The accurate prediction of landslide susceptibility relies on effectively handling landslide absence samples in machine learning (ML) models. However, existing research tends to generate these samples in feature space, posing challenges in field validation, or using physics-informed models, thereby limiting their applicability. The rapid progress of interferometric synthetic aperture radar (InSAR) technology may bridge this gap by offering satellite images with extensive area coverage and precise surface deformation measurements at millimeter scales. Here, we propose an InSAR-based sampling strategy to generate absence samples for landslide susceptibility mapping in the Badong–Zigui area near the Three Gorges Reservoir, China. We achieve this by employing a Small Baseline Subset (SBAS) InSAR to generate the annual average ground deformation. Subsequently, we select absence samples from slopes with very slow deformation. Logistic regression, support vector machine, and random forest models demonstrate improvement when using InSAR-based absence samples, indicating enhanced accuracy in reflecting non-landslide conditions. Furthermore, we compare different integration methods to integrate InSAR into ML models, including absence sampling, joint training, overlay weights, and their combination, finding that utilizing all three methods simultaneously optimally improves landslide susceptibility models.
Accurate prediction of landslide susceptibility relies on effectively handling absence samples in data-driven models. This study investigates the influence of different absence sampling methods, including buffer control sampling (BCS), controlled target space exteriorization sampling (CTSES), information value (IV), and mini-batch k-medoids (MBKM), on landslide susceptibility mapping in Songyang County, China, using support vector machines and random forest algorithms. Various evaluation metrics are employed to compare the efficacy of these sampling methods for susceptibility zoning. The results demonstrate that CTSES, IV, and MBKM methods exhibit an expansion of the high susceptibility region (maximum susceptibility mean value reaching 0.87) and divergence in the susceptibility index when extreme absence samples are present, with MBKM showing a comparative advantage (lower susceptibility mean value) compared to the IV model. Building on the strengths of different sampling methods, a novel integrative sampling approach that incorporates multiple existing methods is proposed. The integrative sampling can mitigate negative effects caused by extreme absence samples (susceptibility mean value is approximately 0.5 in the same extreme samples and presence-absence ratio) and obtain significantly better prediction results (AUC = 0.92, KC = 0.73, POA = 2.46 in the best model). Additionally, the mean level of susceptibility is heavily influenced by the proportion of absent samples.
Global climate change has led to the frequent extreme meteorological events in recent years, triggering severe clustered landslides in mountainous regions. Records of these clustered landslides not only provide post-disaster statistics but also play a crucial role in advancing data-driven regional landslide research and intelligent landslide detection. The Rainfall-induced Landslide in Zixing (RLZX) datasets consist of a landslide inventory map (LIM) and a landslide detection dataset (LDD). RLZX-LIM was created through visual interpretation of 3D scenes before and after the rainfall event, containing 19,403 shallow landslides triggered by extreme rainfall in Zixing City, China, between July 26 and July 28, 2024. We have provided quantitative evaluations of the quality of RLZX-LIM based on reference data obtained from road-aligned surveys and unmanned aerial vehicle (UAV) mapping in the field. RLZX-LDD is further developed using both UAV and satellite images, offering higher quality and robustness, effectively filling the gap in rainfall-induced LDDs. The RLZX datasets have been publicly released for free use to promote related landslide research.
Machine learning is pivotal in data-driven landslide susceptibility mapping (LSM). However, the uncertainty of negative samples and the imbalance between positive and negative samples, which leads to misjudgments and overestimation, remain going challenges. This study introduces a novel framework for LSM that integrates positive-unlabeled (PU) learning with imbalanced learning methods, making full and correct use of vast unlabeled samples. First, prior model based on the spy algorithm is generated to obtain reliable negative (RN) samples which is used to create imbalanced training and testing sets. Subsequently, four imbalanced learning models, namely Synthetic Minority Oversampling Technique-Deep Neural Network (SMOTE-DNN), Adaptive Synthetic-Deep Neural Network (ADASYN-DNN), Balanced Random Forest (BRF), and EasyEnsemble (EE) are employed to process the imbalanced training and testing sets and generate the final prediction models. We have tested our LSM framework using a dataset of regional rainfall-induced landslides that occurred in Beijing, China. The positive impacts of RN samples are evaluated using baseline models and extensive saturation tests with various imbalance ratios are conducted. Imbalanced learning methods enhanced prediction for negative classes, with balance peaks observed in the saturation tests. Balanced Random Forest (BRF) showed the best performance and stability across different imbalance ratios. This framework can improve the prediction accuracy for both positive and negative classes, which has the potential to reduce overestimation and misclassification and holds promise for significantly impacting future modelling strategies in LSM.
26 March 2026 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Landslides occur frequently in mountainous regions worldwide, emergency management and risk analysis urgently need automated, accurate landslide detection (LSD) techniques. AI and data-driven methods are becoming mainstream in LSD, but they overly rely on labeled data and generalize poorly across regions. In recent years, foundation models, exemplified by the Segment Anything Model (SAM), have advanced rapidly and significantly aided downstream tasks like LSD. Leveraging SAM, globally distributed landslide big data Landslide71K, and enhanced three-step transfer learning (TL), we propose LSDSAM and a practical framework for few-shot cross-domain LSD that achieves reliable detection with only single-digit labeled samples in target domains. LSDSAM adds a task-specific branch and a CNN branch to inject landslide edge cues and local features into the SAM encoder. It also introduces a multi-level features fusion enhanced decoder to fuse information across layers and improve performance. The three-step TL involves task-adaptive fine-tuning to acquire landslide source domain knowledge from Landslide71K, domain adversarial fine-tuning to reduce domain bias using unlabeled target domain samples, and few-shot target-specific fine-tuning. In evaluations of landslides caused by rainfall and earthquake, as well as those influenced by complex environments, LSDSAM markedly outperforms other state-of-the-art methods. With only eight labeled samples, it attains mean intersection over union (MIOU) of 70.7%, 81.3%, 71.6%, and 76.0% across four testing datasets, demonstrating strong few-shot generalization. The proposed LSDSAM and TL framework provide a practical solution for application-oriented LSD under limited annotation conditions and offer a reference for future LSD studies in the era of large and foundation models.