Parcel boundary detection based on gradient-based feature from UAV images

Farmland parcel boundary information is crucial for precision agriculture. This study proposes an automated farmland parcel boundary extraction method from UAV images, integrating gradient-based feature extraction and Support Vector Machine classification. First, gradient-based features are extracted by analyzing local discontinuities in a key cue brightness across multiple orientations using oriented half-disc histograms. Second, Support Vector Machine is adopted for boundary classification, leveraging its maximum interval property (matching boundary width characteristics) and binary classification capability (distinguishing boundaries from non-boundaries). Experiments are conducted on multitemporal UAV images of the same area (covering farmland and water) with training data from one time phase and test data from two others. Results show pixel-level accuracy exceeded 88% for both test images, while region intersection over Union ranged from 17.04% to 28.65%. This method achieves high pixel-level extraction accuracy, providing a reliable technical approach for large-scale UAV-based farmland parcel boundary extraction and supporting precision agriculture management such as crop mapping and yield estimation.