Adaptive Stratified Screening and Weakly Supervised UNet with Sparse ALS Refinement for Forest Canopy Height Mapping

Forest canopy height is a key parameter for estimating forest biomass and carbon stocks, as well as a fundamental indicator for assessing forest ecosystem health. Although existing 10 m resolution global canopy height products are free and open-source, they suffer from substantial systematic errors due to the inherent biases of spaceborne LiDAR and the nonlinear relationship between optical imagery and canopy height, making them insufficiently accurate as direct training labels for deep learning. This study proposes a complete framework of “adaptive stratified screening, followed by weakly supervised UNet training and minimal ALS post-processing refinement.” First, using the 12 spectral bands of Sentinel-2 multispectral imagery, the OOB distance is calculated via weighted random forest as an anomaly metric. Through systematic comparison of three anomaly metrics (Cook’s distance, OOB distance, and standardized residuals) and five retention ratios (30–50%), the OOB distance at 40% retention is identified as optimal. An adaptive retention strategy based on tree height stratification is then implemented: for low-tree areas (<20 m), 40% of low-anomaly samples are uniformly retained; for tall-tree areas (≥20 m), the retention ratio is dynamically adjusted according to the proportion of tall-tree pixels (95% when tall-tree proportion < 10%, 70% when 10–30%, and 50% when ≥30%). The screened high-confidence Lang data are then used as supervisory signals to train a UNet model for predicting canopy height across the entire region, producing an initial canopy height map (MAE = 6.97 m, Bias = +4.21 m). Finally, using only a minimal amount of airborne LiDAR data (down to 0.1% area proportion), with a 14-dimensional input comprising the initial predicted height, screening mask values, and the 12 spectral bands of Sentinel-2, a multilayer perceptron is independently trained for each region to perform post-processing refinement of the initial predictions. Experiments based on 68 independent regions from the Open-Canopy dataset, with a strict regional split (34 validation/34 test), show that post-processing refinement reduces the test set MAE from 6.97 m (ASS) to 3.29 m (0.1% ALS, a 52.8% improvement, 58.1% improvement over the original Lang product), Bias from +4.21 m to −0.13 m, and RMSE from 8.95 m to 4.72 m. Even when the ALS proportion is reduced to 0.01%, the MAE remains 3.57 m (a 48.8% improvement). Statistical significance testing (paired Wilcoxon signed-rank test, n = 34) indicates that ASS+MLP(0.1%) significantly outperforms ASS (p < 0.00001). This study provides an effective technical pathway for regional refinement of global canopy height products.