CNN-Based Model for Landslide Susceptibility Assessment from Multispectral Data

Abstract

In this work, a new convolutional neural network architecture is proposed to evaluate the susceptibility to landslides. It is a supervised learning algorithm that has been trained from data whose labels have been obtained by applying a heuristic method that involves geological, geomorphological and land use information. The attributes contemplated the use of multispectral data and spectral indices, in addition to slope and DEM data. Although the cartographic unit in the proposed method is the pixel, the processing was performed at the patch level since it involved the use of spatial information around each pixel. Therefore, the proposed deep learning architecture is characterized by its simplicity and by applying both spatial and channel processing. The proposed method presents similar performance to state-of-the-art methods, achieving an F1 score higher than 88% on test data with low computational cost and pixel-level accuracy.