Semi-Supervised Approach: ClusterFormer for Semantic Segmentation in Remote Sensing Applications

Abstract

Applications of remote sensing have been transformed by deep learning, particularly in semantic segmentation for mapping land use and land cover (LULC). Conventional supervised algorithms, on the other hand, spend a lot of money and time since they require a large amount of labeled data. For semantic segmentation in remote sensing applications, this paper investigates a semi-supervised learning paradigm that makes use of ClusterFormer, a transformer-based architecture intended for effective clustering and attention mechanisms. The suggested method attempts to improve segmentation accuracy while lowering reliance on ground-truth labels by merging a sizable corpus of unlabeled satellite data with a small number of annotated samples. Benchmark LULC datasets and custom high-resolution satellite images are used for experiments, while F1-score and Intersection over Union (IoU) are used as assessment measures. Results demonstrate that the semi-supervised ClusterFormer performs similarly to fully supervised baselines under limited-label conditions, highlighting its potential for scalable remote sensing analysis.