How Much Temporal Information Is Needed? Optimizing Sentinel-2 Time Series for Operational Rice Crop Mapping

Rice crop mapping is essential for precision agriculture, irrigation management, yield forecasting, and assessment of the environmental impacts of rice cultivation. This information is particularly relevant in Mediterranean rice-growing wetlands, such as L’Albufera Natural Park (Valencia, Spain), where crop development is closely linked to seasonal flooding and water-management practices. Although multitemporal Sentinel-2 imagery is widely used for crop classification, the temporal information required to achieve accurate early-season mapping remains insufficiently investigated. This study evaluates the optimal temporal window and spectral variables for operational rice mapping in L’Albufera using Sentinel-2 imagery acquired during the 2023 and 2024 growing seasons. Monthly reflectance composites from visible and near-infrared bands (B2, B3, B4, and B8) and vegetation indices (NDVI, NDWI, EVI, and RVI) were analyzed through temporal profiling, statistical analysis, principal component analysis, and supervised machine-learning classification. Within the classifiers evaluated, performance gains were associated more closely with temporal and spectral feature selection than with classifier family. The strongest performance was achieved using the NDVI, EVI, B3, and B8 from April through June (on the 2023 holdout set, overall accuracy [OA] = 83.8% and rice F1 score = 0.90; in the 2024 cross-year evaluation, OA = 76.5% and rice F1 score = 0.86). Model-agnostic permutation importance and Shapley values were used to identify the most influential predictors, yielding a compact six-predictor set comprising the NDVI from April to June, B8 from May to June, and B3 in June. This reduced six-predictor model retained similar performance in the 2023 holdout set (OA = 82.4%; rice F1 score = 0.89) and in the 2024 cross-year evaluation (OA = 75.4%; rice F1 score = 0.85). These results show that rice fields can be mapped before crop maturity within the study area while reducing temporal and predictor requirements.