Mapping Forest Canopy Height Across Large Areas by Upscaling ALS Estimates with Freely Available Satellite Data

Operational assessment of forest structure is an on-going challenge for land managers, particularly over large, remote or inaccessible areas. Here, we present an easily adopted method for generating a continuous map of canopy height at a 30 m resolution, demonstrated over 2.9 million hectares of highly heterogeneous forest (canopy height 0-70 m) in Victoria, Australia. A two-stage approach was utilized where Airborne Laser Scanning (ALS) derived canopy height, captured over 18% of the study area, was used to train a regression tree ensemble method; random forest. Predictor variables, which have a global coverage and are freely available, included Landsat Thematic Mapper (Tasselled Cap transformed), Moderate Resolution Imaging Spectroradiometer Normalized Difference Vegetation Index time series, Shuttle Radar Topography Mission elevation data and other ancillary datasets. Reflectance variables were further processed to extract additional spatial and temporal contextual and textural variables. Modeled canopy height was validated following two approaches; (i) random sample cross validation, and (ii) with 108 inventory plots from outside the ALS capture extent. Both the cross validation and comparison with inventory data indicate canopy height can be estimated with a Root Mean Square Error (RMSE) of 31% (5.6 m) at the 95th percentile confidence interval. Subtraction of the systematic component of model error, estimated from training data error residuals, rescaled canopy height values to more accurately represent the response variable distribution tails e.g., tall and short forest. Two further experiments were carried out to test the applicability and scalability of the presented method. Results suggest that (a) no improvement in canopy height estimation is achieved when models were constructed and validated for smaller geographic areas, suggesting there is no upper limit to model scalability; and (b) training data can be captured over a small percentage of the study area (6%) if response and predictor variable variance is captured within the training cohort, however RMSE is higher than when compared to a stratified random sample.