In order to improve the utilization of spatial-spectral features in hyperspectral image classification algorithms, this paper investigates an image semantic segmentation technique based on machine learning. The method consists of randomly sampling a window from the training image, using the comparison of the values of two pixels in the window to generate a feature vector, and using it to train a random forest classifier. In the testing stage, the category of each pixel is determined by voting on the leaf nodes of the random forest. The experimental results show that the average classification accuracy of the improved algorithm is improved from 0.3006 to 0.3295, and the global classification accuracy is improved from 0.4430 to 0. 4909. It is concluded that the proposed improved algorithm performs well in semantic image segmentation.