Assessment of Advanced Machine and Deep Learning Approaches for Predicting CO2 Emissions from Agricultural Lands: Insights Across Diverse Agroclimatic Zones

作者：Endre Harsányi, Morad Mirzaei, Sana Arshad, Firas Alsilibe, Atilla Vad, Adrián Nagy, Tamás Rátonyi, Manouchehr Gorji, Main Al-Dalahme, Safwan Mohammed · 发表于：Earth Systems and Environment · 年份：2024 · DOI：10.1007/s41748-024-00424-x · 被引用次数：42 · 研究领域：Soil and Unsaturated Flow、Soil Carbon and Nitrogen Dynamics、Fire effects on ecosystems

Abstract Prediction of carbon dioxide (CO 2 ) emissions from agricultural soil is vital for efficient and strategic mitigating practices and achieving climate smart agriculture. This study aimed to evaluate the ability of two machine learning algorithms [gradient boosting regression (GBR), support vector regression (SVR)], and two deep learning algorithms [feedforward neural network (FNN) and convolutional neural network (CNN)] in predicting CO 2 emissions from Maize fields in two agroclimatic regions i.e., continental (Debrecen-Hungary), and semi-arid (Karaj-Iran). This research developed three scenarios for predicting CO 2 . Each scenario is developed by a combination between input variables [i.e., soil temperature (Δ), soil moisture (θ), date of measurement (SD), soil management (SM)] [i.e., SC1: (SM + Δ + θ), SC2: (SM + Δ), SC3: (SM + θ)]. Results showed that the average CO 2 emission from Debrecen was 138.78 ± 72.04 ppm ( n = 36), while the average from Karaj was 478.98 ± 174.22 ppm ( n = 36). Performance evaluation results of train set revealed that high prediction accuracy is achieved by GBR in SC1 with the highest R 2 = 0.8778, and lowest root mean squared error (RMSE) = 72.05, followed by GBR in SC3. Overall, the performance MDLM is ranked as GBR > FNN > CNN > SVR. In testing phase, the highest prediction accuracy was achieved by FNN in SC1 with R 2 = 0.918, and RMSE = 67.75, followed by FNN in SC3, and GBR in SC1 (R 2 = 0.887, RMSE = 79.881). The performanc...