Abstract: High-resolution remote sensing imagery provides detailed spatial information for farmland change monitoring. However, crop phenological differences, background texture interference, and fragmented plot boundaries often cause false changes, blurred boundaries, and missed detection of small patches. To address these issues, this paper proposes an edge-aware gated learning network, EGF-Net, for high-resolution farmland change detection. The method adopts a MixVision Transformer (MiT) with hierarchical multi-scale representation capability as a shared-weight Siamese encoder backbone to extract contextual semantics and spatial details from bi-temporal images. In the feature fusion stage, a difference-enhanced channel attention neck (DECA) is designed. It constructs bi-temporal difference features through a Concat-Diff strategy, then performs channel attention recalibration to highlight true change responses and suppress non-target interference caused by illumination, soil moisture, and phenological variation. In the decoding stage, an edge-gated context head is constructed. It extracts shallow edge cues from the encoder and introduces them into high-level semantic reconstruction through a gating mechanism, so that boundary localization and structural continuity of changed parcels can be improved. The overall loss combines cross-entropy, Dice, and edge auxiliary losses to alleviate class imbalance and strengthen boundary supervision. Experiments are conducted on the Anyi County field-measured farmland dataset with 0.2 m UAV DOM imagery and the PX-CLCD public farmland change detection dataset with 1 m satellite imagery. The Anyi dataset evaluates the ability to identify fine parcel boundaries and local false changes, while PX-CLCD examines stability under different regional backgrounds and resolutions. Evaluation metrics include Precision, Recall, F-score, IoU, and boundary F1 (B-F1). Quantitative results show that EGF-Net achieves the best performance on both datasets. On the Anyi County dataset, the F-score, Precision, Recall, and IoU reach 98.36%, 98.59%, 98.13%, and 96.78%, respectively. Compared with the MiT baseline, the F-score and IoU are improved by 1.48 and 2.82 percentage points. The B-F1 reaches 65.64%, which is 6.97 percentage points higher than that of MiT, indicating better boundary fitting for fragmented plots. On the PX-CLCD dataset, the F-score, Precision, Recall, and IoU reach 93.65%, 94.23%, 93.07%, and 88.05%, respectively. Compared with MiT, the F-score and IoU are improved by 4.62 and 7.83 percentage points. Ablation experiments verify the effectiveness and complementarity of the proposed modules. Adding the EDGE module improves the F-score and IoU to 97.50% and 95.13%, respectively. Adding the DECA module alone improves them to 97.74% and 95.58%. When both modules are used, the model achieves the best results, with F-score, Precision, Recall, and IoU of 98.36%, 98.59%, 98.13%, and 96.78%, respectively. This demonstrates that DECA mainly enhances change discrimination and suppresses background interference, while EDGE mainly constrains boundary recovery and reduces edge fragmentation. The proposed method also reduces false positives caused by shadows and texture variations. However, several limitations should be noted. First, the original dataset division used 25% overlapping sliding windows and random sample splitting, which may cause spatial autocorrelation and information leakage between training and test samples. Therefore, the reported accuracy may overestimate the model’s generalization ability to unseen regions. Second, the two datasets are independently trained and tested, rather than strictly cross-regional transfer experiments. Third, only two optical images are used, without multi-season, multi-crop, or multi-source data; thus, seasonal tillage changes and permanent cropland changes may still be confused. Fourth, the computational cost and deployment efficiency are not systematically analyzed. Future work will adopt spatially independent dataset partitioning with buffer zones, introduce multi-season and multi-source remote sensing data, develop domain adaptation strategies, and further evaluate the model’s efficiency and uncertainty. The proposed method can provide a methodological reference for subsequent research on spatially independent partitioning and cross-regional transfer.