A self-supervised building extraction method based on multi-modal remote sensing data

This paper addresses challenges in building extraction from remote sensing imagery, including ambiguous edge definition, limited shadow recognition, and heavy reliance on annotated data. To overcome these issues, we propose a self-supervised building extraction method that integrates LiDAR height information with hyperspectral imagery. First, a random forest model selects optimal hyperspectral bands that emphasize building features, reducing dimensionality for efficient processing. Next, we refine the self-supervised learning model Nearest Neighbour based Contrastive Learning Network (NNCNet) into an enhanced version (INNCNet), which performs well in building extraction tasks while minimizing dependence on annotated samples. A connected domain filtering technique is also introduced in the post-processing stage to eliminate misclassifications and noise, improving segmentation accuracy. Evaluation on the Houston2018 dataset demonstrates that the proposed method achieves high accuracy without annotated data, offering a promising approach for large-scale, unsupervised building extraction in remote sensing applications.