Debris-Cast: A Two-Stage Machine Learning Framework for Hurricane Debris Prediction Using Multi-Source Geospatial Data

Accurate hurricane debris estimation is important for contractor mobilization, temporary debris-management site planning, and resource allocation. Existing tools often rely on fixed empirical formulations and limited inputs, which can restrict their ability to represent nonlinear relationships among hazard intensity, the built environment, vegetation, terrain, and flood exposure. This study evaluates a two-stage machine-learning framework for hurricane debris estimation using multi-source geospatial data, observed and post-event hazard inputs, and operational debris-removal records. The framework first classifies whether debris removal was recorded within a spatial unit and then estimates debris volume for debris-present cells. Random Forest and XGBoost models were evaluated for construction and demolition, vegetative, and combined debris from Hurricanes Helene and Milton in Pinellas County, Florida, using H3 hexagonal and equivalent-area square grids at two spatial resolutions. Classification showed higher observed performance at finer resolutions, with the highest AUC of 93.9% for construction and demolition debris using Random Forest at H3 resolution 9. Regression results revealed a scale-dependent tradeoff. Coarser grids generally produced stronger model fit and lower area-normalized error, with R 2 reaching 0.869 for vegetative debris, whereas finer grids provided greater spatial detail and lower raw per-cell MAE. Hexagonal and square grids showed comparable observed performance. The study demonstrates retrospective and screening capability using historical data. Pre-landfall forecasting, transferability across regions and events, and operational deployment require further validation using independent datasets and true forecast inputs.