Deep convolutional neural network-based pixel-wise landslide inventory mapping

This paper reports a feasible alternative to compile a landslide inventory map (LIM) from remote sensing datasets using the application of an artificial intelligence-driven methodology. A deep convolutional neural network model, called LanDCNN, was developed to generate segmentation maps of landslides, and its performance was compared with the benchmark model, named U-Net, and other conventional object-based methods. The landslides that occurred in Lantau Island, Hong Kong, were taken as the case study, in which the pre- and post-landslide aerial images, and a rasterized digital terrain model (DTM) were used. The assessment reveals that LanDCNN trained with bitemporal images and DTM yields the smoothest and most semantically meaningfully LIM, compared to other methods. This LIM is the most balanced segmentation results, represented by the highest F-1 measure among all analyzed cases. With the encoding capability of LanDCNN, the application of DTM as the input renders better LIM production, especially when the landslide signatures are relatively subtle. With the computational setup used in this study, LanDCNN requires similar to 3 min to map landslides from the datasets of approximately 25 km(2) in area and with a resolution of 0.5 m. In short, the proposed landslide mapping framework, featured LanDCNN, is scalable to handle the vast amount of remote sensing data from different types of measurements within a short processing period.

Automated summary

This study presents a feasible alternative method to compile a landslide inventory map using artificial intelligence. The developed LanDCNN model, trained with bitemporal images and DTM, produces the smoothest and most semantically meaningful results among all analyzed cases. The proposed framework is scalable and efficient in handling vast amounts of remote sensing data within a short processing period.