An edge-effect-aware framework for uncertainty-resolved crop mapping and planting-structure monitoring in arid irrigation landscapes

Accurate crop mapping is critical for agricultural water management in the Yellow River Basin under strict water constraints. To address mixed-pixel and noise issues in fragmented landscapes like the Yichang Irrigation Area, this study developed an edge-effect-aware framework. Using Sentinel-2 composites from 2023 to 2025, a 7-dimensional static spectral dataset was constructed. We utilized Optuna optimization to evaluate three deep learning models (SNN, 1D-CNN, LSTM) and Random Forest (RF), coupled with a spatial majority voting filter. Results indicated: (1) The structurally simple SNN demonstrated superior robustness, achieving a peak Overall Accuracy of 0.953 after filtering, outperforming 1D-CNN, RF, and LSTM. By directly mapping static features, SNN avoided the overfitting issues LSTM encountered with non-temporal mixed spectral signals. (2) Spatiotemporal analysis revealed a "sunflower matrix and maize aggregation" pattern. Driven by economic incentives, maize expanded its area proportion from 24% to 30.1%, reflecting trade-offs with water constraints. (3) Boundary-distance analysis confirmed that mixed pixels within the 0–10 m range constitute the primary physical bottleneck limiting accuracy. However, SNN exhibited significantly better stability in these transition zones compared to LSTM. This framework provides uncertainty-resolved crop mapping to support adaptive planting-structure regulation in arid regions.