Hierarchical Bayesian geostatistics for C stock prediction in disturbed plantation forest in Zimbabwe

Abstract

We develop and present a novel Bayesian hierarchical geostatistical model for the prediction of plantation forest carbon stock (C stock) in the eastern highlands of Zimbabwe using multispectral Landsat-8 and Sentinel-2 remotely sensed data. Specifically, we adopt a Bayesian hierarchical methodology encompassing a model-based inferential framework making use of efficient Markov Chain Monte Carlo (MCMC) techniques for assessing model input parameters. Our proposed hierarchical modelling framework evaluates the influence of two but related covariate information sources in C stock prediction in order to build sustainable capacity on carbon reporting and monitoring. The perceived improvements in the spectral and spatial properties of Landsat-8 and Sentinel-2 data and their potential to predict C stock with shorter uncertainty bounds is tested in the developed hierarchical Bayesian models. We utilized the Mean Squared Shortest Distance (MSSD) as the objective function for optimization of sampling locations for equal area coverage. Specifically, we evaluated the models using four selected remotely sensed vegetation indices namely, the normalised difference vegetation index (NDVI), soil adjusted vegetation index (SAVI), enhanced vegetation index (EVI) and an additional distance to settlements anthropogenic variable that justifies from the history of the studied plantation forest in the eastern highlands of Zimbabwe. We evaluated two models making use of Landsat-8 and Sentinel-2 derived predictors using the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Coverage (CVG) and Deviance Information Criteria (DIC). The Sentinel-2 based C stock model resulted in RMSE of 1.16 MgCha−1, MAE of 1.11 MgCha−1, CVG of 94.7% and a DIC of −554.7 whilst its Landsat-8 based C stock counterpart yielded a RMSE, MAE, CVG and DIC of 2.69 MgCha−1, 1.77 MgCha−1, 85.4% and 43.1 respectively. Although predictive models from both sensors show great improvement in predictive accuracy when modelling the spatial random effects, the Sentinel-2 based C stock predictive model substantially outperforms its Landsat-8 based C stock counterpart. The Sentinel-2 based C stock predictive hierarchical model therefore adequately addresses multiple sources of uncertainty inherent in the spatial prediction of C stock in disturbed plantation ecosystems. It is evident from the results of this study that carbon reporting and monitoring can always be improved by scouting for improved and easily accessible remote sensing data and allow forest practitioners to keep track of error across space in resource environments of interest.