Mapping management intensity types in grasslands with synergistic use of Sentinel-1 and Sentinel-2 satellite images

Abstract

Grasslands, being vital ecosystems with significant ecological and socio-economic importance, have been the subject of increasing attention due to their role in biodiversity conservation, carbon sequestration, and agricultural productivity. However, accurately classifying grassland management intensity, namely extensive and intensive practices, remains challenging, especially across large spatial extents. This research article presents a comprehensive investigation into the classification of grassland management intensity in two distinct regions of Poland, NUTS2 - namely Podlaskie (PL84) and Wielkopolskie (PL41), by integrating data from Sentinel-1 and Sentinel-2 satellite imagery. The study leverages the unique capabilities of Sentinel-1, a radar satellite, and Sentinel-2, an optical multispectral satellite, to overcome the limitations of using a single data source. Preprocessed Sentinel-1 and Sentinel-2 data were combined to extract spectral and textural features, providing valuable insights into grassland characteristics and patterns. Supervised classification using the Random Forest algorithm was used, and ground truth data from field surveys facilitated the creation of training samples. In Podlaskie, extensive grasslands achieved an overall accuracy (OA) of 84%, while intensive grasslands attained an OA of 83%. In Wielkopolskie, extensive grasslands exhibited an OA of 84%, while intensive grasslands achieved an OA of 83%. Additionally, the classification metrics, including user’s accuracy (UA), F1 score, and producer’s accuracy (PA), further highlighted the variations in classification accuracy. This comprehensive mapping of grassland management intensity using combined Sentinel-1 and Sentinel-2 data provides valuable insights for conservation agencies, agricultural stakeholders, and land managers. The study’s findings contribute to sustainable land management and decision-making processes, facilitating the identification of ecologically valuable areas, optimizing agricultural productivity, and assessing the impacts of different management strategies. Furthermore, the research highlights the potential of Sentinel missions for grassland monitoring and emphasizes the importance of advanced remote sensing techniques for understanding and preserving these crucial ecosystems.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-83699-4.

Introduction

Grasslands, characterized by an abundance of grasses and herbaceous plants, are among the most valuable ecosystems on Earth. They cover a significant portion of the global land area and play a crucial role in supporting biodiversity, carbon sequestration, soil conservation, and providing various ecosystem services¹. Grasslands are not only ecologically significant but also have immense socio-economic importance, serving as a source of forage for herbivorous livestock, contributing to agricultural productivity, and providing recreational spaces for human activities². However, grasslands are subject to various management practices, which can broadly be classified as extensive and intensive^3,4. Extensive grassland management involves low-intensity grazing or mowing, allowing natural vegetation to persist and promoting biodiversity conservation⁵. These ecosystems have not been substantially modified by fertilization, liming, drainage, herbicide use and introduction of partial or completely renovation of grassland sod. On the other hand, intensive grassland management involves higher stocking densities, frequent mowing, and fertilization, using periodical renovation, aiming to optimize forage production and meet the increasing demands of herbivorous livestock production^6,7. Recent research trends in remote sensing for grassland monitoring^8–10 emphasize enhancing both intra- and inter-seasonal time series for grassland classification^11,12, analyzing plant stress due to drought^13–15, and estimating grasslands aboveground biomass^16,17.

Understanding and mapping management intensity are essential for effective land management, conservation planning, and sustainable agricultural practices^18,19. However, information on grassland management intensity remains limited^20,21. Traditionally, accurately mapping grassland management intensity has been challenging due to the vast spatial extent of grasslands and the limitations of field-based surveys. Recent advancements in remote sensing technology²² and the increased availability of satellite imagery-based programs for determining grasslands such as Copernicus CORINE Land Cover (CLC) and High Resulution Layers (HRL) Grassland have created new opportunities for comprehensive and efficient monitoring of grassland ecosystems. Whilts in Poland the reliability of HRL and CLC data for accurately identifying grasslands using remote sensing techniques has been questioned due to the inconsistent classification criteria they apply^23–26. HRL, developed under the GMES Initial Operations (GIO) Land Monitoring Service, provide detailed land cover data with high spatial resolution. They define grasslands based on physical characteristics, specifically areas dominated by grasses and herbaceous plants. However, this broad definition presents challenges in accurately identifying and distinguishing grasslands across various landscapes. For example, HRL classify grasslands under a wide range of conditions, including pastures, natural grasslands, grass-covered areas within transitional woodlands, alpine meadows, and regions with scattered trees or shrubs. This categorization overlaps with multiple CLC classes (such as 231 for pastures, 242 for complex cultivation patterns, 321 for natural grasslands, and others within the 211 to 244 range), making it difficult to achieve consistency when integrating HRL with CLC data^25,26. The inclusion of complex landscapes like Dehesas, orchards, and areas with dominant grassy cover further complicates the classification process, as these landscapes may not be easily distinguishable through remote sensing techniques. Additionally, relying solely on physical characteristics without considering the ecological roles of grasslands can lead to classification errors, especially in mixed-use landscapes where grasslands are interspersed with other land cover types. The broad and sometimes overlapping definitions in both HRL and CLC data create discrepancies in grassland mapping and monitoring. As a result, while HRL and CLC data are valuable tools for large-scale land cover analysis, their effectiveness in accurately determining grasslands is limited by these classification inconsistencies, raising concerns about the accuracy and precision of grassland monitoring using remote sensing.

Currently, involving satellite radar and optical data together plays a significant role in remote sensing research, especially when the data originates from the same satellite program, such as Copernicus Sentinel. The synergistic use of Sentinel-1 (S-1) and Sentinel-2 (S-2) satellites has become highly popular and widely used for land cover classification^27,28, including forests^15,29, grasslands^30,31, and arable lands^32,33, due to their complementary capabilities and comprehensive coverage. S-1 with its Synthetic Aperture Radar (SAR) technology, provides all-weather, day-and-night imaging, which is crucial for monitoring land cover in various environmental conditions. This radar-based system can penetrate cloud cover and provide consistent observations regardless of weather conditions or lighting, making it invaluable for detecting changes in land cover and assessing the structure of vegetation. On the other hand, S-2 offers high-resolution optical imagery from both S-2 A and S2B, that captures detailed information on vegetation health, land use, and seasonal variations. Its Multi-Spectral Instruments (MSI) provide data across several spectral bands, allowing for precise differentiation between various types of vegetation and land use types. This optical data is particularly useful for assessing vegetation types, monitoring crop conditions, and analyzing land cover changes over time. The high spatial resolution of S-2 imagery enhances the ability to distinguish between different land cover types, making it a powerful tool for detailed land cover classification. Additionally, the integration of these datasets supports better temporal analysis and monitoring, as S-1 frequent revisit times and S-2 detailed spectral information enhance the ability to track and analyze land cover dynamics effectively.

The integration of S-1 and S-2 data in Poland faces several challenges, primarily due to the country’s variable weather conditions and technical limitations. Poland experiences significant cloud cover, averaging 180 days per year³⁴, which complicates the use of Sentinel-2 optical imagery, as it requires clear skies for accurate data collection. This frequent cloud cover limits the availability of high-quality optical images, necessitating the use of scenes with minimal cloud coverage to ensure data reliability. Additionally, the recent termination of the Sentinel-1B mission due to system failure has further exacerbated the situation. This has reduced the frequency of radar data acquisition to one scene every ten days from Sentinel-1 A, which may not always align well with S-2 optical data, affecting the temporal resolution and synchronization of the combined datasets. Together, these factors can hinder the effectiveness of efforts to combine data, impacting the accuracy and consistency of land cover classification and monitoring tasks in Poland.

In this study, we propose an innovative approach by utilizing the synergistic use of S-1 and S-2 satellite imagery to classify grassland management intensity in Poland. The Random Forest (RF) classifier offers a significant improvement over traditional classifications based on CLC and HRL, which often struggle to capture the detailed variability of grassland types. By integrating the radar capabilities of S-1 with the optical imagery of S-2, we effectively overcome the limitations posed by frequent cloud cover in Poland and achieve a more precise classification of grassland management practices. Our approach stands out for its ability to accurately differentiate between extensive and intensive management practices. This is especially important given the diverse structure and layout of grasslands in Poland, which have been historically shaped by State Agricultural Farms (PGR) practices and agricultural systems from the communist era³. These historical practices have created a complex pattern of grassland types and management regimes, making accurate classification challenging. By accounting for both historical and current land use patterns, our method provides a more comprehensive analysis, supporting better decision-making for sustainable grassland management, biodiversity conservation, and the preservation of ecosystem function.

Materials and methods

Figure 1 illustrates the following steps for mapping management intensity types in grasslands using satellite observations and an ensemble learning classification method: (1) collecting Sentinel-1 and Sentinel-2 satellite registrations as well as field observations to label intensive and extensive grassland areas; (2) creating a feature database using spectral bands indicative of grassland characteristics; (3) applying the RF algorithm by splitting data into training and testing sets; (4) assessing the grassland classification using statistical metrics; and (5) creating illustrations that present the spatiotemporal variations of management intensity in Polish grasslands. We provided brief descriptions for context and used color coding to visually distinguish each step.

Fig. 1.

The process of mapping management intensity types in grasslands using the RF algorithm with satellite observations. Figure created by the authors using Inkscape (https://inkscape.org).

Study areas

The study areas, outlined in red in Fig. 2, are located in various regions of the country, each distinguished by unique natural features such as differences in terrain, water bodies, and vegetation structures. These areas range from lowlands to hilly landscapes, which in turn influence the types of soils, their composition, and their characteristics. Additionally, the sites encompass regions with varying levels of human activity, from highly developed areas to more untouched natural environments, making them particularly valuable for research. In the Podlaskie region of northeastern Poland, coded as PL84 under NUTS 2 system, is known for its diverse and rich grassland ecosystems. The region is characterized by vast open plains, meadows, and wetlands, which provide ideal conditions for the development of various grassland communities. The grasslands in Podlaskie Voivodeship are home to a wide range of plant species, including different taxons of grasses, sedges, and flowering plants. These grasslands support a high level of biodiversity, serving as important habitats for numerous bird species, insects, and small mammals. They also contribute to the cultural and agricultural heritage of the region, playing a significant role in livestock farming and traditional land management practices. The presence of rivers like the Biebrza and Bug enhances the region’s ecological diversity by supporting a rich mosaic of aquatic and wetland ecosystems. These river valleys are often humid, fostering unique ecological conditions that are conducive to the growth and development of the extensive grasslands³⁵. Wielkopolskie Voivodeship, located in west-central Poland and coded PL41 under NUTS 2, encompasses a diverse landscape that includes extensive grasslands. The region is characterized by fertile plains, gently rolling hills, and river valleys, providing favorable conditions for the development of diverse grassland ecosystems. Wielkopolskie Voivodeship’s grasslands exhibit a variety of vegetation types, ranging from meadows to pastures. These grasslands support a rich array of plant species, including various grasses, wildflowers, and herbs, contributing to the region’s biodiversity. The grasslands in Wielkopolskie Voivodeship also play a crucial role in supporting agricultural activities, such as livestock farming and hay and silage production.

Fig. 2.

Location of field surveys conducted at grasslands highlighted by bright red dots in Wielkopolskie (PL41) and Podlaskie (PL84) in Poland. Figure created by the authors using QGIS 3.36 Maidenhead software (https://qgis.org/).

The distribution and characteristics of grasslands in Poland pose significant challenges for nature conservation³⁶. These grassland patches are typically small, which increases the risk of local species extinctions. Moreover, their often considerable distance from one another restricts the movement of species between patches, making it difficult for populations to re-establish³. Study has shown that in 2020, the proportion of permanent grasslands within the agricultural land structure was much higher in the Podlaskie Voivodeship compared to Wielkopolskie, with figures of 38.2% and 13.9%, respectively³.

Figure 2 illustrates also the locations of field surveys conducted in 2021. We identified 10 extensive grasslands and 13 intensive grasslands in the Podlaskie region. Conversely, in Wielkopolskie, we documented 15 extensive grasslands and 8 intensive grasslands. Altogether, we examined a total of 46 grassland sites. Our intensive and extensive grassland fields varied in size. In the Podlaskie region, there were four fields smaller than one hectare and nineteen fields between one and ten hectares. Conversely, in the Wielkopolskie region, the study areas included three fields smaller than one hectare, eighteen fields ranging from one to ten hectares, and two fields larger than ten hectares. Figure 3 presents the extensive (left) and intensive (right) grasslands recognized at our study areas. Extensive grasslands are characterized by a low-intensity management approach that aims to mimic natural ecosystems. These grasslands tend to exhibit higher biodiversity, supporting a wide variety of plant species and providing important habitats for numerous wildlife species³⁷. Due to their more natural management practices, extensive grasslands typically undergo fewer cuts, usually ranging from one to two cuts per year³⁸. This allows the vegetation to grow and mature, promoting the persistence of diverse plant communities and providing a suitable environment for many insect and bird species. Extensive grasslands often occur in proximity to natural areas, such as nature reserves, in the forests, or river valleys, contributing to landscape connectivity and the preservation of ecological corridors.

Fig. 3.

Types of grassland management intensity: (a) extensive, (b) intensive. Photos were taken at the Podlaskie study sites during the field campaign on June 30, 2021.

In contrast, intensive grassland management is focused on maximizing production and achieving high yields of biomass for livestock feed⁶. These grasslands are managed with more intensive agricultural practices to optimize forage production. Typically, intensive grasslands undergo three to six cuts per year, ensuring a more frequent harvest of vegetation⁷. This management approach leads to a more homogeneous grassland structure, with a focus on a limited number of high-yielding grass species³⁹. The grasslands are often fertilized to provide additional nutrients for plant growth and production⁴⁰. Special mixtures of grassland seeds are used, carefully selected to enhance productivity and meet the specific nutritional requirements of herbivorous livestock, e.g. ruminants.

On intensively managed grasslands, vegetation adapted to intensive cultivation and regular mowing is present. Dominant grass species include meadow grass (Poa pratensis) and red fescue (Festuca rubra), which thrive under such conditions. Many intensive meadows in Poland also contain leguminous plants, such as red clover (Trifolium pratense) and white clover (Trifolium repens), which are sown to improve the quality of forage. Among other herbaceous plants, meadow cress (Cardamine pratensis) and evening primrose (Oenothera biennis) can be found, both of which have high nutritional value for animals. Horsetail (Equisetum spp.) can also be found in the wetter areas of intensive meadows.

In contrast, extensive meadows in Poland are characterized by a rich diversity of plant species, reflecting lower levels of agricultural input and management. Typical plants found in these meadows include several grasses, such as common bent (Agrostis capillaris), red fescue (Festuca rubra), and timothy grass (Phleum pratense), all of which thrive under extensive management conditions. Leguminous plants like red clover (Trifolium pratense) and white clover (Trifolium repens) are often present, contributing to soil fertility through nitrogen fixation, while bird’s-foot trefoil (Lotus corniculatus) enhances the nutritional value of the meadow. Among herbaceous plants, meadow buttercup (Ranunculus acris) adds vibrant yellow flowers, yarrow (Achillea millefolium) is known for its medicinal properties, and oxeye daisy (Leucanthemum vulgare) contributes to the aesthetic and ecological value of these habitats. Other herbaceous species include sedges (Carex spp.), knapweed (Centaurea spp.), and common ragwort (Senecio jacobaea), which provide important nectar for pollinators.

A particular case of a plant is the common dandelion (Taraxacum officinale), which is commonly found in meadows, both intensively and extensively managed. In intensive meadows, dandelions may be present, although often in smaller quantities, as intensive management and regular mowing can limit their presence. Dandelions can also serve as a food source for various insects, including bees. In meadow ecosystems, their presence contributes to biodiversity and can be beneficial for other plants and animals.

While extensive grasslands prioritize biodiversity conservation and ecological sustainability, intensive grasslands prioritize agricultural productivity and economic efficiency. The differences in management approaches between extensive and intensive grasslands result in contrasting ecological and agricultural outcomes. Extensive grasslands, with their higher biodiversity and less intensive management practices, contribute to the conservation of native plant species, support a wide range of wildlife, and promote ecosystem services such as carbon sequestration and soil conservation. Intensive grasslands, on the other hand, focus on meeting the demands of livestock production, providing high-quality forage and maximizing the yield of biomass for feed purposes. Understanding the characteristics and distinctions between extensive and intensive grasslands is crucial for effective land management decisions. It enables land managers, policymakers, and conservationists to evaluate the ecological trade-offs associated with different management approaches, implement appropriate measures to protect biodiversity and ecosystem services, and develop sustainable agricultural practices that balance production needs with environmental considerations.

Field observations

Field visits were conducted to identify intensive and extensive grasslands and to collect additional ground truth information in selected areas of the Podlaskie and Wielkopolskie Voivodeships (Fig. 2). The information on the locations of both types of grasslands was necessary to train and test the classification for the entire study areas. The locations of grassland fields were collected using a GNSS (Global Navigation Satellite Systems). Each visit included documenting the area with photographs after each cut to confirm the presence of intensive and extensive grassland. Table 1 presents the dates of field trips conducted during the growing season of 2021 at grasslands in Poland. Due to the locations of the meadows, travel time, and weather conditions in 2021, most field trips were conducted in the grasslands of the Wielkopolskie Voivodeship, totaling 20 days during the growing season from April to September. Meanwhile, fewer field trips were made to meadows located far away in the Podlaskie Voivodeship, scattered across many areas, and with unfavorable weather conditions i.e. frequent cloud cover and rainfall, resulting in a total of 10 days during the growing season.

Table 1.

Dates of field visits conducted during the growing season of 2021 at grasslands in Wielkopolskie and Podlaskie voivodeships.

Year 2021	Wielkopolskie	Podlaskie
April	24, 26 (2 days)	10–11 (2 days)
May	8, 10, 21 (3 days)	9–10 (2 days)
June	4–5, 17, 21 (4 days)	29–30 (2 days)
July	2–3, 17, 29, 31 (5 days)	26–27 (2 days)
August	2, 21–22 (3 days)	23–24 (2 days)
September	6, 19, 23 (3 days)	NA
Total no. of days	20	10

Satellite data acquisition

The study area in Podlaskie encompasses three Sentinel satellite imagery tiles: 34UED, 34UFD, and 34UFE. Similarly, the grass fields in the Wielkopolskie region include three Sentinel granules: 33UWT, 33UWU, and 33UXU. The satellite data were accessible from orbits 22, 122, 79, 36, and 136, allowing for image acquisition every 5–6 days. Sentinel satellite images at processing Level-2 A were automatically retrieved using Google Earth Engine (GEE), a cloud-based platform that provides access to geospatial data, analytical tools, and computational resources for satellite imagery and other geospatial data analysis. GEE supports Python and JavaScript and provides tools for data processing, including machine learning algorithms for image classification and time-series analysis⁴¹. For this study, the JavaScript API was utilized within the Earth Engine Code Editor. While the study areas are characterized by a short and variable growing season, prolonged periods of snow and ice, and significant cloud cover, median mosaic were created from Sentinel satellite images to generate comprehensive imagery of the area. Results on the quality of the compositions from Sentinel-2 satellite images confirmed the use of improved cloud-free and composed daily, weekly, or monthly mosaics for regular land monitoring⁴². GEE employs an algorithm to combine Sentinel datasets into a mosaic by calculating a high-dimensional weighted geometric median, which preserves spectral relationships across all bands.

For accurate representation, we utilized the COPERNICUS/S2_SR_HARMONIZED dataset to prepare cloud-free mosaics for 2021, using images collected between March and September. In our study, we used twelve of the S-2 spectral bands, excluding Band 10, which corresponds to the Cirrus filter. The included S-2 bands were as follows: Band 1 (Coastal Aerosol), Band 2 (Blue), Band 3 (Green), Band 4 (Red), Band 5 (Vegetation Red Edge), Band 6 (Vegetation Red Edge), Band 7 (Vegetation Red Edge), Band 8 (NIR), Band 8 A (Narrow NIR), Band 11 (SWIR), Band 12 (SWIR), and Band 9 (Water Vapour).

Additionally, we used the COPERNICUS/S1_GRD dataset, which includes four variables representing two radar polarizations: Vertical-Vertical (VV) and Vertical-Horizontal (VH), from both ascending and descending orbits. This dataset includes a comprehensive preprocessing workflow performed by GEE. The preprocessing steps include calibration, thermal noise removal, terrain correction, and conversion to backscatter coefficients in decibels (dB). First, the orbit file is applied, which updates the orbit metadata with a restituted orbit file (or a precise orbit file if the restituted one is not available). This ensures that the satellite’s position during acquisition is accurately accounted for. Then, GRD border noise removal is applied to remove low-intensity noise and invalid data from the edges of the scene, which helps eliminate artifacts at the image borders. Additionally, thermal noise removal is applied to remove additive noise in sub-swaths, reducing discontinuities between sub-swaths in scenes with multi-swath acquisition modes. Following this, the radiometric calibration step computes the backscatter intensity using sensor calibration parameters from the GRD metadata. This ensures accurate backscatter coefficients for each pixel. Lastly, terrain correction (orthorectification) is performed to convert data from ground-range geometry into σ° using the SRTM 30-meter DEM or the ASTER DEM for higher latitudes (above 60° or below − 60°). This correction accounts for the terrain’s topography, resulting in more accurate reflectance measurements.

To match the spatial resolution of Sentinel-1 to that of Sentinel-2, we applied pixel neighborhood resampling in Google Earth Engine (GEE). Specifically, we used the “ee.Image.resample()” function to resample Sentinel-1 data to the 10-meter pixel size of Sentinel-2. This method resamples the Sentinel-1 pixels based on the surrounding neighborhood, ensuring compatibility between the datasets and allowing for seamless integration of the data for analysis. To generate the S-2 median mosaic, we processed data from 22 optical satellite acquisitions. Similarly, the S-1 median mosaic was constructed using data from 136 radar satellite acquisitions. The georeferencing system used in the study is EPSG:32,634, which corresponds to WGS 84 / UTM zone 34 N.

Variable importance

We investigated the spectral response curves extracted from twelve S-2 spectral bands taken into account during springtime April-May in Poland. Considering that Potočnik Buhvald³¹ and Saadeldin¹⁹ confirmed the utility of combining S-1 and S-2 for grassland mapping, it is important to note that the similarity in optical spectral characteristics among different grassland types can still significantly reduce the accuracy of classification models. This issue has been observed in recent studies, including Yu⁴³, which highlights how closely related spectral signatures can complicate the differentiation between various grassland types. This challenge is especially clear with remote sensing data, where small differences in spectral reflectance often aren’t enough to accurately distinguish between grassland types. These findings are consistent with observations made by Bekkema⁴⁴, who examined S-2 spectral responses across different grassland management types in the Netherlands. During springtime, Bekkema⁴⁴ noted both similarities and discrepancies in the spectral data, underscoring the difficulty in distinguishing various grassland management practices based solely on their spectral signatures. Such discrepancies can lead to variations in classification performance, highlighting the need for improved methodologies to address these challenges. Conversely, Lange²¹ demonstrated that S-2 spectral bands related to vegetation health and structure are powerful predictors, and that both spring and autumn satellite imagery are highly relevant for land-use assessment. Zandler⁴⁵ showed that the 500-repeated Boruta algorithm identified MSACRI, MSAVI, NDVI, and short-wave infrared Band 12 as key variables for monitoring grassland dynamics.

Therefore, to better understand these spectral characteristics and their impact on classification accuracy, the LOESS (Locally Estimated Scatterplot Smoothing) algorithm was employed⁴⁶. LOESS is useful for fitting a smooth curve to data points in a scatterplot, which can help visualize and analyze the underlying trends and patterns in spectral data. This approach allows for a more nuanced interpretation of how different spectral bands relate to grassland types. In addition to LOESS, the Mann-Whitney U test⁴⁷ was used to determine statistically significant differences between spectral channels. This non-parametric test compares differences between groups and highlights significant variations in spectral data. In the analysis, statistically significant differences between S-2 spectral bands are marked in red to facilitate easy identification of critical patterns and discrepancies.

Random forest algorithm

The Random Forest (RF) algorithm⁴⁸ was employed for grassland classification using the extracted features. RF is an ensemble machine learning algorithm that combines multiple decision trees to create a robust classifier⁴⁹. It is well-suited for remote sensing applications, as it can handle a large number of input variables and capture complex relationships between the features and the target classes. The usefulness of the algorithm for satellite-based classification of different land cover classes in forested and agricultural areas in Poland has been documented in recent numerous studies^15,50–52. The classifier was trained on the extracted features from the combined Sentinel-1 and Sentinel-2 datasets. To reduce the impact of variability of clustering of grassland patches, the classification process utilized field-verified polygons. All pixels recognized within the 46 plot fields, comprising 21 intensive grassland units and 25 extensive grassland units in Wielkopolskie and Podlaskie, were used for the RF algorithm. On average, there were 25 pixels per field, with most fields ranging from 10 to 40 pixels, each measuring 10 × 10 m. Consequently, a dataset of 1200 pixels was derived from a single band of the Sentinel mosaic. By utilizing two different polarization combinations VV and VH in ascending and descending modes by Sentinel-1 SAR sensor and twelve spectral bands from Sentinel-2 MSI, we compiled a comprehensive dataset of 19,210 pixels for model training and validation. A stratified random sampling method was then applied to extract pixel sets, ensuring a 60:40 ratio for training and validation purposes. This approach ensured that pixels from any given polygon were only included in one of the sets, maintaining the condition of independence⁵³. The procedure was repeated 100 times, allowing us to assess the range of classification accuracies for different grasslands types in each iteration. Prior to classification, algorithm hyperparameters were fine-tuned using a grid search method in conjunction with 10-fold cross-validation for each scenario, given the varying number of input variables. For the random forest algorithm, the number of trees (ntree) was fixed at 500, with the mtry parameter being adjusted at 6. RF classifier was utilized with the randomForest package in R (version 4.7-1).

Evaluation metrics for classification model

Classification accuracy was primarily evaluated using overall accuracy (OA), which represents the proportion of correctly classified instances across all classes⁵⁴. Additionally, the F1-score was used to provide a more balanced assessment of accuracy. The F1-score is the harmonic mean of producer accuracy (PA) and user accuracy (UA), offering a measure that accounts for both precision and recall⁴⁸. PA reflects the proportion of actual positives correctly identified by the model, indicating how well a class has been predicted⁵⁴. UA, on the other hand, represents the proportion of predicted positives that are correctly classified, showing the reliability of the prediction. By combining these metrics, the F1-score minimizes bias and provides a more comprehensive evaluation of the model’s performance.

To assess the importance of individual spectral bands on the classification results, commonly employed metric such as mean decrease accuracy (MDA) was utilized. It measures the importance of each variable in maintaining the model’s accuracy. In a random forest, individual trees are constructed using different subsets of the data, and MDA assesses how removing a specific variable impacts the overall accuracy of the model. A high MDA value indicates that the variable plays a significant role, as its exclusion leads to a noticeable decrease in model performance. This makes MDA particularly valuable for identifying the most influential features in a dataset, providing insights into which variables are driving the model’s predictions⁵⁵. Whilst our recent studies^15,50,52 in MDA of vegetation communities using remote sensing proved that Sentinel-2 spectral bands most affect RF classification accuracies, we also investigated the MDA for twelve spectral bands extracted from the Sentinel-2 median mosaic, along with four bands from Sentinel-1, for the two analyzed study areas.

Results

Spectral characteristics of extensive and intensive grasslands

The analysis of S-2 spectral bands derived from April-May satellite acquisitions revealed significant disparities in the mean spectral responses between extensive and intensive grasslands, as evidenced by the Mann-Whitney U test (p < 0.0001, alpha = 0.05). Figure 4 illustrates spectral reflectance curves based on mean and standard deviation reflectance values for extensive and intensive grasslands during the spring season in Wielkopolskie and Podlaskie. Spectral response patterns exhibited increased discrepancy, resulting in increased distinctions between extensive and intensive grassland across several spectral bands. It is noteworthy that the comparison of spectral responses between two types of grassland management intensity, as revealed by the statistical analysis using the Mann-Whitney U test, indicated no significant difference between the two independent spectral responses in the green (560 nm, S-2 band no. 3) and red-edge wavelengths (705 nm, S-2 band no. 5) for both types of grasslands in Podlaskie. Additionally, there was no significant difference observed in the ultra-blue wavelength (443 nm, S-2 band no. 1). The results confirm observations made by Bekkema⁴⁴ regarding similarities and discrepancies in the Sentinel-2 spectral responses among various grassland management types during springtime in the Netherlands.

Fig. 4.

Comparison of spectral curves for intensive and extensive grasslands averaged with the loess algorithm (span = 0.35, confidence interval = 0.95). Statistically significant differences between channels are highlighted in red. Figure created by the authors using R 4.7-1 (https://www.r-project.org/) and Inkscape (https://inkscape.org).

Relevance of input data

By analyzing the influence of backscatter and spectral bands on classification accuracy using MDA as a measure, we are able to assess the importance of individual variables (Fig. 5). The values between the Wiekopolskie and Podlaskie Voivodeships are similar and differ by 1–2 MDA values between them. This demonstrates the effective repeatability of the method in different areas and the bands consistency of the Sentinel-1/2 data. The highest values are for the B11 and B12 channels (average of 10 MDA), which are associated with the Short Wave Infrared (SWIR) range. This is followed also by the high values (average 6) of the Near Infrared channels (B6,B7, B8A, B9), while the lowest values are B1 (ultra blue), B3 (green) and B5 (visible near infrared), whose results are on average below 3 MDA. This coincides very significantly with the statistical significance analysis performed for the individual channels. Our MDA analysis confirmed the findings of Zandler⁴⁵, who demonstrated that the 500-repeated Boruta algorithm highlighted the short-wave infrared as one of the most important variables for monitoring grassland dynamics. While analyzing the Sentinel-1 backscatter bands, our findings align with De Vroey³⁰ analysis, which emphasizes the importance of using the VV polarization for detecting mowing frequency, whereas the role of the VH is negligible. Furthermore, the ascending and descending modes do not significantly affect the usefulness of satellite overpasses at the same study area for mapping different grassland-use intensity types, as confirmed in the recent studies^30,31. While comparing two different study areas, we observed differences of about 1 MDA in S-1 ascending and descending modes for mapping grassland management intensity types in Wielkopolskie and Podlaskie, indicating that the role of acquiring radar satellite data from different orbits is minimal.

Fig. 5.

Variable importance of the RF model measured by MDA between Sentinel-1/2 bands in Podlaskie and Wielkopolskie. Figure created by the authors using R 4.7-1 (https://www.r-project.org/) and Inkscape (https://inkscape.org).

Classification performance

In Podlaskie (PL84), the classification performance for extensive grasslands yielded an overall accuracy (OA) of 84% (Table 2). The classification of extensive grasslands exhibited a user’s accuracy (UA) of 80% and an F1 score of 76%. For intensive grasslands, the classification achieved an OA of 83%, with a UA of 70% and an F1 score of 73%. Similarly, in the region of Wielkopolskie (PL41), the classification of extensive grasslands resulted in an OA of 83%. The classification of extensive grasslands demonstrated a UA of 70% and an F1 score of 73%. The performance metrics for intensive grasslands classification in Podlaskie (PL84) revealed an OA of 84%, a UA of 85%, and an F1 score of 88%. Conversely, in Wielkopolskie (PL41), the classification of intensive grasslands achieved an OA of 83%, a UA of 89%, and an F1 score of 88%. Moreover, the article emphasizes that the producer’s accuracy (PA) of extensive grasslands classification in Podlaskie (PL84) reached 71%, while the PA for intensive grasslands classification in the same region attained a higher value of 90%. In Wielkopolskie (PL41), the PA for extensive grasslands classification was 76%, and for intensive grasslands classification, it was 86%.

Table 2.

Confusion matrix for two management intensities of grasslands in Wielkopolskie (PL41) and Podlaskie (PL84) using Sentinel-1 with Sentinel-2 combined data (PA - producer’s accuracy, UA - user’s accuracy, OA - overall accuracy).

	Wielkopolskie		OA:	83%		Podlaskie		OA:	84%
	Extensive	Intensive	UA	F1		Extensive	Intensive	UA	F1
Extensive	413	173	70%	73%	Extensive	1485	364	80%	76%
Intensive	127	1073	89%	88%	Intensive	598	3451	85%	88%
PA	76%	86%			PA	71%	90%

Spatial distribution of two types of grassland use intensity in Poland

The classification maps for grasslands in the Podlaskie and Wielkopolskie regions were generated using combined S-1 and S-2 median mosaics (Fig. 6). The total area of recognized extensive grasslands in Wielkopolskie is 1384.8 km², whereas intensive grasslands cover over 4067.2 km². In relation to the area of the Wielkopolskie, they represent 25.4 and 74.6% of the total area, respectively. Conversely, in Podlaskie, characterized by a dominant lowland landscape with river valleys, there has been a notable increase in both intensive and extensive grasslands, with a total area of 4914.1 km² and 3459.3 km², respectively. They constitute 58.1 and 40.9% of the area of the Podlaskie Voivodeship, respectively. Our satellite remote sensing detection of intensively and extensively managed grasslands in 2021 aligns with observations regarding the vast spatial diversity in their use in Poland^3,36. Analyses conducted by Gabryszuk³ have revealed that the proportion of permanent grasslands in the agricultural land structure in 2020 is significantly higher in the Podlaskie Voivodeship compared to Wielkopolskie, with proportions of 38.2% and 13.9% respectively. In contrast, our calculations revealed that the proportions of all recognized grasslands in the aforementioned regions ranged from 41.9 to 18.2%.

Fig. 6.

Maps of grassland management intensity classification for Wielkopolskie (left) and Podlaskie (right) Voivodeships. Figure created by the authors using QGIS 3.36 Maidenhead software (https://qgis.org/).

Discussion

This study demonstrates the effectiveness of utilizing a combination of optical and radar remote-sensing data to classify intensive and extensive grasslands with a spatial resolution of 10 × 10 m. The accuracy of the mapping approach was assessed by conducting field visits during the growing season of 2021, where photographs were taken to document the current state. While previous studies have recognized the potential of Sentinel-1 and − 2 data for monitoring grasslands in Poland^56–58, particularly when used together, only a limited number have applied multisensor remote sensing for grassland management classification⁵⁹. Many of these studies focused on small, uniform study areas, but the challenge lies in larger, heterogeneous landscapes with varying environmental conditions and management practices. Despite this, the utilization of combined Sentinel-1 and − 2 imagery for detailed grassland use characterization remains largely unexplored. In this study, we aim to evaluate the capabilities of Sentinel-1 and Sentinel-2 combined, in identifying and classifying grassland use intensity across Poland. Our analysis employed the Random Forest algorithm to assess the suitability of different remote sensing features, such as optical spectral reflectance and radar backscatter, from both ascending and descending orbits, in distinguishing between different types of grassland management intensity.

Grassland-use intensity maps

The use of remote sensing and GIS for analyzing grassland types in Poland has been the subject of research in recent years^25,37. In this context, the utility of CLC and HRL for assessing the spatial distribution of grasslands has also been examined, despite the infrequent, every six years, updates of CLC^23,24 and the varying quality of HRL data²⁶. Therefore, to enhance knowledge and achieve greater precision in mapping grasslands in Poland, we propose an approach based on machine learning that utilizes temporal satellite observations during the growing season. The maps presented in Fig. 6 revealed clear spatial trends in grassland use intensity at regions in western and northeastern Poland. In this country, lowland grasslands, as investigated in our study within the Wielkopolskie and Podlaskie Voivodeships, are primarily located near water bodies such as rivers and in local depressions⁶⁰. The remarkable diversity of habitats within grasslands in Poland has prompted the need for their typological division of meadows⁶¹, that categorizes them into four primary groups: (1) wet meadows, commonly found in river valleys on alluvial soils, serve as vital reservoirs for water retention, capable of storing around 10 million cubic meters of water; (2) dry-ground grasslands, situated on local elevations in river valleys, flat non-flooded depressions, mid- and near-peatland elevations, and at the edge of arable lands. Apart from humid habitats, these grasslands generally have low to moderate moisture levels, primarily sustained by rainfall and runoff waters. They occupy various habitats, ranging from fertile to poor soils, including impoverished dry grounds more suitable for afforestation than intensive grassland management; (3) bog grasslands located on peatlands and wetlands, these areas typically exhibit excessive moisture and are dominated by sedges, often accompanied by mosses. While not used agriculturally, bog grasslands hold significant natural value and (4) post-bog grasslands situated on reclaimed peatlands with fairly uniform soils composed mostly of peat or partly decomposed peat and stable moisture levels. These areas are typically drained but can undergo rapid and unfavorable changes without proper grassland management. Following Gabryszuk³ wet meadows constitute 20% of grassland areas in Poland, dry-ground account for 40% of grass areas in the country, while bog and post-bog grasslands make up for 8% and 32% of grassland areas respectively.

Managed meadows and grasslands are typically situated on moderate to poor soils, which are not suitable for cultivation. The spatial distribution of land relief features and soil properties contributes significantly to the clustering of grassland patches³⁶. This finding confirms our observation, notably in the Podlaskie Voivodeship (Fig. 6). Conversely, there are extensive areas in Poland, such as in the Wielkopolskie Voivodeship, where lowland grasslands are notably scarce (Fig. 6). This results from the fact that the Wielkopolska region is known for its agriculture, with major crops including potatoes, sugar beets, and wheat, hence it is referred to as the “Cradle of Europe“⁶². While the spatial patterns remained consistent within the same lowland landscape type across regions in Poland, we found that the method’s reliability persisted despite variations in the clustering of grasslands patches in these two study areas. It’s important to note that the maps didn’t differentiate between mowing and grazing events, unlike previous studies that focused solely on mowing^30,63,64. Similar to previous research⁶⁵, distinguishing between meadows and pastures is crucial for accurately mapping grassland-use intensities, although this is challenging due to limited high-quality reference data. The extensive dataset gathered in this study could be utilized for further exploration, such as refining methods to distinguish between different types of grassland uses using remote sensing techniques.

The study aimed to comprehensively assess and accurately classify the two predominant grassland management intensities, namely extensive and intensive practices, across the specified regions. As reported by Chang²⁰ grassland management intensity (intensive or extensive) play a crucial role in the greenhouse gas balance and surface energy budget of this biome, both at field scale and at large spatial scale. The generated classification maps constitute valuable resources for various applications and decision-making processes. The synergistic use of Sentinel-1 and Sentinel-2 satellite datasets for grassland classification offers a promising and efficient approach for comprehensive and accurate monitoring of these critical ecosystems, enhancing our understanding of their distribution, dynamics, and ecological significance.

Algorithm performance with suggestions for improvements

Classification accuracy was evaluated using measures such as Overall Accuracy (OA), Producer Accuracy (PA), and F-score (F1). It was observed that extensive grassland consistently exhibited lower producer and user accuracies compared to intensive grassland. For instance, the user accuracy for extensive grasslands in Podlaskie was recorded at 80% applying RF, object-based classification, 2 classes and with 20 selected features, whereas for intensive grassland, it stood at 85%. Similarly, in Wielkopolskie, the user accuracy for extensive grassland dropped to 70% using RF, object-based classification, 2 classes, with 9 selected features, while for intensive grassland, it remained high at 89%. The lower accuracies observed for extensive grassland were attributed to its higher biodiversity value, which varies depending on soil texture, the vast study area with diverse environmental conditions, and the imbalance in the training dataset with 39% of polygons representing extensive grassland and 61% representing intensive grassland. Therefore, further research is warranted to explore methods for expanding and balancing the number of reference training samples to enhance grassland management intensity classification. This could involve incorporating additional attributes obtained from topographic data³¹, phenocams⁶⁶ or webcams⁶⁴ particularly in areas where the current samples are not adequately representative. Performance metrics of the model demonstrate its reliability despite two factors: first one on the spatial patterns showing inconsistency across different lowland landscape types in various regions of Poland, and second one on variations in weather conditions affecting vegetation growth and the availability of satellite imagery.

Table 3 presents the basic characteristics of researches on grassland mapping carried out at study areas in Europe. Numerous studies have applied the SVM algorithm for classifying grassland management or intensity levels^67–69. While Marcinkowska-Ochtyra⁵¹ tested three common and well-known machine learning algorithms SVM, RF and CNN on mapping grasslands habitats in Poland, the results differed slightly among them, 88%, 86% and 84% respectively. According to Table 1, most studies^{19,21,44,64,68–70} examining remote sensing of grasslands management intensity have heavily relied on optical satellite data, particularly using vegetation indices to assess grassland conditions effectively. While radar-based parameters such as backscatter amplitudes, interferometric coherence, and polarimetry-based decomposition have been investigated in some studies^30,63, they have mainly been applied to detect grassland mowing rather than classifying grassland types in general. Our study emphasizes the possibility of simultaneous use of the spectral data from twelve Sentinel-2 bands and Sentinel-1 backscatter for effectively mapping and monitoring grassland utilization. The combination of SAR backscatter and spectral reflectance provides valuable insights into distinguishing between intensively and extensively managed grasslands. It should be noted that utilizing vegetation indices such as NDVI^45,65,69,71, EVI⁶⁴, and LAI⁷⁰ has been significant in classifying grassland types. Nevertheless multispectral responses from S-2 bands used in our study offer several advantages over traditional vegetation indices for mapping grasslands intensity due to their comprehensive and detailed representation of the Earth’s surface. Firstly, S-2 provides data across thirteen spectral bands, covering a wide range of the electromagnetic spectrum from the visible to the near-infrared and shortwave infrared. This rich spectral information allows for the detailed discrimination of different grassland management types based on their unique spectral signatures, which is often more nuanced than what can be captured through a few indices^71–73.

Table 3.

A brief overview of recent satellite remote sensing applications with machine learning algorithms in grassland classification.

Authors	Data	Study area	Algorithm	Object	No. of classes	Best accuracies (%)
Our study	Sentinel-1 Sentinel-2	Poland	Random Forest	Management intensity types	2	83–84
70	RapidEye	Germany	Decision Tree	Management intensity types	4	86.1
44	Sentinel-2	Netherlands	Decision Tree	Management intensity types	2	84.3
68	MODIS	EU-27	Support Vector Machine	Management intensity types	3	77–80
21	Sentinel-2	Germany	Convolutional Neural Network	Management intensity types	3	66–85
69	Formosat-2	France	Support Vector Machine	Management types	3	75
31	Sentinel-1 Sentinel-2 Topography Data	Slovenia	Random Forest	Management intensity types	2	70–84
19	Sentinel-1 Sentinel-2	Ireland	Convolutional Neural Network Random Forest	Management intensity types	3	92.8 84.8
64	Sentinel-2 Landsat 8	Switzerland	mowingDetection_UDF	Management intensity types	4	78
65	Sentinel-2	Switzerland	k-means	Thematic classes	2	68.8–79.7
51	Sentinel-2	Poland	Support Vector Machine Random Forest Convolutional Neural Network	Habitats	3	88 86 84
76	Sentinel-2	France	Random Forest Support Vector Machine	Habitats	7	78 71
67	Sentinel-2	Italy	Support Vector Machine	Habitats	4	93–95

Additionally, relevance analysis of input data, performed using methods such as the MDA in our study or the Boruta algorithm used by Zandler⁴⁵, confirms the necessity of selecting specific bands or band combinations tailored to the classification of particular grassland-use intensity types. This flexibility allows for the extraction of more relevant information for distinguishing between classes, compared to relying on general-purpose vegetation indices like NDVI^45,65,69, EVI⁶⁴ or LAI⁷⁰. Moreover, in complex ecosystems or in areas where land cover types have similar spectral characteristics in the visible and near-infrared parts of the spectrum, the additional bands provided by S-1 and S-2, such as those in the shortwave infrared and backscatter VV and VH, can be crucial for accurate classification. These bands can help identify subtle differences that are not apparent in traditional vegetation indices. In the case of similar spectral profiles for extensive and intensive grasslands in Wielkopolskie, as shown in Fig. 4, it highlights the need to search for additional variables that would enable a clear distinction between these two types of grasslands. It is important to note that Wielkopolskie is characterized by a predominance of cultivated land, whereas intensive grasslands are much more scattered, often narrow in width, and generally occur in low densities compared to those in Podlaskie^3,36. In contrast, intensive grasslands are primarily found in the valleys of the Notec and Warta rivers, along with their tributaries and artificial channels. These characteristics imply challenges in clearly delineating the boundaries between the two types of grasslands.

Furthermore, Tesfaye⁷⁴ and Dusseux⁷⁵ have confirmed that the utilization of multiple spectral bands can mitigate the limitations of vegetation indices. For instance, in high biomass areas like grasslands, where the indices may become saturated and no longer increase linearly with biomass, the inclusion of additional spectral bands proves invaluable. This ensures more accurate classification in densely vegetated areas. Lastly, the temporal dynamics of land use and vegetation can be better understood using multiple bands, allowing for monitoring changes over time, assessing the impacts of seasonal variations, and improving classification accuracy by considering the phenological state of vegetation. In summary, while vegetation indices have their utility, the comprehensive spectral information offered free-of-charge by multispectral satellite sensors like S-2 or Landsat 8⁶⁴ enables a more nuanced and accurate classification of land cover maps. This approach leverages the full potential of satellite imagery for environmental monitoring and management.

In this study, in addition to the spectral response from S-2, we utilized two backscatter polarizations, VV and VH, from both ascending and descending modes of S-1. This choice is attributed to several key factors. First, VV and VH polarizations are particularly sensitive to the vertical structure of vegetation, especially diversified at wetlands and grasslands⁷¹. Grasslands typically have a vertical component, such as grass height and density, and these polarizations capture variations in vegetation structure more effectively, allowing for better differentiation between grassland types³⁰. In contrast, HH and HV polarizations may not respond as effectively to vertical vegetation features, which can lead to less accurate assessments of wetland and grassland conditions^59,63. Another advantage of VV and VH polarizations is their ability to reduce the effects of soil backscatter, especially in moist conditions^56,64. This reduction allows for a clearer signal from vegetation, which is particularly important in grasslands, where fluctuating soil moisture levels can impact backscatter readings. Additionally, the combination of data from both modes enhances our ability to differentiate between backscatter signals from soil and vegetation. This capability is especially beneficial in alluvial extensive grasslands, where soil moisture levels can fluctuate and affect backscatter readings. Furthermore, we considered the fact that De Vroey³⁰, using only the descending modes of S-1 VV and VH to estimate coherence, demonstrated in his study that, depending on various grazing scenarios, only 56% of all parcels and 59% of the mown parcels were identified correctly in terms of mowing dynamics. Most errors consisted of false detections in grazed parcels and omitted spring mowing events. Similarly, Weber⁶⁴ indicates that the relationship between radar data and mowing events appears to be much weaker, potentially leading to a higher incidence of false positives compared to optical satellite imagery.

While Potočnik Buhvald³¹ demonstrated that both types of data provide valuable information for mapping grassland-use intensity types, especially NDVI and S-1 backscatter values, our analysis confirmed that the optical spectral bands from S-2 are characterized by significantly higher MDA values compared to the radar backscatter from S-1 (Fig. 5). While this research study³¹ analyzed MDA, focusing on temporal variability and satellite data acquisition using solely NDVI from optical spectra, we investigated MDA utilizing all twelve optical bands, highlighting new valuable information on B7, B8, B11, and B12 from S-2. Therefore, based on our results, it can be assumed that S-2 data provide a more precise utility for monitoring grassland-use intensity types, as they do not misclassify intensively managed grasslands with frequent mowing events. This is also relevant for recognizing extensive alluvial grasslands, which typically have only one mowing event at the end of the growing season. In Wielkopolskie, no distinct difference in MDA values was observed when using S-1 VH backscatter. However, it is worth noting that only the VV polarization, in both descending and ascending modes, showed significant improvements in distinguishing between grassland types, while the VH polarization did not demonstrate such notable differences, as observed in recent studies^30,31. Based on the results obtained, it can be concluded that monitoring grassland-use intensity types should primarily rely on remote sensing data, particularly multispectral images, which prove to be more versatile. This approach yielded satisfactory results for all the studied grassland types. By employing MDA, we can effectively reduce the dimensionality of the input data and identify the satellite image attributes that are most significant for constructing a model distinguishing between intensively and extensively managed grasslands within a specific year. Furthermore, the MDA criterion generated as output from RF classifier has demonstrated high efficiency in feature selection.

Applications for grasslands management and biodiversity conservation

These maps can support the analysis of spatial and temporal trends in grassland use intensity, enabling an investigation into its correlation with biodiversity indicators at the regional scale, specifically in areas such as Wielkopolskie and Podlaskie. Remote sensing-based intensity assessments can contribute to effectively capturing various production conditions and management techniques, serving as reliable predictors of plant species richness and ecological indicator values. Previous studies by Bekkema⁴⁴ and Abdollahi⁷⁷ have highlighted the necessity of using satellite imagery, particularly in spring before the first mowing date, to achieve accurate grassland classification. Our research supports this assertion, showing that spectral responses during April and May period play a crucial role in distinguishing between intensively and extensively managed grasslands, capturing the dynamic growth phase of vegetation. This is why field observations before the first mowing date in the end of May / at the beginning of June 2021 in Poland are so crucial to gather materials for verification and assess the accuracy of grasslands classification. Despite unfavorable weather conditions, it was possible to go on field trips for 9 and 6 days in Wielkopolskie and Podlaskie respectively during this period (Table 1). Radar backscatter further enhances classification accuracy, especially in regions with high cloud cover, such as the temperate climate of Poland. This is particularly relevant given the frequent cloud cover in Poland, averaging 180 days per year³⁴.

Classification of grassland maps holds significant value for ecology and conservation efforts aimed at preserving natural habitats. Firstly, these maps provide essential information about the distribution and extent of grasslands, which are critical ecosystems supporting diverse flora and fauna^78,79. By accurately classifying grassland areas, ecologists gain insights into the spatial patterns of biodiversity, allowing for targeted conservation efforts to protect vulnerable species and their habitats. According to Szymura³⁶ the distribution of grasslands and their characteristics across Poland presents significant challenges for nature conservation. Typically, these grassland patches are relatively small in size, which heightens the risk of local species extinction. Additionally, they are often situated at considerable distances from one another, impeding the movement of populations between patches and hindering their ability to re-establish³. In numerous regions in Poland grassland patches are separated by vast distances, exacerbating these challenges^39,57,58. The existing systems of protected areas and ecological corridors are insufficient for effectively safeguarding grasslands³⁶. There is an urgent requirement to formulate a comprehensive landscape-scale plan dedicated to grassland protection to support conservation endeavors. This plan should encompass both protected areas and ecological corridors to facilitate the connectivity of habitats³⁷. The development of such a plan hinges upon the availability of accurate and readily accessible remote sensing data regarding the distribution of grassland patches in Poland^56,59. Given the rapid decline of semi-natural grasslands and the ongoing global environmental changes, the establishment of such a system is imperative⁶⁸. Urgent action is needed to address these issues and ensure the preservation of grassland ecosystems for future generations.

Additionally, monitoring grassland management intensity allows farmers to evaluate the health and condition of grassland ecosystems over time. By observing changes in land cover and vegetation composition, areas experiencing degradation or being overtaken by invasive species can be identified^52,80. This information is crucial for implementing management strategies to restore degraded grasslands and maintain their ecological integrity. This integration of ecological knowledge into land management practices promotes the preservation of biodiversity while meeting the needs of local communities and stakeholders³⁷. In summary, classification of grassland maps is invaluable for ecology and conservation efforts, providing essential information for understanding, protecting, and managing these vital ecosystems. By accurately mapping grassland habitats and monitoring changes over time, conservationists can effectively prioritize conservation actions, restore degraded areas, and promote sustainable land management practices to ensure the long-term health and resilience of grassland ecosystems.

Limitations of the study and future research directions

Although we used S-1 and S-2 satellite data to evaluate their effectiveness in mapping grassland-use intensity types in two study areas, our approach still has some limitations. We defined management intensity types based on mowing frequency, rather than on plant indicators, to distinguish between extensive and intensive meadows. The S-1 dataset, which is widely used for detecting mowing events, is generally considered reliable. By combining it with optical data from S-2, the spectral responses proved highly useful as predictors for classifying grassland types using the RF algorithm (Fig. 5). As discussed, previous studies often rely on either S-1 radar data or S-2 optical data, but rarely both, which highlights the broader possibilities for satellite imagery selection. Choosing satellite data is not a zero-sum decision, meaning we cannot definitively favor one data type over the other. Our analysis demonstrated the necessity of using both S-1 and S-2 data.

However, several limitations should be addressed in future research. The first is the sample size, which is currently restricted to two study areas in Wielkopolskie and Podlaskie. Collecting field observations is time-consuming and requires collaboration with private Polish farmers and agricultural agencies to obtain permissions. In this study, we collected data from these two regions to support satellite-based mapping of grassland types. However, the methodology, particularly the spectral curve similarities observed in Wielkopolskie (Fig. 4) and the MDA differences between the two regions (Fig. 5), suggests the need for additional training data. This would improve the calibration and validation of the mapping model across a broader area, such as at the national level. Expanding the dataset is crucial for thoroughly testing the Random Forest algorithm, which has so far only been applied at a more localized administrative level. Additional data could include land registry information with parcel boundaries from the Land Parcel Identification System (LPIS) and current land cover/use data to identify fallow and abandoned lands.

The transferability of the model we implemented is a critical aspect of our research, as it determines the model’s applicability across different regions and conditions. We employed a RF algorithm, known for its robustness and flexibility in handling various types of data. This model was trained on the field data collected in the Wielkopolskie and Podlaskie, allowing it to effectively capture the complex relationships between S-1 backscatter and S-2 spectral bands and management intensity. The results indicated that while the model performed well in familiar settings, its accuracy diminished in regions significantly different from the training areas, such as Wielkopolskie (Table 2). Therefore, further calibration and adaptation may be necessary to enhance its applicability in diverse study areas in Poland, thereby improving the model’s effectiveness in broader agricultural and ecological contexts.

Another limitation that also represents a potential direction for future research is the consideration of specific plant species as indicators for mapping management use intensity in grasslands. However, achieving this would require very high-resolution multispectral satellite data with a pixel size of 1 × 1 m, which would allow for the identification of individual species and capture the spatial variability of their distribution. In previous years, this approach has been attempted using WorldView, GeoEye-1, and Pleiades-1 A/1B, which demonstrated superior spatial resolution, and these were compared with hyperspectral data collected from satellite, airborne and unmanned aerial vehicle platforms. Nevertheless, the primary limitation is the cost of acquiring such data, which is available commercially, and in the case of airborne hyperspectral data, typically requires custom orders, often resulting in just one flight throughout the entire growing season.

In conclusion, there are numerous opportunities for significant contributions to future research. This includes exploring alternative data sources and establishing distinct management intensity classes through the implementation of new indicators for differentiating various types of grasslands-use intensity. A key component remains the application of machine learning, coupled with robust modeling and performance analysis, which enables a thorough assessment of the validity of the approaches employed in these studies. These opportunities can help advance our understanding of grassland ecosystems and contribute to more effective management practices.

Conclusions

The classification of grassland management intensity using the combination of S-1 and S-2 satellite imagery in the regions of Podlaskie (PL84) and Wielkopolskie (PL41), Poland, reveals insights into the ecological and socio-economic dynamics of these vital ecosystems. Grasslands, with their pivotal roles in biodiversity conservation, carbon sequestration, and agricultural productivity, pose challenges in accurate classification, particularly across large spatial extents. Leveraging the unique capabilities of Sentinel-1’s radar imagery and Sentinel-2’s multispectral data, this study employs a synergistic approach to overcome these challenges. Preprocessed Sentinel-1 and Sentinel-2 data are combined to extract spectral and textural features, enhancing our understanding of grassland characteristics and patterns. Supervised classification using the Random Forest algorithm, trained with ground truth data from field surveys, reveals distinctions in classification performance between the two regions. Extensive grasslands achieve overall accuracies of 70–76%, while intensive grasslands obtain 86–90% producer accuracy. Additionally, classification metrics such as user’s accuracy, F1 score, and producer’s accuracy highlight variations in classification accuracy. The combination of radar and optical data improves classification accuracy, providing complementary information crucial for distinguishing between grassland management intensities. The mapping of grassland management intensity offers valuable insights for conservation agencies, agricultural stakeholders, and land managers, facilitating sustainable land management and decision-making processes. Furthermore, the study underscores the potential of Sentinel missions for grassland monitoring and emphasizes the importance of advanced remote sensing techniques for understanding and preserving these critical ecosystems. Overall, the comprehensive classification of grassland management intensity, integrating Sentinel-1 and Sentinel-2 satellite data, contributes to advancing our knowledge of grassland ecosystems and supports their conservation and sustainable management.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, M.B., M.K., K.D.-Z.; methodology, M.B. and M.K.; software, M.K.; validation, M.B., M.K., K.W. and P.G.; formal analysis, M.B. and M.K.; investigation, M.B., M.K., K.D.-Z., P.G., B.G.; resources, M.B., M.K. and K.W.; data curation, M.B., M.K. and P.G.; writing—original draft preparation, M.B. and M.K.; writing—review and editing, M.B., M.K., P.G., B.G.; visualization, M.B. and M.K.; supervision, K.D.-Z. and P.G.; project administration, K.D.-Z., and P.G.; funding acquisition, K.D.-Z. and P.G. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by Polish-Norwegian Research Programme, the GrasSAT project (grant agreement no. NOR/POLNOR/GrasSAT/0031/2019-00).

Data availability

The datasets used and analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.