Noise-Aware Regional Contrast Semantic Segmentation of Remote Sensing Images Using Crowdsourcing Labels

Remote sensing semantic segmentation enables automated identification and monitoring of land cover categories in remote sensing images, providing precise geographic information support for urban planning, resource management, environmental protection, and broader geographical and environmental research. With the advancement of deep learning, methods based on convolutional neural networks (CNNs) have significantly improved pixel-level automatic classification accuracy in remote sensing image semantic segmentation. However, this progress relies on extensive, high-quality fine annotations, which are costly and time-consuming to acquire. To address these challenges, we propose utilizing easily accessible and cost-effective crowdsourced data from OpenStreetMap (OSM) to generate learning labels for network training. However, due to the open nature of crowdsourced data, the generated labels may contain noise interference, leading to decreased network performance during training. Meanwhile, supervised contrastive learning has shown strong potential in handling noisy labels, with previous research proposing selective contrastive learning strategies to address noisy labels in classification tasks. However, such research has been limited to classification, as segmentation poses memory explosion issues. To further tackle this problem, we introduce a semantic segmentation algorithm based on crowdsourced annotation and selective contrastive learning. The core idea involves selecting confident regions in the regional dimension, allowing pixel samples within these regions to participate in subsequent learning, thereby better capturing meaningful pixel features for classification while reducing sample quantity. By leveraging contrastive learning ofpaired features and selecting confident pairs from noise pairs for supervised contrastive learning, we enhance intra-class compactness, thereby improving network robustness and reducing noise interference. Experimental validation on the proposed crowdsourced semantic segmentation datasets demonstrates favorable outcomes.