Implementation of Partial Cross Entropy Loss for Point-supervised Remote Sensing Image Segmentation

Abstract

Weakly supervised semantic segmentation has emerged as a promising approach for reducing the annotation burden associated with dense pixel-level labelling. Among weak supervision strategies, point supervision offers an attractive compromise by requiring only a small subset of labelled pixels while preserving meaningful semantic information. However, effective optimisation under sparse supervision requires loss functions capable of excluding unlabelled regions from the training process. This study presents an implementation framework for point-supervised remote sensing image segmentation using Partial Cross Entropy Loss. Dense segmentation masks obtained from the LoveDA dataset were converted into sparse point annotations through random pixel sampling with a point ratio of 1%. A subset containing 200 image-mask pairs was constructed and partitioned into training and validation sets. A U-Net architecture with a ResNet-34 encoder was employed as the segmentation backbone, while Partial Cross Entropy Loss was implemented using the ignore-index mechanism available in PyTorch to restrict optimization to labelled pixels only. Functional validation was performed through forward propagation, loss computation, and gradient backpropagation. Successful parameter updates and finite loss values confirmed the correct integration of sparse point supervision with the encoder-decoder segmentation network. The proposed framework provides a simple, modular, and reproducible implementation for point-supervised semantic segmentation and establishes a foundation for future investigations involving weakly supervised and semi-supervised remote sensing image analysis.