Non-Parametric Machine Learning for Predicting Soil Carbon Storage: A Comparative Analysis of Ensemble Methods

Abstract

Soil carbon storage is important factor to our environmental sustainability but estimating it correctly is a difficult task because there are intricate relationships between soil attributes, climate variables, urbanisation land uses and geospatial variables. This study examines the possibility of non
parametric machine learning models for the predication of Total Organic Carbon (TOC) in soil. We trained and compared four Machine Learning Models- Random Forest (RF) Model, Multilayer Perceptron (MLP) Model, K-Nearest Neighbours (KNN) and Gradient Boosting Machines (GBM)– on a
dataset containing soil physicochemical characteristics, climate variables and urbanisation indicators. The performance of the four models was compared on the basis of Root Mean Square Error (RMSE). The finding indicates that GBM performed best, recording the minimum test RMSE, whereas KNN records the maximum error. The results demonstrate the power of ensemble approaches to soil carbon predication, offering a strong, data driven tool for environment monitoring and land management. Further research will investigate Deep Learning methods and other Geospatial data to further improve predication accuracy.