Development of Neural Networks and Performance Appraisal of Supervised Learning Models for Predicting Organic Carbon in Soils Under Different Cropping Systems

Accurate prediction of soil organic carbon (TOC) is crucial for assessing soil health, fertility, and ecosystem functions and services. The purpose of the present study was to evaluate the performance of various supervised learning models in comparison to neural networks for accurately capturing the complex relationships between soil attributes and TOC in soils under three wheat-based ecosystems (viz. wheat-rice/maize/cotton). The study aimed to identify the most robust and precise models for predicting TOC levels, with the ultimate goal of informing sustainable land management practices and improving soil fertility and health in agricultural systems. We evaluated the performance of eight supervised learning models viz. linear (LR), ridge (RR), lasso (L_AR), elastic net (E_NR), decision tree (DT), random forest (RF), gradient boosting (GBR) and support vector regression (SVR) vis-à-vis artificial neural networks (ANNs) for capturing the complex relationships between soil attributes and TOC in soils under three wheat-based ecosystems. These results revealed that the proportion of micro-aggregates (< 0.25 mm) was significantly (p < 0.05) higher in soils under rice-wheat system, compared with the other systems which themselves did not differ significantly. However, there was a non-significant (p < 0.05) difference in the proportion of macro-aggregates (> 0.25 mm) amongst the compared wheat-based systems. The proportion of total water stable aggregates (WSA) was significantly higher by ~ 13.1–18.7% for maize-wheat, compared with other systems. Maize-wheat soils exhibited ~ 21.6% reduction in dehydrogenase (DHA) activity as compared to rice-wheat soils. However, cotton-wheat soils exhibiting ~ 28.8% decrease in alkaline-P (Alk-P) activity, compared with rice-wheat soils. These results indicate that ANN_{8 − 15−1−1} model consisting of Tansig activation function and Levenberg-Marquardt (LM) algorithm exhibited superior performance in accurately predicting TOC levels. Model (ANN_{8 − 15−1−1}) validation of a comprehensive data-set underscored its robustness and precision with mean squared error (MSE) values of 1.055E-13, 4.950E-11 and 5.561E-03 for training, validation, and testing data-sets, respectively. Nonetheless, high values of the correlation coefficient (R; 0.9964**, 0.9845**, and 0.9989**; p < 0.01, respectively), underscored the model’s effectiveness in capturing the complex relationships between soil attributes and agricultural practices. Amongst the various supervised learning models evaluated, DT model outperformed, exhibiting the highest coefficient of determination (R²) of 0.997** (p < 0.01). These findings highlight the importance of employing advanced modeling techniques, e.g., ANN and supervised learning algorithms for precise prediction of TOC concentration in agricultural soils. These insights have significant implications for sustainable land management practices, enabling policymakers to make informed decisions regarding soil fertility and health, thereby contributing to the mitigation of land degradation and climate change impacts.