Mapping landslides through a temporal lens: An insight towards multi-temporal landslide mapping using the U-Net deep learning model

Abstract

Repeated temporal mapping of landslides is essential for investigating changes in landslide movements, legacy effects of the landslide triggering events, and susceptibility changes in the area. However, in order to perform such investigations, multi-temporal (MT) inventories of landslides are required. The traditional approach of visual interpretation from cloud-free optical remote sensing imageries is time consuming and expensive. Recent endeavours exploring Convolutional Neural Networks and deep learning models have made rapid and accurate mapping of landslides feasible but have not been applied for multi-temporal landslide mapping in the Himalayas, yet. Earlier models used a standard supervised learning approach, with a small landslide inventory over a limited area used for training , which is then utilized to predict landslides in nearby areas. We propose a new strategy, using geographically separate training samples to design a standard approach which can be utilized to create multi-temporal landslide inventories. RapidEye images of 5-metres spatial resolution are used to generate MT landslide inventories in the study area of Rasuwa district, Nepal. We test the effectiveness of the model by training with only 55 landslides and predicting for a different area. Then, using the weights attained from this first training phase, we use transfer learning to map landslides over a time period between 2013 and 2019 in the Rasuwa district. We also adopt data augmentation techniques to add more training samples, leading to higher overall accuracies ranging from 58% in 2015 to 80% in 2017. We also perform a spatial comparison between the manual (observed) and predicted inventories to evaluate the differences between landslide densities and overall landslide statistics of landslide area distribution. The benefit of a transfer learning-based model training is that it circumvents the need for generating annual inventories for training a deep learning. A single event based inventory is enough to generate landslide inventories over a number of years, at least until landslide preparatory conditions do not change significantly. This application can enable automated workflows to generate MT landslide inventories of particular areas as the basis for landslide evolution and movement change analysis.

HR-GLDD: A globally distributed dataset using generalized DL for rapid landslide mapping on HR satellite imagery

“…The proposed dataset is evaluated through several state-of-the-art Deep Learning segmentation models. In the past years, the U-Net (Abderrahim et al, 2020) has been used in several landslide detection applications which yield generally the most reliable results (Bhuyan et al, 2022;Meena et al, 2022c;. Therefore, we decide to use it as a benchmark model when training on the proposed dataset.…”

“…By overcoming these challenges, automated pipelines that address these issues can considerably reduce the requirement for human involvement and pave the way for the development of reliable real-time mapping and monitoring of natural hazards at the continental and global scales. Based on the quality of GLDD, reliability of automated pipelines and rapidly growing availability of HR satellite imagery, we can realistically envision mapping of landslide instances and contribute towards generating and updating landslide inventories at largescales, spatially and potentially, also temporally (Bhuyan et al, 2022).…”

Generating multi-temporal landslide inventories through a general deep transfer learning strategy using HR EO data

“…Figures 10 and S1, S2, S3 and S4a) show that for all the cases except Papua New Guinea, the surface area of the predicted landslide are quite similar to that of the reference ground truth. The differences stem from the fact that modelled landslides suffer from geometric fragmentation which is persistent throughout all the study areas 56 . Therefore, landslide areas are often under-represented in the modelled outputs.…”

“…According to the United States Geological Survey (USGS), the main shock generated strong ground shaking reaching up to a maximum Peak Ground Acceleration (PGA) of 0.87 g 61 . Various studies examined the spatial distribution of co-and/or post-seismic landslides 35,62,63 as well as their evolution over time via MT landslide inventories 36,56,64,65 . Here, we only focus on a subset of the area affected by the co-seismic landslide event (Fig.…”

“…Here, we only focus on a subset of the area affected by the co-seismic landslide event (Fig. 2a), investigated in another study 56 for MT inventory generation. However, the current work promotes updated products than the previous study as this research employs a much advanced and richer model.…”