Forest Classification Using Random Forests with Multisource Remote Sensing and Ancillary Gis Data

This research evaluates the utility and performance of a machine learning decision tree classification technique – random forests – for forest classification using remote sensing and ancillary spatial data, across a large area of heterogeneous forest ecosystems in Victoria, Australia. Random forest classification models for forest extent, type and height were trained using 786 2 km x 2 km aerial photograph interpreted (API) land cover maps, distributed across the Victorian Forests and Parks Monitoring and Reporting Information System (FPMRIS) public land forest and parks systematic sample point grid network. API land cover data was adjusted to a summer 2008 baseline using wildfire and forest management GIS data. Model variables included atmospheric, terrain and BRDF corrected Landsat TM image bands 1-5 and 7; MODIS derived annual NDVI range; topography layers; and modelled climate surfaces. Open-source GIS software GRASS and statistical analysis package R, were used to build the random forest classification model. Cross-validation of the model was undertaken using an out-of-bag bootstrap sample. Reserved sample test data was used to assess model performance. Preliminary results are encouraging, with forest classification achieving an overall accuracy of ~73% (Kappa: 0.72). The most important variables for forest classification were elevation and near and mid-infrared reflectance. This research demonstrates the utility of random forests for classifying large areas of environmentally and structurally diverse forest ecosystems. * Corresponding author.