In-situ hyperspectral remote sensing feature extraction of selected common tropical rainforest species

Abstract

Hyperspectral remote sensing has potentials in solving dilemma due to high diversity of tropical tree species during tree spatial distribution mapping for forest management and conservation. This research aims to establish a multi-level tree species classification strategy which has capability in dealing with high diversity of tropical tree species in Malaysia. Three research objectives were formed namely, 1) to evaluate the influence of spatial scale in within species spectral variability of tropical tree, 2) to examine the effectiveness of multi-level classification strategy in improving tree species classification accuracy, and 3) to study the influence of spatial scale and species grouping methods in multi-level tree species classification. A total of 20 tropical tree species and in-situ hyperspectral remote sensing data were collected at tree branch and leaves spatial scales. Spectral variation analysis has revealed a significant influence of remote sensing data spatial scale on within species spectral variability where tree branch spatial scale data dominated the upper range of this variability in the majority of the tree species in this research. Support Vector Machine (SVM) and Maximum Likelihood Classifier (MLC) methods were adopted in the multi-level classification strategy to classify tree species using 32 vegetation indices extracted from in-situ hyperspectral data. The multi-level classification strategy has resulted in a 5% improvement in the classification accuracy from the ordinary classification for both SVM and MLC classifiers. The improvement was marked from 69.41% to 74.56% and from 64.98% to 69.53% in SVM and MLC tree species classifications respectively. Four tree species classification scenarios were designed in combinations of two spatial scales data (i.e. leaves spatial scale and tree branch with leaves spatial scale) with two species grouping modes to study the influence of these variables on the performance of multi-level SVM classification. Tree species data at tree branch spatial scale has proven its influence on the classification accuracy where SVM produced the accuracy at 77.21% and 72.79% for leaves spatial scale and tree branch with leaves spatial scale respectively at the first level in the multi-level classification strategy. Later, the multi-level SVM classification strategy has made a 2% improvement in the classification accuracy for tree species classification scenarios in the next two levels of classification. Two designed tree species groupings namely mode A (grouping based on individual classification accuracy) and mode B (grouping based on individual misclassification error) have presented influence on the multi-level SVM classification performance. The influence was shown in the number of sub-groups and tree species in sub-groups formed by the two grouping modes. Out of the four tree species classification scenarios, the multi-level SVM classification strategy has the best performance in the case of leaves spatial scale with species grouping mode A with a classification accuracy recorded at 79.2%. This research has proven multi-level classification strategy has its capability in handling a high number of tropical tree species with promising accuracy in tree species spatial distribution mapping.