An evaluation of hyperspectral and multispectral data for mapping invasive species in an African Savanna.

Abstract

Invasive alien plant (IAP) species affects a range of ecosystem types in various regions of the world. Therefore are now considered one of the main phenomena causing global change. Invasive alien plants (IAP’s) cause considerable impacts on ecosystem processes and functions, biodiversity, agriculture and human well-being. Parthenium hysterophorus is an IAP which is widely spread across the globe. It is difficult to control and eradicate, and has detrimental impacts on the natural environment and human health. However, there is no record of accurate and up-to-date information on the distributions and extent of P. hysterophorus. This study evaluated the capability of hyperspectral and multispectral data for mapping P. hysterophorus in northern KwaZulu-Natal province, South Africa. First, the study sought to determine an optimal subset of bands from canopy hyperspectral data for discrimination of P. hysterophorus from its co-existing species. A novel hierarchical approach that integrates statistical filters and a wrapper technique has been proposed to select optimal bands to solve the problem of high spectral dimensionality and improve classification accuracy. A non-parametric algorithm, Support Vector Machines (SVM) showed inferior classification accuracy, i.e. 76.19% and 78.57% when using 20 best spectral bands from SVM – Recursive Feature Elimination (SVM-RFE) and entire dataset (n = 1633), respectively. On the other hand, superior overall accuracy of 83.33% was achieved when using ten spectral bands identified by the hierarchical approach. Next, SVM classifier was adopted to evaluate the capability of multispectral data (i.e. Operational Land Imager, OLI and SPOT 6) for determining the distribution and patch sizes of P. hysterophorus. The results showed that SPOT 6 had a higher overall accuracy of 83.33% than OLI, i.e.76.39%. While SPOT 6’s the higher spatial resolution was useful for better characterisation of the distribution and patch sizes, the study found that the spectral configuration of OLI was more important in identifying possible locations infested by P. hysterophorus. Overall, the study demonstrated that fewer spectral bands selected by the proposed hierarchical approach have the greatest potential for reliably discriminating IAP species using airborne and satellite hyperspectral sensors. The study also demonstrated that the current information needs on IAP’s can be addressed using accessible multispectral data, valuable for effective land management, site specific weed management, and site prioritisation.