Phenology-Guided Weakly Supervised Cropping Structure Mapping with Phenological Similarity Constraints

To address the challenges of cumulative pseudo-label noise and substantial phenological differences across regions in cropping structure mapping, this study develops a phenological knowledge-guided weakly supervised semantic segmentation framework, termed PhenoStruct-WSF. The framework first constructs typical phenological curves for major annual cropping patterns and uses TWDTW to measure the phenological similarity between samples and templates. A two-class margin indicator, ∆i, is then proposed, and phenological confidence is derived through probability calibration, after which high-confidence samples are selected as anchors under posterior-FDR control. Subsequently, the anchors are embedded into a structure-aware weakly supervised segmentation network, PhenoStruct-Net, forming a two-stage training strategy of “fidelity first, expansion later”, in which robust pseudo-label expansion is achieved through phenological confidence gating and adaptive weighting. Experiments were conducted in three representative study areas, namely Handan in Hebei Province, Jingzhou in Hubei Province, and Siping in Jilin Province. The anchor samples selected using posterior-FDR achieved accuracies above 94% in all three study areas. PhenoStruct-WSF consistently achieved overall accuracy (OA), mean intersection over union (mIoU), and macro F1-score (mF1) above 89%, 81%, and 89%, respectively, across the three regions, with an average OA improvement of 4.52 percentage points over the comparison methods. Sample sensitivity experiments indicate that the proposed method maintains a stable trend of accuracy improvement under different training sample sizes, demonstrating good adaptability to limited samples. The ablation experiments further show that phenological confidence anchors, posterior-FDR control, and the confidence-weighting mechanism can effectively suppress pseudo-label noise propagation and improve the robustness of crop mapping in complex agricultural landscapes.