UNet Segmentation for Accurate Parcel Delineation: Improvement with temporal multi-spectral images

Abstract. Agricultural parcel delineation is critical for generating cadastral maps that underpin sustainable land management, precision agriculture, and data-driven policymaking. While satellite imaging provides a scalable solution, most existing approaches rely on static RGB or single-date spectral data, neglecting the temporal dynamics of agricultural landscapes. This study introduces TempAgriBound, a novel temporal multispectral dataset designed to advance parcel boundary detection by capturing both spectral and phenological features. The dataset comprises a dense time-series of Sentinel-2 multispectral imagery (10 bands at 10m resolution) and derived spectral indices (NDVI, NDWI, SAVI, etc.), spanning the entire 2023 growing season in Brittany, France (a region characterized by diverse crop rotations and fragmented landholdings). We propose a 3D U-Net architecture optimized for spatiotemporal feature extraction, which processes multi-spectral time stacks to exploit crop growth stages and seasonal spectral variations. For comparison, a 2D U-Net baseline using single-date RGB composites was implemented. By systematically evaluating these models, we aim to determine the differential effects of temporal spectral information on parcel boundary detection. These findings underscore the synergistic value of temporal resolution and spectral diversity in automated parcel mapping, particularly in regions with complex crop patterns. The study advances scalable precision agriculture tools and provides actionable insights for integrating temporal-spectral data into national land registries.