StruMPL: Multi-task Dense Regression under Disjoint Partial Supervision and MNAR Labels
Researchers have developed StruMPL, a novel multi-task dense regression model designed to estimate forest aboveground biomass (AGB) using disparate data sources. The model integrates satellite lidar data, which provides structural information but lacks biomass estimates, with ground-based plot data that offers biomass figures but is subject to bias and missingness. StruMPL addresses these challenges by employing a shared encoder with regression, imputation, and propensity heads for spatial correction, alongside a learnable physics module to enforce known allometric laws between variables. AI
IMPACT Introduces a new method for integrating heterogeneous data sources in regression tasks, potentially improving ecological modeling accuracy.