SIDEST: A sample-free framework for crop field boundary delineation by integrating super-resolution image reconstruction and dual edge-corrected Segment Anything model

•Proposed a sample-free method (SIDEST) for large-scale recognition of crop fields.•Sentinel-2 imagery was enhanced to 2.5 m and PlanetScope to 1 m for delineation.•Dual-edge corrected SAM reshaped unsupervised image segmentation.•SIDEST was evaluated over Jiangsu province of China and on a public Danish dataset.•It outperformed existing delineation methods and deep learning networks. Digital crop field boundaries are fundamental for parcel-level cropland monitoring and precision crop management. The use of very high-resolution (VHR) remote sensing images represented a popular approach for crop field boundary delineation, but mapping crop fields over large regions remains challenging. Moreover, few studies have been dedicated to reforming sample-free delineation methods with ever-advancing deep learning techniques, such as the Segment Anything Model (SAM). This study proposed a sample-free method for delineating crop field boundaries from free Sentinel-2 (S2) or low-cost PlanetScope (PS) images by integrating Super-resolution Image reconstruction and Dual Edge-corrected Segment anyThing model (SIDEST). With the VHR images from a Chinese satellite Gaofen-2, the spatial resolutions of S2 and PS images were elevated separately to 2.5 m and 1 m through the generative adversarial network. Subsequently, a novel image segmentation algorithm was developed by correcting the SAM output twice with edge intensity information. Crop fields were identified automatically using the key phenological information extracted from bi-temporal satellite imagery. SIDEST was evaluated over three test areas in Jiangsu Province of eastern China and on an independent public dataset by comparison with existing algorithms. The results demonstrated that the super-resolution module improved the clarity of PS and S2 imagery, which led to an increase in the intersection over union (IoU) by 10 % for crop boundary delineation. Furthermore, the dual edge-corrected SAM outperformed edge detection and multi-scale segmentation algorithms, enabling SIDEST to surpass classical object-based image analysis techniques. On the public dataset, SIDEST also exhibited an increase of 8.1 % in IoU over advanced deep learning-based models for crop field recognition. Ultimately, county-wide (IoU = 0.87, F1-score = 0.93) and provincial maps (IoU = 0.79, F1-score = 0.88) of crop fields were generated with SIDEST from PS and S2 imagery, respectively. SIDEST has the potential to become a valuable and cost-effectiv