Enhancing Crop Field Segmentation with Super-Resolved Sentinel-2 Imagery: A Deep Learning Approach Leveraging Multitemporal Edge Detection

This study explores the potential of integrating super-resolved Sentinel-2 imagery to improve crop field segmentation, addressing the challenges in accurately delineating small and/or irregularly shaped agricultural fields. By combining multitemporal edge detection with deep learning techniques, segmentation accuracy is significantly improved. Super-resolution, implemented via an Enhanced Super-Resolution Generative Adversarial Network (ESRGAN), increases the spatial resolution of Sentinel-2 imagery from 10 meters to 2.5 meters, enabling finer detail extraction. A ResUNet model trained on standard 10 m resolution datasets demonstrates strong performance when applied to super-resolved imagery, showcasing its adaptability. Further improvements are achieved when models are specifically trained on super-resolved datasets. Results demonstrate enhanced segmentation in complex agricultural landscapes, such as the Fucino Plain (Italy), Punjab (India), and Suzhou (China), with ground truth-based validation in Quebec (Canada).