Early-season delineation of agricultural fields using a fully convolutional multi-task network and satellite images

: The accurate delineation of agricultural fields is essential for crop condition monitoring, yield estimation, and irrigation management. While satellite imagery offers a cost-effective solution for large-scale mapping, traditional boundary detection methods often face challenges such as false edges and suboptimal segmentation. To address these limitations, we present an operational multi-task deep learning approach using the ResUNet-a d7 model, leveraging freely available Sentinel-2 Level-3A data, which ensures enhanced temporal and spatial consistency for large-scale applications. Additionally, we introduce a novel post-processing method based on Gaussian Mixture Models (GMM) to refine boundaries between adjacent fields, enabling precise extraction of individual fields. Extensive assessments were conducted across 14 geographically diverse sites spanning two years of data, along with spatio-temporal experiments to evaluate the model's transferability to unseen regions and new acquisition year. The results demonstrate that the ResUNet-a d7 model achieves a high weighted F1 score of approximately 92%, highlighting its strong performance and reliability. This study provides a scalable and robust solution for early-seasonagricultural field delineation, paving the way for operational applications in irrigation detection, crop type identification, and crop health monitoring.