Mapping carbon content in a mountainous grassland using SPOT 5 multispectral imagery and semi-automated machine learning ensemble methods

Abstract

Traditional methods of determining and monitoring CO2 fluxes are limited to a small area defined by the fetch of the instrumentation and depend on extensive and expensive fieldwork. Remote sensing, however, enables monitoring studies in a wide area at constant time periods and is likely to be more robust and adapted to a wider range of environmental conditions than the measurements taken from conventional instruments. In this study, we investigated the potential of using a multispectral sensor with moderate spectral capabilities (4 bands) to remotely detect and map the content of carbon over a mountainous rural rangeland in KwaZulu-Natal, South Africa. The results showed that the spatial and spectral resolution of the SPOT 5 satellite can effectively map carbon across a rangeland using two contemporary ensemble models. More specifically, the study demonstrated that random forest (RF) produced the best correlation coefficient of 0.71 with an RMSE of 9 when compared to the correlation coefficient of 0.68 and RMSE of 20 when using the stochastic gradient boosting (SGB). Overall, this study provided a framework for utilising efficient ensemble methods to automatically predict grass carbon over a mountainous region while exploiting the SPOT 5 spectral bands. The study found that the SWIR, Green, and NIR spectral regions were the top three most effective bands.