Enhanced landslide susceptibility prediction with 3D ALOS PALSAR imagery and neural networks: A data-efficient framework

Abstract

Landslide susceptibility mapping (LSM) founded on DEM is a growing research field with profound implications for human safety and infrastructure preservation. Many existing methods rely on extensive input data to enhance predictive accuracy. This paper aims to introduce a remote sensing-data-requirement framework for LSM. Our approach exclusively leverages a single ALOS PALSAR image, comprising three key steps: (1) Pre-processing, (2) derivation of explanatory variables, and (3) neural network modeling. To begin, we extracted 22 input variables from the ALOS PALSAR image. These variables played a pivotal role in developing the Neural Network (NN) predictor. The predictor structure consists of 22 variables in the input layer, 150 neurons in the hidden layer, and a single output layer. Our model was trained using 5,829 sample points, and subsequently, it was employed to generate landslide susceptibility (LS) map with 745,810 points. Based on the Overall accuracy metric, the model exhibited impressive performance accuracy, achieving 89.3% training and 82.3% testing accuracies. Additionally, it demonstrated a strong performance of 95.22% during training and 84.7% during testing according to the ROC curve. In conclusion, the implementation of our proposed method underscores its ability to develop remarkable accuracy model with remote sensing-data-requirement. This framework offers valuable insights for future progress in regions with challenging conditions and extensive data coverage. Moreover, it effectively handles data quality inconsistencies and data updating issues.