Semisupervised Building Damage Assessment in Polarimetric SAR Imagery via a Graph-Enhanced Network

Polarimetric Synthetic Aperture Radar (PolSAR) holds significant potential for post-disaster building damage assessment. However, conventional methods often rely on simple change indices that neglect feature correlations and employ pixel-level classification models that lack spatial contextual information, leading to limited accuracy and insufficient robustness. To address these limitations, this paper proposes a semi-supervised building damage assessment framework that enhances performance by fusing multi-level change features with cross-graph spatial contextual information. The framework first constructs a more comprehensive multi-level change feature input that is highly sensitive to building damage. Second, a lightweight network, GESNet, centered around a Graph Context Module (GCM), is designed. It overcomes the receptive field limitations of traditional convolutional networks by constructing graphs across adjacent image patches and propagating contextual information. Furthermore, this study establishes an iterative semi-supervised learning procedure by integrating Fuzzy C-Means (FCM) clustering with a self-training strategy to effectively mine supervisory information from unlabeled data. Experimental results on ALOS-2 and Sentinel-1 datasets demonstrate that the proposed framework significantly outperforms traditional methods, supervised baselines, and other state-of-the-art semi-supervised models across all evaluation metrics. Compared to the next-best semi-supervised model, our method achieves a 5.3% improvement in Balanced Accuracy (BA) and a 0.051 increase in the Kappa coefficient, while also exhibiting enhanced robustness against noise and variations in the number of training samples.