Growth Before Differencing: Direct Transition Learning for Forest Biomass Monitoring

Mature forests can accumulate above-ground biomass after canopy closure, but annual growth is small relative to standing stock. Differencing two satellite stock estimates can therefore suppress the signal that carbon-monitoring systems seek. We test direct transition supervision using repeated individual-tree allometry from 238 mature-forest intervals at 20 National Ecological Observatory Network sites. A growth-jump transition model predicts surviving-tree growth from AlphaEarth embedding displacement, optical, radar, lidar, and climate predictors; evaluation holds out entire sites. Predicted-stock differencing yields 0.54 megagrams of above-ground biomass per hectare per year against 3.33 observed survivor growth. Direct supervision predicts 3.34 and reduces plot mean absolute error from 3.67 to 1.50, with a paired site-bootstrap reduction of 2.17 and a 95 percent bootstrap interval from 1.58 to 2.73. Site-mean transfer is clearer, with a correlation of 0.79 and an R-squared value of 0.44, although plot R-squared is 0.11. AlphaEarth alone is weak, and adding it to multisensor predictors produces mixed, modest gains. Mortality-driven net change remains unresolved. The practical contribution is a target-design principle for forest-carbon monitoring, not a net-flux or crediting product.