Landslide detection using visualization techniques for deep convolutional neural network models


HACIEFENDİOĞLU K. , Demir G., BAŞAĞA H. B.

NATURAL HAZARDS, vol.109, no.1, pp.329-350, 2021 (Journal Indexed in SCI) identifier identifier

  • Publication Type: Article / Article
  • Volume: 109 Issue: 1
  • Publication Date: 2021
  • Doi Number: 10.1007/s11069-021-04838-y
  • Title of Journal : NATURAL HAZARDS
  • Page Numbers: pp.329-350
  • Keywords: Deep learning method, Convolutional neural networks, VGG-19, Resnet-50, Inception-V3, GradCAM, ScoreCAM, Landslide, INSAR, AREAS

Abstract

Landslides occur when masses of rock, earth, and other debris move down a slope. The slope of an area is directly responsible for the magnitude of the landslide. Being able to identify regional locations more likely to be impacted by landslides is essential if interventions to prevent loss of life and liberty are to be implemented. To further this objective, studies have been carried out using deep learning methods to assess the likelihood of landslides in a localized area. This study seeks to illuminate the reliability in using the deep learning method to effectively detect landslide zones for the purpose of enacting proactive interventions. Pre-trained models of Resnet-50, VGG-19, Inception-V3, and Xception, all of which are established deep learning approaches, were used to automatically detect regional landslides and then compare results. In addition, Grad-CAM, Grad-CAM + + , and Score-CAM visualization techniques were considered depending on the deep learning methods used to accurately predict the location of landslides. The present research focuses on the landslides that took place in the Gundogdu area of Rize, a city on the Black Sea cost of Turkey, from August 26 to 27, 2010, where unfortunately a significant number of individuals lost their lives. As a large number of landslide scene images are needed in order to facilitate a learning model's deep learning, images from the area were obtained by aircraft and then organized as a dataset. Non-landslide scenes were added as a separate class in the training dataset to estimate the landslide regions more accurately. In total, 80% of the data will be used for training the model, while 20% will be used for testing the model that is built out of it. The experimental results were evaluated with the receiver operating curves and f1-score applicable to landslide detection characteristics. Obtained results show that both Resnet-50 and VGG-19 had a success rate of over 90%. Results also effectively demonstrate how the best visualization techniques for localizations are Grad-CAM and Score-CAM.