Overview
Landslides are among the deadliest natural disasters on the planet: roughly 4,600 deaths and over $4 billion in damage worldwide every year. India alone lost more than 4,000 people to landslides between 2004 and 2016, on top of roughly $400 million in economic damage annually — and a lot of it is predictable, we just don't warn people fast enough. The systems that exist today are too slow to help: NASA's LHASA, one of the standard models, is computationally heavy and refreshes only once a day, even though satellite and radar data update hourly, and none of these models take advantage of modern AI forecasting, which can fold in live weather and terrain data far more accurately at a fraction of the compute cost.
LandSight is a real-time landslide prediction model designed to be fast, scalable, and cheap enough to actually deploy, built around a four-stage pipeline. Feature engineering started by identifying the strongest predictors of landslides from over a decade of data — precipitation, slope, and elevation came out as the core features — pulling topographic data (slope, elevation, lithology) and precipitation records from the Global Landslide Catalog (GLC) and EarthEnv into a single combined dataset that hadn't existed in one place before, layered with secondary features like soil type and vegetation (NDVI) to capture how risk shifts across landscapes. Model selection tested random forests, XGBoost, logistic regression, and deep learning, scoring each on accuracy and compute cost since the model needed to be deployable in low-resource settings — XGBoost won on both counts. Because landslide data is scarce and heavily imbalanced, I generated synthetic samples, adding 1,400 non-landslide events so the model could generalize across varied terrain instead of memorizing a handful of examples. Finally, I integrated live predictions from Google's GraphCast weather model, which is what turns LandSight from a static risk map into an actual hourly early-warning system for high-risk regions.