A Multitemporal Mountain Rice Identification and Extraction Method Based on the Optimal Feature Combination and Machine Learning

Abstract

The quick and precise assessment of rice distribution by remote sensing technology is important for agricultural development. However, mountain rice is limited by the complex terrain, and its distribution is fragmented. Therefore, it is necessary to fully use the abundant spatial, temporal, and spectral information of remote sensing imagery. This study extracted 22 classification features from Sentinel-2 imagery (spectral features, texture features, terrain features, and a custom spectral-spatial feature). A feature selection method based on the optimal extraction period of features (OPFSM) was constructed, and a multitemporal feature combination (MC) was generated based on the separability of different vegetation types in different periods. Finally, the extraction accuracy of MC for mountain rice was explored using Random Forest (RF), CatBoost, and ExtraTrees (ET) machine learning algorithms. The results show that MC improved the overall accuracy (OA) by 3–6% when compared to the feature combinations in each rice growth stage, and by 7–14% when compared to the original images. MC based on the ET classifier (MC-ET) performed the best for rice extraction, with the OA of 86%, Kappa coefficient of 0.81, and F1 score of 0.95 for rice. The study demonstrated that OPFSM could be used as a reference for selecting multitemporal features, and the MC-ET classification scheme has high application potential for mountain rice extraction.