Objectives Landslides are common natural disasters that cause significant casualties and economic losses worldwide. The small baseline subset (SBAS) interferometric synthetic aperture radar (InSAR) technique is widely used for the early identification of landslides. However, conventional SBAS-InSAR processing is often compromised by the presence of non-monotonic deformation pixels, which introduce substantial noise and obscure the precise delineation of landslide-prone regions. These non-monotonic deformation signals may arise from atmospheric disturbances, unmodeled topographic errors, or other environmental factors, making it challenging to differentiate actual landslide displacement from noise. Consequently, accurately extracting landslide deformation areas remains a significant challenge in InSAR-based geohazard monitoring.

Methods To overcome this limitation, an advanced landslide deformation identification framework is proposed, integrating statistical filtering and clustering techniques. First, the Spearman rank correlation coefficient is calculated to evaluate the monotonicity of displacement time series at each pixel. A statistical threshold is then applied to remove pixels exhibiting non-monotonic deformation trends, ensuring the retention of only those showing a consistent increase or decrease in displacement over time. This process substantially reduces noise interference while preserving meaningful deformation signals. Subsequently, a hybrid clustering framework is employed, combining the hierarchical navigable small-world (HNSW) network with the density-based spatial clustering of applications with noise (DBSCAN) algorithm. The HNSW network enhances the efficiency of high-dimensional nearest neighbor searches, while DBSCAN effectively clusters pixels with similar deformation characteristics, further mitigating spatial noise. This combined approach facilitates the precise identification of landslide deformation areas while maintaining computational efficiency.

Results Experimental validation confirms that the proposed method eliminates over 96% of background noise, significantly improving the accuracy of landslide detection. The filtering process enhances the signal-to-noise ratio of SBAS-InSAR results, allowing for the reliable identification of deformation zones across diverse topographic and climatic conditions.

Conclusions Additionally, the method demonstrates superior performance in detecting landslides characterized by slow displacement rates or weak signal responses, which are often overlooked by traditional approaches. By improving the precision and reliability of landslide hazard assessment, the proposed methodology offers an effective and scalable solution for large-scale landslide monitoring.