Résumé
In this work, we introduce a comprehensive framework for the evaluation of agricultural parcel delineation methods that rely on binary semantic segmentation models. The framework consists of three main components: a binary semantic segmentation model trained specifically on parcel interiors, a watershed-based post-processing approach for parcel instance retrieval, and the application of Hoover metrics. We compare several state-of-the-art deep learning architectures. Our analysis reveals that traditional metrics, such as IoU and F1-score may fail to capture important qualitative differences between methods, particularly in their ability to separate adjacent parcels. Consequently, we adopt Hoover metrics to provide more nuanced insights into model performance by categorizing instance detection outcomes. This approach offers a pathway to transition from overall area-based metrics to instance-level analysis for agricultural parcel assessment.