Commercial forest species discrimination and mapping using image texture computed from WorldView-2 pan sharpened imagery in KwaZulu-Natal, South Africa.

Abstract

Forest species discrimination is vital for precise and dependable information, essential for commercial forest management and monitoring. Recently, the adoption of remote sensing approaches has become an important source of information in commercial forest management. However, previous studies have utilized spectral data or vegetation indices to detect and map commercial forest species, with less focus on the spatial elements. Therefore, this study using image texture aims to discriminate commercial forest plantations (i.e. A. mearnsii, E. dunnii, E. grandis and P. patula) computed from a 0.5m WorldView-2 pan-sharpened image in KwaZuluNatal, South Africa. The first objective of the study was to discriminate commercial forest species using image texture computed from a 0.5m WorldView-2 pan-sharpened image and the Partial Least Squares Discriminate Analysis (PLS-DA) algorithm. The results indicated that the image texture model (overall accuracy (OA) = 77%, kappa = 0.69) outperformed both the vegetation indices model (OA = 69%, kappa = 0.59) and raw spectral bands model (OA = 64%, kappa = 0.52). The most successful texture parameters selected by PLS-DA were mean, correlation, and homogeneity, which were primarily computed from the red-edge, NIR1 and NIR2 bands. Lastly, the 7x7 moving window was commonly selected by the PLS-DA model when compared to the 3x3 and 5x5 moving windows. The second objective of the study was to explore the utility of texture combinations computed from a fused 0.5m WorldView-2 image in discriminating commercial forest species in conjunction with the PLS-DA and Sparse Partial Least Squares Discriminate Analysis (SPLS-DA) algorithm. The accuracies achieved using SPLS-DA model, which performed variable selection and dimension reduction simultaneously yielded an overall accuracy of 86%. In contrast, the PLS-DA and variable importance in the projection (VIP) produced an overall classification accuracy of 81%. Generally, the finding of this study demonstrated the ability of image texture to precisely provide adequate information that is essential for tree species mapping and monitoring.