Abstract: In semantic segmentation of remote sensing imagery based on full convolutional neural network （FCN）, the upsampling and downsampling mechanism in FCN led to the loss of edge details of ground objects in segmentation results. To solve this problem, a land cover classification method based on object-oriented segmentation results to optimize FCN classification of high-resolution remote sensing images was proposed. Firstly, the high-resolution remote sensing imagery was initially classified by FCN model. Then, object-oriented segmentation results were used to optimize the initial classification results obtained based on FCN. Experimental results showed that the method in our paper could not only retain the details of the ground object edge effectively, but also eliminate the phenomenon of salt and pepper in the initial extraction results of FCN. Finally, the visual effect of the classification results was optimized and the classification accuracy was improved.

Key words: object oriented, full convolutional neural networks, high resolution remote sensing image, land cover classification