A Joint Semantic Segmentation Loss Function for Imbalanced Datasets

Semantic segmentation is one of the most important applications in remote sensing image analysis. Since remote sensing datasets are often highly imbalanced in terms of class distribution, specialized loss functions such as focal loss are required. In this paper, a loss function that combines weighted focal loss with Jaccard loss has been developed. This loss function has been used to train U-Net and DeepLabV3+ semantic segmentation models on the recently introduced Land-cover.ai dataset, which has a high level of class imbalance. It has been observed through our experiments that the combined loss function leads to a performance improvement.