Using Weight Reliability Masks on Imbalanced Datasets for Satellite Image Segmentation

This article addresses the issue of imbalanced datasets in satellite image segmentation problems, where machine learning models often neglect minority classes in favor of majority ones. We propose using spatial weight masks for the loss function computation taking into account the reliability index of individual pixels. This approach enhances segmentation quality, significantly improving metrics for minority classes. Additionally, a dataset expansion method using generative adversarial networks (GANs) is explored, showing slight improvements in recognizing less represented crop types in the dataset, and it is compared with the proposed method. The simultaneous usage of weight masks and generative networks is investigated.