Comparative evaluation of Sentinel-1, Sentinel- 2, and fused time series for agricultural field boundary delineation using Deep Learning

Details

Supervisors

Defourny, Pierre ; Deffense, Quentin

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

Agricultural field boundary delineation (AFBD) is fundamental for agricultural monitoring, food security assessment, land management, and the implementation of agricultural policies. Recent advances in Earth Observation (EO) and deep learning have enabled the automatic extraction of agricultural parcels from satellite imagery. Sentinel-2 optical imagery has demonstrated strong performance for AFBD, but its effectiveness is limited in regions affected by persistent cloud cover. Sentinel-1 Synthetic Aperture Radar (SAR) provides weather-independent observations and offers complementary information through backscatter intensity and interferometric coherence. However, the contribution of SAR coherence to AFBD remains largely unexplored, particularly in temperate agricultural landscapes. This thesis comparatively evaluates Sentinel-1, Sentinel-2, and fused Sentinel-1/Sentinel-2 time series for AFBD in Wallonia, Belgium. Multi-temporal Sentinel-1 SAR and Sentinel-2 optical imagery acquired between 2018 and 2021 were combined with reference parcel boundaries derived from the Walloon Land Parcel Identification System (LPIS). A PTAViT3D-based deep learning architecture was implemented and evaluated under five configurations: cloud-free Sentinel-2, Sentinel-2 with less than 50% cloud contamination, Sentinel-1 coherence only, Sentinel-1 coherence combined with backscatter, and fused Sentinel-1/Sentinel-2 inputs. Performance was assessed using pixel-level segmentation metrics and polygon-level geometric evaluation. Cloud-free Sentinel-2 achieved the strongest performance (MCC 0.919 for extent, 0.695 for boundaries), with only marginal degradation under 50% cloud contamination, confirming the robustness of attention-based temporal modelling to cloud cover. Sentinel-1 alone reached an MCC of 0.852 for extent and 0.509 for boundaries when combining coherence, backscatter, and dual-orbit acquisitions; interferometric coherence is shown to carry sufficient discriminative information to detect field extent, although it remains less suited for precise boundary localisation when used in isolation. The fused Sentinel-1/Sentinel-2 model did not surpass the optical baseline, with SAR sensitivity to intra-field surface heterogeneity producing over-segmentation. The results clarify the operational trade-offs between optical, SAR, and fused inputs for AFBD in cloud-prone temperate agricultural landscapes.