Super-Resolution Assisted Farmland Boundary Extraction from Medium-Resolution Satellite Image: A Real-ESRGAN and YOLO Segmentation Framework

The study addresses the issue of insufficient spatial resolution in remote sensing images for farmland boundary identification in precision agriculture. It proposes an innovative framework based on Real-ESRGAN super-resolution reconstruction and YOLO instance segmentation, as well as a method for farmland boundary extraction that combines super-resolution reconstruction with deep learning segmentation, to improve the accuracy of farmland identification in medium-resolution satellite images. Taking the agricultural area of Nanxiong City, Guangdong Province, as the study region, the study constructed a manually annotated farmland boundary dataset. The Real-ESRGAN model was employed to perform blind super-resolution reconstruction on the GF-2 satellite image, and the YOLO instance segmentation model was used for farmland boundary extraction. The results indicate that after performing blind super-resolution reconstruction of the GF-2 satellite image using the Real-ESRGAN model, the spatial resolution of the GF-2 satellite image was improved from 4 m to 1 m. By simulating real-world complex degradation through a high-fidelity degradation model, Real-ESRGAN significantly enhances robustness against various practical degradations in remote sensing images, reconstructing a high-quality GF-2 satellite image rich in textural details. The texture and boundary details of the GF-2 satellite image are significantly enhanced, effectively mitigating field merging and boundary discontinuities caused by aliasing effects. After extracting farmland boundaries using a YOLO instance segmentation model based on a multi-task architecture, the super-resolution reconstructed images achieved an average PSNR of 26.19 dB and an average SSIM of 0.7676. In the farmland boundary recognition task, the values of mAP@0.5 were between 0.70 and 0.75, the values of mAP@0.5:0.95 were between 0.55 and 0.60, and training and validation losses were around 2.00 and 2.50, respectively. The validation results indicated that the model did not overfit and possessed good generalization ability. A comparison of the models revealed that the Real-ESRGAN super-resolution model outperforms the EDSR+OpenCV super-resolution model both numerically and visually. The study adopts the highest boundary recognition accuracy achieved on the validation set as the final performance metric for the model. The framework provides a cost-effective solution for precise farmland boundary recognition. Its backbone network is tailored to super-resolved images; the Neck module enhances cross-scale boundary responses, and the segmentation head accurately models sub-pixel-level geometric topology. Super-resolution technology effectively enhances the spatial information representation of medium-resolution remote sensing images, offering a low-cost solution for large-scale farmland boundary extraction and providing strong support for precision agricultural management.