Robust Deep Learning for Accurate Landslide Identification and Prediction

Landslide is the most common natural risk in mountainous regions on all five continents and they can pose a serious threat in these areas. Strong earthquakes, unusual weather events such as storms and eruptions of volcanoes, and human-caused events such as creating roadways that crossed the slopes are the main causes of landslides and they cause significant dangers to residential properties and society as a whole. The Landslide4sense dataset is used for identifying landslides, which contains 3799 training samples and 245 testing samples. These image patches are taken from the Sentinel-2 sensor, while the slope and Digital Elevation Model (DEM) are from the ALOS PALSAR sensor. Data was gathered from four distinct geographical areas namely Kodagu, Iburi, Taiwan, and Gorkha. We use Deep Learning (DL) models such as ResNet18, U-Net, and VGG16 to predict the landslide. By comparing the above models with the evaluation metrics like loss, precision, recall, F1 score and accuracy, ResNet18 model is selected as the best model for landslide identification.