Exploring Innovative Machine Learning Models for Soil $\text{CO}_{2}$ Emission Forecasting

Soil respiration, the natural process by which carbon dioxide (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{CO}_{2}$</tex>) is released from the soil into the atmosphere, is the largest flux of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{CO}_{2}$</tex> from the Earth's surface. Despite its significance, accurately predicting <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{CO}_{2}$</tex> emissions caused by soil respiration remains challenging, and the influence of environmental variables are not yet fully understood. Early studies using artificial intelligence (AI) to predict soil respirations often relied on linear models or overly simplified datasets which lacked real-world applications. This study introduces a novel integration of XGBoost and Shapley Additive exPlanations (SHAP) to achieve both high predictive accuracy and detailed feature interpretability. In this study, an innovative ensemble model, XGBoost, is developed to estimate annual soil respiration (Rs) using spatial, temporal, climatic, and other independent variables. Additionally, a deep neural network (DNN) model is explored. Among the machine learning models evaluated, XGBoost shows the best performance, with a 4% reduction in Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) and a 3.7 % increase in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex> compared with the Random Forest model, which was so far believed to be the state-of-the-art method. To aid in model explainability, SHAP is used to give feature importance at both the global and individual levels of analysis. Results indicate that Mean Annual Temperature (MAT) and Mean Annual Precipitation (MAP) are the most dominant environmental factors, with geographic variables such as latitude and longitude also significant. Other strong predictors include Study Year, Soil Type, soil composition-related features, Elevation, and Basal Area (BA). This study fills the gap between predictive modeling and interpretability, providing an explainable AI framework that is scalable for realworld environmental research.