A geo-domain constrained similarity framework for sample augmentation in landslide susceptibility mapping

Landslide susceptibility mapping (LSM) commonly relies on the assumption of geographical similarity. However, existing sample augmentation methods assume global similarity and often overlook regional differences in landslide-forming mechanisms, limiting model generalization in spatially heterogeneous mountainous regions. To address this gap, we proposed a Geo-domain Constrained Similarity (GCS) framework and validated it using 336 historical landslides in Zhenxiong County. The framework first partitions the study area into homogeneous geo-domains through spatial clustering and then develops two complementary similarity metrics: SWCFD, which characterizes environmental background similarity, and CF-BE-HMD, which quantifies landslide-forming mechanism similarity. Based on these metrics, three sample augmentation strategies were designed to identify representative training samples from real geographical units rather than synthetic feature-space samples. The results showed that GCS successfully identified geo-domains with distinct controlling factors, supporting the necessity of zonal modeling. The combined strategy increased the Accuracy of both Random Forest and Support Vector Machine models from approximately 73% to over 83%, while improving the AUC from about 0.80 to above 0.92. It also enhanced prediction stability by reducing uncertainty associated with spatially heterogeneous samples. The proposed framework provides an effective sample optimization strategy for LSM, geohazard investigation, and regional landslide risk management in complex mountainous areas.