Constructing a Large-Scale Landslide Database Across Heterogeneous Environments Using Task-Specific Model Updates

Abstract

Preparation and mitigation efforts for widespread landslide hazards can be aided by a large-scale, well-labeled landslide inventory with high location accuracy. Recent smallscale studies for pixel-wise labeling of potential landslide areas in remotely-sensed images using deep learning (DL) showed potential but were based on data from very small, homogeneous regions with unproven model transferability. In this paper we consider a more realistic and practical setting for large-scale heterogeneous landslide data collection and DL-based labeling. In this setting, remotely sensed images are collected sequentially in temporal batches, where each batch focuses on images from a particular ecoregion, but different batches can focus on different ecoregions with distinct landscape characteristics. For such a scenario, we study the following questions: (1) How well do DL models trained in homogeneous regions perform when they are transferred to different ecoregions, (2) Does increasing the spatial coverage in the data improve model performance in a given ecoregion (even when the extra data do not come from the ecoregion), and (3) Can a landslide pixel labeling model be incrementally updated with new data, but without access to the old data and without losing performance on the old data (so that researchers can share models obtained from proprietary datasets)? We address these questions by extending the Learning without Forgetting framework, which is used for incremental training of image classification models, to the setting of incremental training of semantic segmentation models (e.g., identifying all landslide pixels in an image). We call the resulting extension Task-Specific Model Updates (TSMU). TSMU semantic segmentation framework consists of an encoder shared by all ecoregions to capture the similarities between them, and ecoregion-specific decoders to capture the nuances of each ecoregion. This framework is continually updated using a threestage training procedure for each new addition of an ecoregion without having to revisit data from old ecoregions and without losing performance on them.A national compilation of landslide inventories by the U.S. Geological Survey (USGS) was used to develop database for this study. We focused on space-visible landslides within four ecoregions. These landslides were manually identified and labeled using high-resolution satellite images from Google Earth. The database contains 496 labeled and georeferenced pre-event/postevent images pairs, which corresponding to 1,918 landslide records in the USGS landslide inventory. Using the TSMU framework, we conduct extensive experiments on four ecoregions in the United States to address the aforementioned questions.

ISSLIDE: A New InSAR Dataset for Slow SLIding Area DEtection With Machine Learning

“…The Landslide4Sense [19] competition provides 3,799 annotated optical patches for post-event landslide detection. The dataset from [9] contains 1,918 landslides from the United States Geological Survey to implement their detection on optical images. Authors of SEN12-FLOOD [4] gathered Sentinel-1 and 2 images as well as labels provided by Copernicus Emergency Management Service to create a flood dedicated multimodal dataset with 412 time series.…”

“…If getting access to remote sensing images is straightforward with the Copernicus program, annotating this amount of data is not only costly and time consuming but requires also strong expert knowledge. This is the reason why land-cover classification and change detection datasets are often based respectively on pre-existing maps [6]- [8] or inventories [4], [5], [9] for labelling. Early warning systems are more challenging and, in particular, detecting ground deformations or motions requires both geomorphological expertise and radar interferometry knowledge.…”

Enhance the Accuracy of Landslide Detection in UAV Images Using an Improved Mask R-CNN Model: A Case Study of Sanming, China

“…In this study, we only selected the visible band of the UAV as the input data, and the experimental results show that the interference from the background noise of the surrounding features cannot be completely eliminated. Referring to other landslide detection studies [56], fusing more data from other regions can improve the accuracy of the model detection. In subsequent studies, we can train and validate by accessing additional landslide datasets to improve the model and increase its generalization to new test areas.…”

Mapping Landslide Susceptibility Over Large Regions With Limited Data

“…The landslide inventories used to train the models should outline areas within the modeling domain with geomorphic evidence of landsliding. Despite the proliferation of new automated landslide mapping techniques that use increasingly available remote sensing data (e.g., Benz & Blum, 2019; Ghorbanzadeh et al., 2021; Nagendra et al., 2022), landslide inventories required for LSSMs are still lacking over most of the world.…”