Christobaltobbin/SAM-Africa-Fields

Fine-tuning SAM 2 for smallholder agricultural field boundary delineation in sub-Saharan Africa # SAM AFRICA FIELDS Fine-tuning Segment-Anything-Model (SAM 2) for smallholder agricultural field boundary delineation in sub-Saharan Africa. ## Motivation Standard SAM was trained on 11 million images, but none of them are the irregular, small-area farm plots typical of African smallholder landscapes. The result is poor field boundary delineation in contexts where accurate mapping matters most, for food security monitoring, human-wildlife conflict mitigation, and EUDR compliance in conservation areas like the Kavango-Zambezi (KAZA) TFCA. This project addresses that gap through domain adaptation of SAM 2 using LoRA fine-tuning on curated African field boundary datasets. ## Background This work extends the findings of my MSc. thesis: *Enhancing Land Cover Classification in Southern Africa Using Multi-Sensor and Multi-Algorithm Approaches* (Tobbin, 2025, Julius-Maximilians-Universität Würzburg), which demonstrated that vanilla SAM achieved only 7% IoU on Sentinel-2 and 2% IoU on PlanetScope imagery over Binga District, Zimbabwe; primarily due to the model's inability to handle small, irregular field geometries common in the region. ## Related Work M.Sc. Thesis Repository: Christobaltobbin/Segment_An…