Ex-ante analyses using machine learning to understand the interactive influences of environmental and agro-management variables for target-oriented management practice selection

Abstract

Conservation management in dryland agriculture preserves water, improves soil health and yields. To comprehend the complex interactions of conservation management and environmental factors in a rainfed forage system of the US Great Plains, distinguish the superior influence of conservation over conventional management, and have a different perspective from simulation modeling, machine learning (ML) and artificial intelligence models were adapted in 2022. The variables in this study included ten years of daily recorded weather data and yield values simulated by the DSSAT model suite, considering four years of actual data on aboveground and belowground biomass, depth-wise carbon, water content, various physicochemical soil parameters, and management practices (Sarkar and Northup 2023). Two optimized ML models, Random Forest and AdaBoost, were found to perform better, when the algorithms of six ML models- namely Decision Tree, Random Forest, Bagging, Gradient Boosting, AdaBoost and XGBoost were tuned with different hyperparameters, validated and trained before predicting the biomass yields. Feature Importance plotting by these two models revealed the five most influencing similar variables, which were in different orders: average maximum temperature during daylight hours, total soil water, seasonal average minimum temperature, cumulative potential evapotranspiration and CO2. Hence, SHapley Additive exPlanation (SHAP) algorithm was adopted to dive into the database and clarify the interaction effects of management practices especially tillage and soil cover with different environmental variables. Interestingly, the SHAP model indicated soil cover as the 5th most important variable, followed by maximum temperature during daylight hours, cumulative potential evapotranspiration, seasonal minimum temperature and CO2. The interaction plotting of SHAP analysis also manifested that intensity of tillage and use of no soil cover could be detrimental. Considering the rising atmospheric CO2 levels and temperatures, along with depleting soil water, no-till practices with a springtime cover of grass peas or field peas and the addition of 100 % residue can be acclaimed for high water-use efficiency and increased aboveground biomass of rainfed sorghum sudangrass in drylands. We recommend using impeccable dataset, particularly from diverse agro-environmental systems with various tillage practices and soil covers, before regional adoption. Additionally, exploring the impacts on diverse soil types is advisable before selecting a sustainable management strategy for precision agriculture.