An Explainable Predictive Approach for Investigation of Greenhouse Gas Emissions in Maritime Canada's Potato Agriculture

An Explainable Predictive Approach for Investigation of Greenhouse Gas Emissions in Maritime Canada's Potato Agriculture

This study aims to develop an optimized expert system that effectively models and predicts greenhouse gas (GHG) emissions from potato crops, integrating experimental data and advanced computational methods. This research seeks to significantly contribute to mitigating climate change impacts and improving food security. We address the challenges of precise GHG data collection in Maritime Canada's potato cropping system by utilizing a high-precision LI-COR instrument, ensuring accurate and reliable measurements for this study. In this effort, the first stage of the investigation comprised measuring experimental soil properties and greenhouse gas (GHG) emissions, specifically carbon dioxide (CO2) and nitrous oxide (N2O) in the potato cropping system, from two fields on Prince Edward Island, Canada. A novel interpretable glass-box intelligent framework was designed in the second part. This approach includes the best subset extra trees (BSET) feature selection, weighted aggregated sum product assessment (WASPAS), gradient-based optimization (GBO) algorithm, and k-nearest neighbours (KNN). The BSET-WASPAS feature selection method first attained four most appropriate input combinations for each target among existing seven input features. Afterwards, the optimal combinations were utilized to feed the KNN-GBO model. Additionally, three comparative machine learning (ML) approaches were considered to validate the main framework: leveraging the extra trees and GBO (Extra-GBO), classical KNN and Random Vector Functional Link (RVFL). The SHapley Additive Explanations (SHAP) tool was employed in the last phase to determine the contribution of features in the primary model. The WASPAS is used to individualize statistical metrics like correlation coefficient (R), root mean squared error (RMSE), and Reliability, aiming for easier and superior model identification. In monitoring CO2, the KNN-GBO|Combo 2, owing to its exceptional performance in terms of R = 0.9940, RMSE = 0.2426, Reliability = 97.2973, and WASPAS = 3.39E-6, outperformed the KNN, Extra-GBO, and RVFL, respectively. Moreover, in the N2O scenario, KNN-GBO|Combo 3 regarding (R = 0.9910, RMSE = 0.0940, Reliability = 91.7632, and WASPAS = 4.93E-6) resulted in the most promising performance compared to the KNN, Extra-GBO, and RVFL.