In order to solve the problem that the ability of panicle recognition by each channel or index of digital image color space is not clear, in rice yield estimation based on UAV image, an effective panicle characteristicselecting method was developed. The field experimental data were collected from super rice achievement transformation base of Shenyang Agricultural University in 2017 and 2018, including highresolution digital image collected with UAV and the number of panicles in each sampling square in rice plots. In order to identify the panicle recognition ability of channels or index in the RGB and HSV color space, a triclassification image sample library of rice panicle, leaf and background was firstly constructed, and features extraction was performed by using the best subset selection (BSS) algorithm. The BSS extracted the seven characteristic parameters which were suitable for panicle segmentation of japonica rice in Northeast China, and used as input to panicle segmentation model based on BP neural network. The recognized panicle pixels from segmentation model were clustered by connected component analysis and the number in each sampling square was estimated, which can be compared with field measurement results for quantitively error analyzing. The results showed that the best subset selection based feature extraction performed best when the number of the feature was 7 (features were R,B,H,S,V,GLI and ExG, respectively), and the latitude was 3m. The corresponding minimum MSE of cross validation is 0.0363. The rice panicle segmentation model can effectively achieve the extraction of japonica rice panicle in Northeast China, with the average RMSE and MAPE of rice panicle number extraction in three flight altitude images taken by 3m, 6m and 9m were 9.03 and 10.60%, 11.21 and 14.88%, 13.10 and 17.16%, respectively.