To identify cropland, grassland and shrubland, we used spectral information retrieved from all Sentinel-2 pixels available for Switzerland between April–November of 2017–2019 with initial cloud coverage lower than 80 % (as reported in the Sentinel-2 metadata). We used the Sentinel-2 imagery preprocessed to surface reflectance (Level 2A) within GEE (GEE repository ee.ImageCollection(“COPERNICUS/S2_SR”)) and applied a cloud masking procedure (GEE repository ee.ImageCollection(“COPERNICUS/S2_CLOUD_PROBABILITY”). This procedure uses deep learning semantic segmentation (Garcia-Garcia et al., 2017) to set the cloud occurrence probabilities within Sentinel-2 scenes (Zupanc, 2017) and has been documented to outperform other cloud masking procedures. Additional scripts were used to mask shadows and snow coverage (see the linked GEE codes for more information). After preprocessing, the number of images available within the time period varied substantially over Switzerland due to cloudiness, especially in the Alps, and the acquisition pathways of the Sentinel-2 satellite sensors (Fig. 2).

To identify the spectral parameters of cropland and grassland, we used parcel-level training data indicating the occurrence of different crop types and permanent grassland derived by cantonal (state) authorities based on the reporting of farmers in the years 2017–2019. We have considered parcels with reported crop usage (e.g. wheat, spelt, corn, potatoes, sugar beets) within a reported year to be cropland. Due to the practise of crop rotation which is frequently applied on Swiss agricultural parcels, as well as different crop type within an individual field, it is likely that the spectral characteristics of a given single parcel changed within the 3-year study period. However, the training data for each parcel are for one time point only within the study period, and therefore such temporal variations were not considered within the training data. Cropland is considered to be an area covered by crops within a given year for particular training data, while grasslands are considered to be areas explicitly reported as permanent grasslands. Permanent grasslands are likely to have similar phenological trajectories over the multiple growing seasons considered within the study period. Parcels classified as annual grasslands within the reported statistics were considered to be cropland because we considered annual grassland to be a part of the crop rotation management.

The parcel-level training polygons were selected manually from different areas across Switzerland. The overall requirement for selection of a polygon was its visual homogeneity (e.g. no trees within the cropland or grassland fields). We also considered the variability of croplands and grasslands by including parcels that reported different land use within those two classes (e.g. different cropland types). To avoid selecting pixels of mixed land use, which are likely to occur in heterogeneous landscapes, especially at parcel edges, we shrunk the training polygons by a 10 m “inside” buffer distance (buffer size defined according to the resolution of Sentinel-2 bands). In total, 1378 polygons covered by more than 400 000 10 m pixels were selected to train the classification model.

In addition to the cropland and grassland data, we provided the random forest (RF) model with training data on shrubland, forest, wetland and water surfaces. These training data allowed us to better define the ranges of reflectance of different LC in the training data. In the accuracy assessment we only include the shrubland as it is a widespread LC class in Swiss mountainous areas and, due to its vegetation cover with minimal seasonal variation in reflectance, is likely to be considered grassland by our predictive models. Other LC classes were further masked out using different maps and were not an object of accuracy testing. Training data on different LC classes were generated from 70, 25 and 13 polygons with an average area of 11, 30 and 1177 ha for shrubland, wetland and water surfaces, respectively. These polygons were digitized manually from the Sentinel-2 imagery guided by the Swiss Land Use Statistics.

We extracted multiple indices for every Sentinel-2 pixel overlapping the training polygons. The indices used were chosen based on previous mapping of Swiss agricultural areas (Kolecka et al., 2018; Pazúr et al., 2021) and followed general assumptions about the differences between the phenology of cropland and grassland over a year (e.g. low growth on cropland in autumn) and low variability of permanent grassland over time. The selected indices characterize phenology either over the whole growing season or over a particular part of the growing season, or they relate to summary statistics of the spectral bands of Sentinel-2 (Table 1). Similar indices have been found useful for land cover classification over large areas (Pflugmacher et al., 2019) and mapping phenological responses in grassland areas or cropland types (Ghazaryan et al., 2018; Gómez Giménez et al., 2017). Furthermore, we included variables characterizing the terrain properties of each pixel, such as elevation, slope and terrain orientation retrieved from the 30 m Shuttle Radar Topography Mission data (GEE repository ee.ImageCollection (“USGS/SRTMGL1_003”) (Farr et al., 2007).

To account for different environmental conditions (i.e. climate and terrain) within Switzerland, the classification models were calculated separately for two strata defined by biogeographic regions. The biogeographical regions of the Jura Mountains and the Swiss Plateau together form one stratum (regions 3 and 4 on Fig. 1), while the Alpine regions form a second stratum (regions 1, 2, 5 and 6 on Fig. 1). While we tested several approaches to defining the strata, including a single national model and different combinations of the biogeographical regions, we found that the two strata chosen gave the best results and most straightforward approach to taking into account the climatic and topographical differences between the more lowland areas and the mountainous areas. In particular the whole of Switzerland approach resulted in biases for crop detection towards the Plateau and higher inaccuracies in mountainous areas, particularly in mixed shrubland grassland areas.

We used the random forest (RF) classifier implemented within the GEE platform (library ee.Classifier.smileRandomForest) for the classification of cropland, grassland and shrubland. The default parameters of the GEE algorithm were used except for the number of trees (300 trees were used) and the number of variables per split (10 variables per split were used). These variables were set according to the accuracy cross-validation checks of different combinations of their values (Kuhn, 2008). We also produced a separate layer of non-vegetated areas, i.e. sealed surfaces, rocks and water bodies for masking purposes. Assuming that those surfaces have little or no phenological response, we classified these areas using the 95th percentile of the distribution of all normalized difference vegetation index (NDVI) values recorded over the mapping period. Using two different thresholds, we found the non-vegetated areas under 2000 m a.s.l. on those pixels where the 95th percentile of recorded NDVI value was lower than 0.7. In areas above 2000 m a.s.l., the non-vegetated areas occurred where the 95th percentile of recorded NDVI value was lower than 0.6 (Fig. 3).

To increase the accuracy of the final map we masked out woody vegetation (forest, tree lines and single trees) from the model output with the vegetation height model (VHM) of Switzerland (Ginzler and Hobi, 2015). Within areas classified as cropland, woody vegetation was masked out using a vegetation height threshold of 3 m. This was done to prevent the exclusion of crops that potentially reach 3 m in height, especially towards the end of the growing season (the period when the imagery on which the VHM is based was captured). Within grasslands, a lower threshold of 1 m was applied in order to mask out vineyards, shrubs and hedgerows. In addition, we converted all cropland areas smaller than 2000 m2 to grassland as we considered this a minimum size for cropland parcels. This allowed us to eliminate the effects of misclassification of LC on mixed pixel areas, such as in settlement hinterland or at high elevations, where small patches of grassland may appear within areas of low vegetation cover, e.g. in the transition zone between grassland and rocks. As it is likely that all the areas of forest were masked out using the VHM, we converted all the unmasked pixels of forest to shrubland as we found large uncertainties of distinguishing between those two classes within our model.

Map accuracy was assessed against three separate testing datasets: the “parcel-level+” testing dataset, derived from parcel-level data and digitized polygons of other LC classes derived in a similar manner to that described in Sect. 2.2 (Training data section), and two independent datasets, the ALL-EMA (Agricultural Species and Habitats' Monitoring Programme) (Riedel et al., 2018) testing dataset and the Swiss Land Use Statistics testing dataset, described further below. Since these datasets were produced independently and for differing purposes, each dataset has differences in extent, nomenclature and ground-truth information.

The distribution of the values of the indices derived from Sentinel-2 and the elevation model within the generated samples indicated differences between cropland, grassland and shrubland (Fig. 4). Substantial differences between these LC classes were found for individual spectral bands (NIR, SWIR1, SWIR2), phenological indices (NDVI, EVI and its derivatives) and terrain indices (elevation, slope).

Using those indices, we found better separability of cropland and grassland for the Jura and Plateau stratum compared to the Alps. Indices that characterize phenology over the growing season generally have higher values on grassland areas than cropland areas except for the standard deviation of NDVI, which is lower on grassland. For the Jura and Plateau stratum the standard deviation of NDVI was found to be one of the most important indices in the classification model (Fig. 5). Grasslands within both strata were also characterized by a negative skewness of the distribution of NDVI values. This relatively high proportion of higher NDVI values reflects the higher greenness of these areas. For shrubland, the values of these indices were within a very narrow span of the value distribution. This suggests the phenological behaviour of shrubland over the growing seasons is less variable than that of cropland or grassland. Good separability between cropland and grassland was also possible through the phenological indices ndvi_pc_05 and ndvi_pc_25, which represent the lower percentiles of the distribution of NDVI recorded through all 3 years of Sentinel-2 data. Especially for the Jura and Plateau stratum, both indices were much higher on grassland than on cropland and highly contribute to the accuracy of the resulting classification model. Within the samples generated for the Jura and Plateau stratum, using the single phenological indices, such as the ndvi_pc_05, may successfully separate cropland and grassland classes. By ensuring the robustness of such metrics for example by including multiple growing seasons, the model generated for the Jura and Plateau stratum may be applicable on flat and less complex terrains elsewhere.

However, we observed a wide range in spring seasonality (sSpring), especially on Alpine grasslands. This observation may be due to the low phenology of the alpine grassland at the beginning of the growing season or snow artefacts, which were missed from the snow mask applied in the classification especially during the snow melting period. Elevation also substantially aided the separation of grasslands from cropland and the classification accuracy, particularly within the Alps stratum. This observation can be considered relevant to mountainous areas, where cropland areas are generally only feasible at comparably lower elevations. Generally, the complex terrain conditions of the Alps, described by the terrain parameters, substantially influenced the separability of the cropland and grassland classes which might limit the transferability of the model outside of the Swiss Alps.

Comparison between the derived map and the selected test datasets showed high accuracy of the modelling outputs (Fig. 6). The comparison of the output map with the test data derived from the selected parcel-level+ polygons resulted in the highest overall accuracies, 0.93 for the Jura and Plateau region and 0.92 for the Alps region. This accuracy assessment (parcel-level+ data) was also based on samples that were the most equally distributed across the different classes (prevalence ratios). The lowest accuracies were found for the comparison to the ALL-EMA dataset in the Alps regions for the presence of cropland and shrubland, 0.44 and 0.31 (sensitivity ratio), and absence of grasslands, 0.38 (specificity ratio). The prevalence rates, however, demonstrate that these results relate to a limited number of samples. Due to low occurrences and a large number of samples, the prediction of absence of cropland in the Alps was of highest accuracy (0.99, specificity ratio) compared to the ALL-EMA testing data. The accuracy assessment of the shrubland mapping resulted in similarly high accuracy rates for both the parcel-level+ data and the Swiss Land Use Statistics. These similarities are to be expected since the parcel-level+ training data for shrubs were selected from a subset of the Swiss Land Use Statistics. These accuracy assessments only consider areas of overlap between the output map and cropland, grassland, and shrubland as defined by the given testing dataset. Potential bias of the output map outside of this extent, e.g. occurrence of cropland, grasslands and shrublands in areas of woody vegetation, lakes or non-vegetated areas, is not considered.

Furthermore, we assessed the nationwide accuracy of cropland and grassland mapping using all Swiss Land Use Statistics interpretation points (n= 4 128 498). This accuracy assessment resulted in high overall agreement with our classification at 85 % with a user's accuracy of 88 % and 67 % and a producer's accuracy of 72 % and 81 % for cropland and grassland, respectively (Table 2). The lower accuracy may be due to nomenclature differences, which mean that the definitions of cropland and grassland are not fully aligned (e.g. agricultural crop rotation between crop and grassland on the same parcel). Moreover, the Swiss Land Use Statistics are determined from aerial imagery captured over an extensive time span which, depending on location does not match with the period of our mapping data and the spatial resolution of the output map (10 m). By summarizing the spatial distribution pattern of agreement/mismatch within 1 km2 blocks (Fig. 7), a higher degree of mismatch was found in mountainous and sparsely vegetated areas, as well as in wine-making areas. By comparing the cropland and grassland with the Swiss Land Use Statistics, our mapping outcome performed well also in comparison with different land cover mapping products with global coverage (Fig. 8). Specifically, in the case of accuracies of cropland and grassland, the mapping output outperformed the Copernicus Global Land Cover 100 m (PROBA, Buchhorn et al., 2021), Global Land Cover Map in 30 m resolution (GLC30, Zhang et al., 2021) and Global Land Cover Map in 10 m resolution (GLC10, Gong et al., 2019).