Rice Extraction in Sanjiang Plain Based on Remote Sensing Data and Optimal Feature Spaces

Rice is one of the main food crops in China, Timely and accurate acquisition of the spatial distribution of rice is of guiding significance for food security and agricultural development. In this study, a total of 40 candidate features in five categories: spectral, tasseled cap transformation, texture, terrain, and radar were constructed based on Sentinel-1, Sentinel-2, and terrain data, respectively. The Pearson-Relief_F algorithm was used for feature selection and weight evaluation, and was combined with random forest, support vector machine, and classification regression tree for comparison experiments to filter the optimal feature combination and optimal classification model according to the classification effect. The results show that, 13 optimal feature spaces such as red edge normalized vegetation index (NDVIre1) and (MTCI), combined with random forest features have obvious advantages for the identification of rice growing areas. They can accurately extract the spatial distribution information of rice in the Sanjiang Plain, with the overall accuracy of 95.71% and the Kappa coefficient of 0.95 by ten-fold cross-validation. The accuracy of the selected 13 optimal features is slightly higher than that of all 40 features, with a 67.5% reduction in feature dimensionality. The study can provide a reference for the accurate extraction of rice areas in large regions.