Automated temporal landslide detection: Testing whether human-in-the-loop machine learning can bootstrap generalizability

Landslides cause more than 4,000 deaths and $20 billion in economic losses each year, with fatalities projected to increase 140–160% by mid-century as climate change intensifies extreme rainfall. Risk-informed adaptation – deciding where to build, where to harden infrastructure, where to relocate vulnerable communities – requires knowing not just where landslides might occur, but how often they actually do. This temporal frequency information does not currently exist anywhere in the world. Existing remote-sensing approaches require prior knowledge of when landslides happened, thus limiting their use to post-event mapping. To address this gap, we are testing whether a three-dimensional convolutional neural network applied to monthly satellite image stacks can discover both the location and timing of landslides without prior knowledge of event dates. Year 1 is a rigorous feasibility test across four or more diverse datasets to determine whether systematic data curation can bootstrap model generalizability. If successful, the approach will provide foundational frequency data to support operational hazard assessment and adaptation planning for the more than 300 million people living in landslide-prone areas worldwide.