Extracting Summer-Harvested Crops in the Baojixia Irrigation District Using CycleGAN and Transfer Learning

Remote sensing-based mapping of crop planting structures in irrigation districts plays a vital role in forecasting regional production and optimizing water resource allocation. However, optical satellite imagery is limited by insufficient spatial resolution when applied to fragmented farmland landscapes, while unmanned aerial vehicle (UAV) imagery is limited by spatial coverage and high data processing costs. To address these bottlenecks, this study proposed a cross-scale collaborative extraction framework utilizing the CycleGAN network and transfer learning to map summer harvest crops (winter wheat and rapeseed) in the Baojixia Irrigation District for the year 2023. First, multiple semantic segmentation models—including U-Net, DeepLabv3+, SegFormer, and HRNet—were evaluated on a joint satellite–UAV dataset, with U-Net selected as the optimal backbone. Next, CycleGAN was introduced to perform style translation from the UAV domain to the satellite domain. This step generated high-fidelity, satellite-like images that preserve UAV-derived high-resolution spatial details, which were subsequently used to pre-train the U-Net backbone, significantly reducing the labor of manual annotation. Finally, the model was fine-tuned with real satellite images to achieve precise crop extraction. Results indicated that this framework accelerates model convergence and improves segmentation accuracy. The proposed method achieved an mIoU of 85.09%, an mPA (Recall) of 91.63%, a Precision of 91.97%, an Accuracy of 93.57%, and an F1-Score of 91.80%, outperforming the baseline U-Net model by 2.98%, 2.00%, 1.66%, 1.52%, and 1.83%, respectively. By successfully transferring high-resolution prior knowledge into the satellite feature space, this study provides a cost-effective and highly accurate solution for crop identification in complex agricultural landscapes, breaking the spatial limitations of UAV remote sensing.