Supervised tree-ensemble classification

Supervised tree-ensemble classification answers the question: what land-cover class does each pixel belong to, based on patterns learned from labelled examples? The method trains a decision-tree ensemble on spectral and temporal features derived from satellite imagery, then applies the trained ensemble to produce wall-to-wall class probability maps.

The Hansen et al. 2013 Global Forest Change (GFC) product is the anchor implementation.^[1] The algorithm computes per-pixel temporal spectral metrics from annual Landsat growing-season composites: statistical ranks, means, and linear regression coefficients of red, NIR, SWIR, and thermal bands, plus derived indices including NDVI, NBR, and NDWI. These metrics serve as features for a bagged decision-tree ensemble trained on geographically distributed labelled reference samples. Each tree in the ensemble votes on the class membership of a pixel (tree cover, forest loss, forest gain); the median vote probability across the ensemble is thresholded to assign the final class label. Annual updates apply the same pipeline to each successive year's Landsat stack, providing a record of global forest cover change from 2000 to present at 30m resolution.^[2]

The GLAD-L alert system (Hansen et al. 2016) extends the same bagged decision-tree approach to near-real-time operation, applying the ensemble to the incoming Landsat stream with a 4-observation confirmation rule.^[3] The GFC ensemble uses bagged decision trees, which aggregate bootstrap samples of the training data across a tree ensemble.^[1]

The method degrades in persistently cloudy tropical regions, complex mixed-forest landscapes with ambiguous spectral signatures, and areas with rapid phenological variation that confounds the temporal metrics.