Combining spatial response features and machine learning classifiers for landslide susceptibility mapping

Abstract

Reliable landslide susceptibility mapping (LSM) is essential for disaster prevention and mitigation. This study develops a deep learning framework that integrates spatial response features and machine learning classifiers (SR-ML). The method has three steps. First, depthwise separable convolution (DSC) extracts spatial features to prevent confusion of multi-factor features. Second, spatial pyramid pooling (SPP) extracts response features to obtain features under different scales. Third, the high-level features are fused into prepared ML classifiers for more effective feature classification. This framework effectively extracts and uses different-dimension features of samples, explores ML classifiers for beneficial feature classification, and breaks through the limitation of fixed input sample sizes. In the Yarlung Zangbo Grand Canyon region, data on 203 landslides and 11 conditioning factors were prepared for availability verification and LSM. The evaluation indicated that the area under the receiver operating characteristic curve (AUC) for the proposed SR and SR-ML achieved 0.920 and 0.910, which were 6.6% and 5.6% higher than the random forest (RF, with the highest AUC in ML group) method, respectively. Furthermore, the framework using 64×64 size inputs had the lowest mean error of 0.01, revealing that samples considering landslide scales could improve performance for LSM.

Keywords

• Landslide susceptibility
• Feature fusion
• Hybrid models
• Machine learning