Deep Learning-Based Grassland Classification Using Multi-Modal Sentinel-1, Sentinel-2, and Street-Level Imagery
Journal
SSRN
Date Issued
October 2, 2025
DOI
10.2139/ssrn.5884802
Abstract
Accurate grassland classification is important for sustainable land management and ecological monitoring, particularly in policy contexts such as the EU's Common Agricultural Policy (CAP). This study explores the application of deep learning models to classify grasslands using multi-modal remote sensing data comprising Sentinel-1 SAR, Sentinel-2 multispectral, and street-level images. Each of these data sources provides complementary information: Sentinel-2 provides rich spectral information, Sentinel-1 provides structural and moisture-related information regardless of the weather conditions, and street-level images provide ground-level views with high detail. Our results show that Sentinel-2 alone provides strong classification accuracy. Adding Sentinel-1 provides small but consistent gains, and even when Sentinel-2 is unavailable, Sentinel-1 alone still provides quite accurate results. While the inclusion of street-level imagery does not surpass the performance of satellite-only models, it offers complementary value in certain cases, such as visually complex or ambiguous parcels. The results show the value of fusing satellite and ground-level data for improved grassland classification, demonstrating the potential of deep learning for setting up scalable and high-accuracy environmental monitoring systems.
File(s)![Thumbnail Image]()
Name
ssrn-5884802.pdf
Size
1.66 MB
Format
Adobe PDF
Checksum (MD5)
5975627bb6c9f3655010308ccc5fed75