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

Preparation and mitigation efforts for widespread landslide hazards can be aided by a large-scale, well-labeled inventory with high location accuracy. Recent smallscale studies pixel-wise labeling of potential areas in remotely-sensed images using deep learning (DL) showed but were based on data from very small, homogeneous regions unproven model transferability. In this paper we consider more realistic practical setting large-scale heterogeneous collection DL-based labeling. setting, remotely sensed are collected sequentially temporal batches, where each batch focuses particular ecoregion, different batches focus ecoregions distinct landscape characteristics. For such scenario, study the following questions: (1) How well do DL models trained perform when they transferred to ecoregions, (2) Does increasing spatial coverage improve performance given ecoregion (even extra not come ecoregion), (3) Can pixel incrementally updated new data, without access old losing (so that researchers share obtained proprietary datasets)' We address these questions extending Learning Forgetting framework, which is used incremental training image classification models, semantic segmentation (e.g., identifying all pixels an image). call resulting extension Task-Specific Model Updates (TSMU). TSMU framework consists encoder shared capture similarities between them, ecoregion-specific decoders nuances ecoregion. This continually threestage procedure addition having revisit them.