Observation-Aware Graph Transfer for Ladder Fuel Density Under Coverage Distribution Shift

Climate change is increasing the frequency and severity of wildfire and other disturbances, making forest structural conditions increasingly dynamic and creating a growing need for timely information to support climate-adaptive fuel management. Ladder fuel density (LFD), a key component of vertical forest structure, is associated with wildfire behavior and can inform fuel-reduction interventions. However, repeated high-resolution airborne laser scanning (ALS) surveys are costly, and existing gap-filling approaches rely on assumptions of local similarity, spatiotemporal continuity, or mechanistic regularity that disturbance can violate. We investigate when spatial and structural relationships remain informative once these assumptions break down, constructing graph-based representations of this relational information as an early-stage instantiation. Across four forward-consistent reference-target combinations in the Caples Creek watershed, local interpolation (RFSI) contributed negatively in every case while structural graph context remained positive throughout, which is a failure driven by training/deployment coverage mismatch, not sparse coverage itself. We plan to test whether this reflects a genuine methodological gap, decompose the graph's contribution from correlated covariates, extend the analysis across wider temporal gaps, and propose a heterogeneous multi-relational graph architecture with an observation-support-dependent gating mechanism to learn this weighting directly. If successful, this approach could shorten the lag between disturbance and updated fuel-condition information, helping prioritize fuel-reduction resources in time as conditions shift under climate change, with potential extension to other vegetation-risk domains such as utility corridor management.