An Integrated Deep Learning Approach to Estimate Canopy Height and Uncertainty by Combining Seasonal Optical, SAR and Limited GEDI LiDAR Data over Ontario’s Managed Forests, Canada

Accurate canopy-height mapping is essential for forest carbon monitoring, yet spatially continuous estimates over northern forests remain limited by sparse LiDAR sampling and unreliable optical data. We present a probabilistic deep learning framework that integrates seasonal Landsat optical, Sentinel-1 C-band SAR, and yearly ALOS-PALSAR-2 L-band SAR composites with GEDI training targets to produce 30 m canopy-height and pixel-level uncertainty maps over Ontario, Canada. A ResUNet ensemble is trained with a Laplace negative log-likelihood (NLL) loss; aleatoric and epistemic uncertainty are captured through a mixture of Laplace distributions, calibrated post hoc by Platt scaling and validated against airborne LiDAR (ALS P98). Under matched architecture and inputs, replacing Gaussian with Laplace NLL improved R2 by 18.6%, reduced RMSE by 13.7%, and cut systematic bias from 1.43 to 0.14 m. The primary model reached R2=0.70, RMSE = 3.65 m, and bias = 0.23 m, exceeding GEDI’s own footprint-level agreement with the independent ALS reference within Ontario’s managed forest zone. Seasonal compositing raised R2 by 12.9% over summer-only inputs, and SAR-only configurations outperformed four global products in every forest class. The framework also delivers a calibrated, pixel-level uncertainty layer that reflects local structural difficulty beyond canopy height and can serve as a confidence layer for downstream carbon-stock and forest-management applications.