Estimating Basal Area of Spruce and Fir in Post-Fire Residual Stands in Central Siberia Using Quickbird, Feature Selection, and Random Forests

The occurrence and spatial arrangement of post-fire alive residual stands affect the recolonization of trees and animals of burned areas in boreal ecosystems. Because the analysis of residual stands in the field is prohibitively expensive we lack understanding on how residual stands are distributed and why. Here, we explore the use of high-resolution Quickbird satellite imagery in conjunction with in-situ measurements and machine learning techniques to map basal area of spruce and fir for two fire areas in Central Siberia, and analyze the distribution of residual stands with respect to topography. First, an advanced feature selection algorithm which combines a genetic algorithm with guided local search is wrapped around the Random Forests regression technique, to identify suitable variable subsets out of a large number of candidate variables that were derived from Quickbird data. Second, we train and apply Random Forests using the derived variable subsets to the two fire areas to generate spatially explicit estimates of basal area for spruce and fir. Third, we analyze species specific differences and the relationship between basal area and topography using a high resolution digital elevation model from ASTER data. Our results show that the main gradients of species specific basal area can be reproduced using Quickbird data but stress the importance of variable selection. We find associations of residual stands with topography depressions and channels exhibit larger prevalence of residual stands than ridges or plateaus, the latter being more often subject to severe fires. We further found that the relationship between basal area and elevation tends to be reversed inside the burned area in comparison to the surrounding unburned forest. Our results suggest that local topography may control the sensitivity of ecological processes to a changing fire regime with more severe fires, and highlight the synergistic use of high resolution satellite remote sensing and machine learning methods for fire ecological applications.