An effective corn and soybean mapping model integrating phenological and biophysical information based on seasonal median composite satellite imagery

Accurate and timely monitoring of corn and soybean distribution is crucial for ensuring food security, nutritional diets, and environmental conservation. While crop-specific classifiers have garnered considerable attention, distinguishing between corn and soybean remains challenging due to their similar phenological trajectories, spectral characteristics, and the lack of effective indicative features. Efforts to improve their separability typically rely on integrating multi-source data, which incur substantial temporal and economic costs, and often falls short of meeting the demands of large-scale, time-sensitive applications. To address these limitations, we present a highly-effective model for mapping corn and soybean that leverages two robust time-series-independent crop indices, derived from a seasonal median composite Sentinel-2 imagery, that integrate phenological and biophysical characteristics. We first develop an integrative index to identify corn by capturing carotenoid accumulation patterns across key phenological stages. Subsequently, we construct an unsupervised classifier that synergistically integrates the discriminative features of the proposed corn index and a soybean-specific mapping index during their peak growth season. We evaluate the model across five environmentally diverse study sites (A–E) spanning three major agricultural countries. Applying a grid search method in site A, we identify two optimal thresholds, which are then applied to all study sites without site-specific calibration. The model achieves an average overall accuracy (OA) of 88.90% and F1 score of 0.88., outperforming a random forest classifier with an average OA of 82.77% and F1 score of 0.80. Threshold sensitivity analysis indicates the optimal thresholds for each study site yields an average OA gain of 0.49% compared with the applied thresholds, demonstrating the model's robustness to threshold variations. Our results highlight the potential of efficient, crop-specific indices to enable low data-dependent classification frameworks for large-scale, time-sensitive crop mapping.