A Bayesian geo-statistical approach for plantation forest productivity assessment after the fast-track land reform in Zimbabwe.

Abstract

The principal objective of the current study was to investigate how the new generation multispectral remote sensing, along with variants of the Bayesian hierarchical geostatistical methodology, could handle prediction uncertainty of carbon (C) stock. The assessment of C stock prediction uncertainty was conducted in a managed and disturbed plantation forest ecosystem located in Manicaland province of Zimbabwe. To achieve this, the study made use of ancillary data from the multispectral (Landsat-8 and Sentinel-2) remote sensing platforms, which informed the application of different inferential and methodological variants within the Bayesian hierarchical geostatistical framework. Allometric equations suited for the target plantation tree species in the sampled region were used to derive C stock from Above ground Biomass (AGB) sampled on 500 m2 circular supports. These Bayesian geostatistical models utilized a combination of Landsat-8 and Sentinel-2 derived vegetation indices, along with climatic and topographic variables. The study found that the Normalized Difference Vegetation Index (𝑁𝐷𝑉𝐼), Distance to settlements (𝐷𝐼𝑆𝑇), and Soil Adjusted Vegetation Index (𝑆𝐴𝑉𝐼) played crucial roles in influencing the spatial distribution of C stock in the studied region. Enhanced Vegetation Index (𝐸𝑉𝐼) is an insignificant predictor for both Landsat-8 and Sentinel-2 driven C stock predictions. Among the tested Bayesian approaches, the spatially varying coefficient (SVC) model, the multi-source data-driven Bayesian geostatistical approach, and the frequentist geostatistical framework were examined. Regardless of the various specifications for independent variables in the predictive C stock modelling within the Bayesian framework, 𝑁𝐷𝑉𝐼 and 𝐷𝐼𝑆𝑇 emerged as significant predictors of the modelled response variable. The non-stationary and Sentinel-2 driven Bayesian hierarchical model, with 𝑁𝐷𝑉𝐼 and DIST covariables, proved to be the most effective prediction model in the studied plantation forest ecosystem in Zimbabwe. This best-performing C stock predictive model was subsequently used to predict C stock under both current (1970-2000) and future (SSP5-8.5) 2075 climate scenarios. The results of the Bayesian constructed hierarchical model indicate a significant shrinkage of forest C stock density and distribution under the future SSP5-8 (2075) business-as-usual climate projection. Basically, the findings of this study highlight the critical role of new generation multispectral remote sensing and Bayesian geostatistical approaches in assessing and predicting carbon stock uncertainty in forest ecosystems. These insights have significant implications for informed land management strategies, aligning with the goals and recommendations of the Intergovernmental Panel on Climate Change (IPCC) to effectively address climate challenges and enhance sustainable land management practices.