Abstract: Taking Sentinel-2 multispectral image of Zhutuo, Yongchuan district of Chongqing as an example,a random forest classification model was constructed, and rice spatial distribution was extracted using single-time and multi-time feature variable sets as variables,and the response degree and extraction accuracy of rice to different spectral feature sets were analyzed. The results show that the rice distribution in the study area is relatively scattered and complex,which is basically compatible with the typical landform of the region. Rice at tillering stage have more distinguishing characteristics than that at filling stage,and the extraction accuracy can be effectively improved by using multi-temporal data. Traditional MLC and SAM classifiers have limited precision in extracting ground objects,while remote sensing model based on machine intelligence classification has an overall precision of more than 90% and kappa coefficient of more than 0.80,which provides a reference for crop information extraction in southwest mountainous areas.

Key words: random forest, rice extraction, remote sensing, southwest region