Landslide Detection in Google Earth Engine Using Deep Learning Methods

Rapid mapping of landslides by deep learning (DL) methods using high resolution satellite images had proven to be effective. Still, the acquisition and pre-processing of satellite images to be used in DL methods are often time consuming. Further, the transferability of DL methods from the images where the model was trained to the other regions also reduces the model's overall accuracy. To address these issues, we have used Google Earth Engine (GEE) modules to pre-process the satellite images and their derivate. We have used them as input layers to the U-Net and ResU-Net DL models in Python script on Google Colaboratory. To increase the transferability of the model while improving the accuracy for the training area, we have used change detection-based input layers such as NDVI difference, the spectral angle between pre- and post-landslide satellite images and PC-4 band of change detection-based PCA after applying radiometric normalisation. The model was trained in the Kodagu region (Karnataka) and tested in the Puthumala Region (Kerala). The metrics evaluated for the U-Net and ResU-Net models have shown an improved performance of the ResU-Net model by 3% precision, 4.8% recall, 2.7% F1-score and 2.8% Mathews Correlation Coefficient. Through the ResU-Net model, we achieved the maximum recall value of 93.7% for the Kodagu region. The radiometric normalization applied on the testing area based on the training area was found to be effective by equating similar spectral signatures of surface features, thereby increasing the overall accuracy. The ResU-Net model was additionally tested in three diverse regions with reported and un-reported landslide such as the Taliye Village region, Raigad District (Maharashtra), Pandali Village region, Mandi District (Himachal Pradesh) and a no landslide region near Kodagu area (Karnataka), respectively. The ResU-Net model provided similar accuracies for these areas, thus validating the model’s transferability. The pre-processed layers of GEE helped to reduce the processing time, allowing multiple test runs with different training areas and input layers.