Comparative Analysis of Machine Learning Models for High-Resolution Mapping of Soil Organic Carbon Stocks Using Remote Sensing Variables in Northern Ghana

Abstract

Accurate and comprehensive knowledge of spatial soil characteristics is crucial for environmental modelling, risk assessment, and decision-making. The utilization of Remote Sensing data for Digital Soil Mapping has proven to be a cost-effective and time-efficient alternative to traditional soil mapping methods. However, the potential of Remote Sensing data in enhancing understanding of local-scale soil information in West Africa remains largely untapped. This research aimed to explore the use of satellite data, and laboratory-analysed soil samples to map the distribution of organic carbon (SOC) in Northeastern Ghana. Three statistical prediction models, namely Random Forest, Xtreme Gradient Boosting, and Naïve Bayes were employed and compared. To ensure robustness, internal validation was performed using cross-validation techniques. Analysis of model performance statistics indicated that the RF and XG techniques exhibited slightly superior performance compared to the Naïve Bayes Algorithm, with RF yielding the highest accuracy in most cases. One limitation of Naïve Bayes was its inability to effectively capture non-linear relationships between dependent and independent variables, leading to less accurate predictions of soil properties in unsampled locations. Among the spectral predictors, precipitation data was found to be the most significant in Random Forest and Xtreme Gradient Boosting models, while Soil Organic Matter, Soil Bulk Density, Biomes, and NDVI emerged as prominent terrain/climatic variables in predicting soil properties. Furthermore, the results highlighted Precipitation, Soil Bulk Density, Soil Organic Matter, and Land Surface Temperature as significant predictors in the Naïve Bayes Algorithm. With the growing availability of freely accessible Remote Sensing data, the enhancement of soil information at local and regional scales in data-scarce regions like West Africa can be achieved with relatively minimal financial and human resources.