Forest structure monitoring with small footprint LIDAR-optimized spectral remote sensing

Forest structural content of hypserspectral imagery was evaluated over oak-hickory forests within the Ozark National Forest in north-central Arkansas, USA, and evaluated for prediction of basal area. A LIDAR-assisted assessment of the structural information contained in the hyperspectral imagery was used in a machine learning process to define the spectral derivatives that would best predict biophysical variables. NASA Hyperion hyperspectral satellite derivatives were used to develop rule sets for predicting normalized height percentile (NHP) surfaces from a Leica Geosystems ALS50 small footprint LIDAR point cloud. The most successful predictors of forest structure were subsequently tested in rule sets to predict basal area measured in situ. Selected hyperspectral indices and bands from the minimum noise fraction transformation (MNF) were converted to rule sets using the Cubist machine learning decision tree. The machine learning phase was able to predict the LIDAR normalized height percentiles with accuracies between 2.08--3.69 meters based on the root mean squared error. The results indicate the hyperspectral data contains valuable information that can predict the canopy and in particular understory characteristics of the forest. The prediction of the lowest NHP layer (representing understory) consistently resulted in the highest accuracy of 2.08 meters. The results suggest, at the 30 x 30 m measurement scale, that orbital hyperspectral imagery can be used as a first step in the monitoring of forest structural variables of interest. Continued development of rapidly calibrated biophysical remote sensing techniques will allow timely and accurate assessment of forest conditions across large geographic regions.