Crop Field Boundary Detection Using 3d Convolutions in Multi-Spectral Multi-Temporal Hr Satellite Images

The advent of new satellite missions offering high spatial, spectral, and temporal resolution has significantly enhanced the possibility to monitor vegetation and agricultural practices. The High-resolution (HR) Satellite Image Time Series (SITS) enables a deeper understanding of crop fields behavior and precise boundary detection. While Convolutional Neural Networks (CNNs) have demonstrated effectiveness in crop fields-related analyses, existing methods for crop boundary detection often focus on mono-temporal image analysis, overlooking valuable multi-temporal information in SITS. To address this gap, we propose the utilization of a UNet-based three-dimensional (3D) CNN architecture, allowing for the simultaneous modeling of spatial-temporal information within multi-spectral multi-temporal SITS. Additionally, we explore various CNN-based U-Net models to further validate the proposed approach in accurately detecting crop field boundaries. The method is evaluated in an agricultural area in Germany using 12 Sentinel-2 Level-2A images and has demonstrated promising results.