/emission-inventories-accounting-for-climate-protection/treibhausgas-emissionen-durch-landnutzung-landnutzungsaenderung-und-forstwirtschaft-lulucf\u003c/span\u003e\u003cspan address=\"https://www.thuenen.de/en/thuenen-topics/climate-and-air/emission-inventories-accounting-for-climate-protection/treibhausgas-emissionen-durch-landnutzung-landnutzungsaenderung-und-forstwirtschaft-lulucf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFederal Statistical Office (2025) The database of the Federal Statistical Office. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www-genesis.destatis.de/datenbank/online\u003c/span\u003e\u003cspan address=\"https://www-genesis.destatis.de/datenbank/online\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTetteh GO, Schwieder M, Blickensd\u0026ouml;rfer L, Gocht A, Erasmi S (2024) Agricultural land use (raster): National-scale crop type maps for Germany from combined time series of Sentinel-2 and Landsat data Zenodo \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17197830\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17197830\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTetteh GO, Schwieder M, Blickensd\u0026ouml;rfer L, Gocht A, Erasmi S (2025) Agricultural land use (raster): National-scale crop type maps for Germany from combined time series of Sentinel-2 and Landsat data Zenodo \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17197871\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17197871\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Thünen Institute of Farm Economics, Bundesallee 63, 38116 Braunschweig, Germany","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Agricultural lands, Long-term monitoring, Remote sensing, Time series analysis, Deep Learning, Land-use classification, Long-term dataset","lastPublishedDoi":"10.21203/rs.3.rs-9074257/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9074257/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWe present nationwide annual agricultural land-use maps of Germany from 1990 to 2023, created from Landsat and Sentinel-2 images using a deep learning approach. Based on farmers\u0026rsquo; parcel-level declarations from 2006 to 2022, we extracted annual training samples for 13 crop classes and one grassland class. These samples were used to train a multi-year one-dimensional convolutional neural network, which was subsequently applied to generate the annual land-use maps. Overall map accuracies ranged between 85% and 93%. Dominant classes such as grassland, rapeseed, winter cereals, sugar beet, and maize were detected with high accuracy (\u0026ge;\u0026thinsp;90%). Conversely, minor classes such as fallow land and plantation were predicted with low accuracy (\u0026le;\u0026thinsp;52%). Comparison of map areas with agricultural statistics over the entire study period revealed high correlations for most classes, particularly maize (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.978). The presented maps provide an essential basis for analyzing long-term trends in agricultural land-use. They can be used to fill temporal gaps in national agricultural statistics and to disaggregate those statistics to higher spatial units.\u003c/p\u003e","manuscriptTitle":"Nationwide annual agricultural land-use maps of Germany from 1990 to 2023 derived from satellite imagery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-10 17:01:59","doi":"10.21203/rs.3.rs-9074257/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ede1caa3-15e0-4339-aaeb-38451d6148df","owner":[],"postedDate":"March 10th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":64190562,"name":"Artificial Intelligence and Machine Learning"},{"id":64190563,"name":"Geographic Information Systems"},{"id":64190564,"name":"Agricultural Economics \u0026 Policy"},{"id":64190565,"name":"Environmental Policy"}],"tags":[],"updatedAt":"2026-03-11T20:08:56+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-10 17:01:59","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9074257","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9074257","identity":"rs-9074257","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}