Effect of Image Band Number and Band Multicollinearity on Mapping Accuracy of Tree Diversity and Evenness

Abstract

Satellite imageries are increasingly being used in the studies of plant diversity due to cost effective wider spatial coverage. Multispectral or hyperspectral imageries consist of different spectral bands where some bands might be multicollineared among them and deemed redundant to use in the classification/analysis. This study investigates how the accuracies of maps of tree diversity and evenness of an area are affected when redundant and multicollineared bands are not included in image classification. We classified two medium resolution imageries; Landsat Thematic Mapper (TM) and Advanced Land Observation Satellite (ALOS) Advanced Visible and Near Infrared Radiometer type 2 (AVNIR-2), in combination with four ancillary environmental variable layers namely elevation, slope, aspect and solar radiation in backpropagation Neural Network (NN).  Map accuracies obtained by using all bands/all multicollineared bands of TM and environmental variables were compared and contrasted with the corresponding accuracies obtained by using all bands/all non-multicollineared bands of ALOS with environmental variables. Results show that use of all available bands of TM/ALOS does not increase the accuracies of the tree diversity and evenness maps significantly than using its necessary fewer bands. Also the use of multicollineared bands do not increase map accuracies significantly as long as the principal bands of the imageries (band no 4 of TM and ALOS, 0.76–0.90 μm) responsible for capturing vegetation diversity and evenness are not excluded from classification. In a nutshell, in remote sensing tree diversity and evenness using TM and ALOS data, number of image bands and multicollinearity among them are not factors to affect map accuracy significantly.