Mapping natural forest cover, tree species diversity and carbon stocks of a subtropical Afromontane forest using remote sensing.

Abstract

Natural forests cover about a third of terrestrial landmass and provides benefits such as carbon sequestration, and regulation of biogeochemical cycles. It is essential that adequate information is available to support forest management. Remote Sensing imageries provide data for mapping natural forests. Hence, our study aimed at mapping the Nkandla Forest Reserve attributes with Remote Sensing imageries. Quantitative information on the forest attributes is non-existent for many of these forests, including the sub-tropical Afromontane Nkandla Forest Reserve. This does not support scientific and evidence based natural forest management. A review of literature revealed that progress has been made in Remote Sensing monitoring of natural forest attributes. The Random Forest (RF) and Support Vector Machine (SVM) were applied to Landsat 8 in classifying the land use land cover (LULC) classes of the forest. Each of the algorithms produced higher accuracy of above 95% with the SVM performing slightly better than the RF. The SVM, Markov Chain and Multi-Layer Perceptron Neural Network (MLPNN) were adopted for a spatiotemporal change detection over the last 30 years at decadal interval for the forest. There were consistent changes in each of the four LULC classes. The study further conducted a forecasting of LULC distribution for 2029. Aboveground carbon (AGC) estimation was carried out using Sentinel 2 imagery and RF modelling. Four models made up smade of Sentinel 2 products could successfully map the AGC with high accuracies. The last two studies focused on tree species diversity with the first evaluating the influence of spatial and spectral resolution on prediction accuracies by comparing the PlanetScope, RapidEye, Sentinel 2 and Landsat 8. Both the spatial and spectral resolution were found to influence accuracies with the Sentinel 2 emerging as the best imagery. The second aspect focused on identifying the best season for the prediction of tree species diversity. Summer imagery emerged as the best season and the winter being the least performer. Overall, our study indicates that Remote Sensing imageries could be used for successful mapping of natural forest attributes. The outputs of our studies could also be of interest to forest managers and Remote Sensing experts.