On the Spectral–Phenological Features for Crop Mapping Under Complex Planting Patterns: A Case Study in Jiangsu Province, China

Accurate crop mapping in fragmented agricultural landscapes is challenged by overlapping crop calendars and redundancy among multi-source time-series variables. Using Sentinel-1/2 imagery from December 2022 to December 2023, we constructed 275 season-specific spectral–phenological feature–month variables (125 for summer crops and 150 for winter crops) for rice, maize, soybean, winter wheat, and winter rapeseed in Jiangsu Province, China. An auxiliary binary Random Forest (RF) was used to estimate out-of-bag (OOB) permutation-based predictive contributions and construct search priors. A prior-guided genetic algorithm (GA) then identified compact subsets, with crop-specific five-class RF models used both to evaluate candidate subsets and to produce the final classifications. A fixed stratified 80/20 development–validation split was maintained throughout the analysis, with the validation subset reserved for final assessment. August and April were the principal discriminative periods for summer and winter crops, respectively, while VH backscatter and SWIR-related indices, particularly STI and NDTI, showed recurrent predictive contributions across crops. On the independent validation subset, the optical/vegetation-index scheme, SAR-only scheme, and the complete feature library achieved mean target-crop F1-scores of 78.42%, 83.74%, and 86.96%, respectively. The GA-selected subsets retained 9–39 variables and achieved a mean five-class overall accuracy of 91.77% and a mean target-crop F1-score of 93.95%. After non-target classes were merged into a single background class, the integrated seasonal maps achieved overall accuracies of 81.20–95.03% on the same validation subset. Supplementary classifier comparisons indicated that subset effects depended on the crop and learning algorithm. The findings support crop-specific, interpretable dimensionality reduction within the RF workflow, while broader transferability requires multi-year and multi-region evaluation.