NeighborLoss: A Loss Function Considering Spatial Correlation for Semantic Segmentation of Remote Sensing Image

House segmentation of remote sensing image based on deep learning has become the main method because it can automatically extract features. However, accuracy is affected not only by network model, but also loss function, existing functions, except Binary Cross Entropy, are designed to deal with imbalanced dataset, no new research improving Entropy for balanced and all function treat each pixel in isolation, without considering spatial correlation between its neighbor pixels. To solve this problem, a named NeighborLoss proposed. Firstly, used get prediction results pixel. According whether eight neighboring pixels consistent prediction, different weights given Finally, weighted average value cross entropy batch taken as final value. We use semantic networks SegNet, PSPNet, UNET ++, MUNet both respectively houses open data set WHU dataset sensing. The show that compared MIoU, Precision, Recall, Accuracy improved. From predicted graph, more accurate edge house, especially corner house. effective segmentation.