A machine learning framework reconstructs continuous biomass growth trajectories for managed Douglas-fir forests across the Pacific Northwest. People often assume that satellites monitor forest growth continuously over decades, but orbital lasers only capture isolated snapshots of canopy height. The new system solves this gap by training a recurrent neural network on ecosystem simulations and adjusting the predictions with spaceborne laser measurements.
Growing trees absorb carbon dioxide from the atmosphere and store it inside wood and forest soils as stands mature. Orbiting instruments record tree height at single moments, similar to seeing separate still photos instead of a full movie of forest development. The recurrent model processes daily weather, soil composition, and stand ages to calculate how wood mass accumulates year after year. This setup enforces strict mass balance rules so calculated wood growth never drops below zero or violates physical laws.
Researchers calibrated the Ecosys ecosystem model against carbon flux towers at two sites before training the neural network on regional growth data. They fine-tuned the model using spaceborne lidar observations collected between 2019 and 2023 across more than 2,300 harvest plots. The pretrained network matched simulated cumulative biomass trajectories with an R-squared of 0.98 and eliminated growth overestimates on held-out test plots.
The researchers state that this framework converts sparse laser endpoints into continuous, site-specific growth curves for carbon accounting and timber harvest planning. The reconstructed trajectories preserve regional biomass gradients shaped by local precipitation, temperature, and elevation across the Pacific Northwest.