Research on Unbalanced Sample Segmentation of Remote Sensing Image

Abstract This article proposes a framework for the unbalanced sample segmentation of remote sensing images on an open data set. For the few sample categories, improved by converting it into binary segmentation and combining with weighted cross-entropy loss, and then merged with the segmentation result of the sufficient sample categories. Finally, the mIoU (mean Intersection of Union) of the 8 categories is increased by 4.5% compared to the results before the improvement, especially, the results of the few sample category road and grassland are increased by 10.2% and 9.7%. The experiments show that the framework can greatly improve the segmentation performance of the few sample categories, and have a good guiding significance for the problem of multiclass unbalanced sample segmentation.