Cross-Regional Classification of Rice Cropping Systems Based on Within-Year Seasonal Composition Using Multimodal Remote-Sensing Time Series

Rice cropping-intensity products indicate the number of rice-growing seasons within a year but cannot reveal the seasonal compositions that define rice cropping systems. This limitation constrains detailed characterization of regional rice production patterns. To address this gap, this study developed a framework for rice cropping-system classification based on combinations of early-, middle-, and late-season rice cultivation. The framework integrates SAR-assisted reconstruction of cloud-contaminated Sentinel-2 time series using Sentinel-1 observations, Phenology-Constrained Time-Series Splicing (PTS) to construct PTS samples for underrepresented rice cropping-system classes, and an Ordered-Cycle Query Transformer (OCQT), which encodes annual-global and chronologically ordered-cycle representations with learnable season-query tokens for eight-class classification. On the spatial-block validation set, OCQT trained using real samples together with PTS samples achieved an overall accuracy of 0.9040, a macro-F1 of 0.8643, and a region-balanced macro-F1 of 0.8275, outperforming OCQT trained using only real samples. Across all four independent regional test sets, this configuration increased weighted-F1 by an average of 3.98 percentage points, with a maximum gain of 9.88 percentage points. This study advances rice monitoring from cropping intensity to explicit characterization of within-year seasonal composition and provides a framework for detailed agricultural monitoring using multi-source remote sensing and ordered-cycle representation learning.