| dc.description.abstract | Accurate crop field delineation from satellite imagery is a critical component of agricultural monitoring. However, most existing models are developed and evaluated in large-scale, industrial agricultural regions, where field boundaries are relatively regular and high-quality annotated data is more readily available. In contrast, smallholder regions—where fields are smaller, more irregularly shaped, and often lack precise geospatial labels—remain underrepresented in both data and model performance. This thesis investigates model architectures, loss functions, and learning paradigms for improving segmentation performance in smallholder settings. Using datasets from Austria, India, and Rwanda, we evaluate several model configurations including ResUNet++ with Dice+BCE and Tanimoto+BCE losses, a meta-learned ResUNet++ using Model-Agnostic Meta-Learning (MAML), and SAM2 ViT-H, a large vision transformer released by Meta, evaluated in a zero-shot setting. We introduce a data processing pipeline that converts vector field boundaries from the FTW dataset into highresolution image–mask pairs suitable for supervised learning. Quantitative and qualitative results reveal that models trained on industrial-scale data perform poorly in smallholder regions without adaptation. SAM2 exhibits strong zero-shot performance, especially on larger fields, while ResUNet++ models trained directly on India perform more consistently across small-to-medium sized fields. MAML yielded underwhelming performance under resource constraints, highlighting the need for further tuning. These findings underscore the importance of geographically diverse, well-aligned training data and support the case for developing globally representative agricultural segmentation datasets. | |
