Identification of Rice in Southern Mountainous Area Based on Object-Oriented and Machine Learning Methods

WANG Yingying; DUAN Liangxia; ZHAO Yining; SUN Guangrui; YANG Lihong; ZHOU Qing; XIE Hongxia

doi:10.15933/j.cnki.1004-3268.2025.04.015

Abstract：The phenology information of rice varies with terrain.In view of the cloudy and rainy conditions and complex terrain in the southern mountainous areas,it is of certain research value to find a remote sensing identification method for rice.With Yongshun County,a typical mountainous terrain in Hunan Province,as the research area,combined with Sentinel-1 SAR and Sentinel-2 MSI data,the time series curve was generated by the images of four key phenology periods of rice to grasp the growth trend of rice.Firstly,the object-oriented method was used to segment the images of transplanting period and harvesting period;secondly,the feature variables optimized by the feature space optimization algorithm(FSO)were input into four models for classification such as random forest(RF);finally,the results were compared.Otherwise,according to the transplanting order,the rice samples were divided into early moving and late moving samples,the images of transplanting period and harvesting period were reclassified to explore the influence of rice transplanting time on the classification of the two images.The results showed that,compared to imagery from the transplanting stage,imagery from the harvesting stage offered better classification accuracy and was more suitable for rice mapping.The object-oriented FSO-RF model achieved the best classification results for imagery during the harvest period,with an overall accuracy of 93.19%and a kappa coefficient of 0.901.The time of rice transplanting had little effect on the image classification at harvest stage,but had a great influence on the image classification at transplanting stage.The reason was that the earliest transplanted rice was easy to be similar to the spectral characteristics and texture characteristics of some dry land crops and woodlands,and there was a phenomenon of confusion and mis-classification.In order to improve the recognition accuracy of the image during the transplanting period,it is necessary to improve the time resolution of the image,add more texture features or mask the ground objects.