Ex-ante analyses using machine learning to rank variables of interactions among soil, atmospheric and agricultural management

Authors: SARKAR, RESHMI - Texas A&M Agrilife; Northup, Brian; LONG, CHARLES - Texas A&M Agrilife

Submitted to: European Journal of Agronomy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/6/2024
Publication Date: 11/18/2024
Citation: Sarkar, R., Northup, B.K., Long, C. 2024. Ex-ante analyses using machine learning to rank variables of interactions among soil, atmospheric and agricultural management. European Journal of Agronomy. 162. Article 127432. https://doi.org/10.1016/j.eja.2024.127432.
DOI: https://doi.org/10.1016/j.eja.2024.127432

Interpretive Summary: Conserving soil moisture is an important practice for producers in the Southern Great Plains for crop production in drought-affected environments. Two important tools for saving moisture are systems of conservation tillage, and the use of cover crops or green manures. These management techniques can preserve soil water, provide soil organic matter, and enhance nutrient cycles which improves crop yields. However, it can be difficult to understand how forms of management affect these factors, given the larger effects of growing conditions, and how these conditions interact with management. To understand how interactions between management and the environment influenced crop yields, we applied a series of techniques in machine learning to help refine an understanding of the most important variables that drive forage production in a rainfed system. We applied six different models of machine-learning, and other data processing tools, to predict yields generated by sorghum-sudangrass that was managed with different tillage systems and types of cover crops. The goal was to define the impacts of forms of management, soil conditions, and growing conditions on forage yields. We found two optimized models (Random Forest and AdaBoost) performed best. Both models showed the five most influential variables on yields were the same, but the two models showed different orders of importance for these variables. The most important variables were; average maximum temperature during daylight hours, total amount of soil water, average minimum-temperature during growing seasons, cumulative potential evapotranspiration (movement of water from soil and plants to the atmosphere), and concentrations of carbon dioxide in the atmosphere. The different forms and types of management that were applied were among the least important variables. A deeper series of tests were then applied to fully examine how tillage systems and cover crops influenced forage yields. These tests showed cases where using cover crops was the fifth most important factor influencing yields, though air temperatures, potential evapotranspiration and atmospheric carbon dioxide still had much greater effects. Combining cover crops with conservation tillage showed small influences, though tillage systems were still among the least influential feature. While the environment in which crops are grown is the primary driver of forage yields by sorghum-sudangrass in rainfed agriculture, no-till combined with a spring-grown cover crop may provide small improvements in water use efficiencies and yields of sorghum-sudangrass. We suggest more tests of data from production systems that include a wider variety of tillage systems and types of cover crops, to identify management practices that are adapted to the Southern Great Plains.

Technical Abstract: Methods of conservation management may preserve soil water, improve soil health, and enhance crop yields in the U.S. Southern Great Plains. However, the effects of forms of management on these factors can be difficult to define, given the large-scale effects of environmental conditions, and how these conditions interact with management. To comprehend subtle interactions in the impacts of management and environmental factors on crop yields, an ex-ante analysis using machine learning was used to refine an understanding of the important variables in a rainfed forage system. Six models of machine-learning were tuned, validated, and trained to predict biomass yields of systems producing sorghum-sudangrass, with an aim to define the impacts of management, simulated soil, and environmental parameters. Optimized models of Random Forest and AdaBoost performed best of the six tested. Plots of Feature Importance for these models revealed the same five most influential variables, but the models arrayed the variables in different orders: average maximum temperature during daylight hours (Avg-TMAX-DL), total soil-water (TSWD), seasonal average minimum-temperature (Avg-TMIN-SN), cumulative potential evapotranspiration (CumPET), and carbon dioxide concentrations (CO2) in the atmosphere. Forms of management were among the least influential. A deeper test with the SHapley Additive exPlanation (SHAP) algorithm was applied to clarify the effects of management practices involving tillage systems and the use of cover crops. These results showed use of cover crops was the 5th most important variable affecting yields, though Avg-TMAX-DL, Avg-TMIN-SN, CumPET and CO2 had greater impacts. Combining cover crops with conservation tillage showed some beneficial influence, though tillage systems were among the least influential attribute. While environmental factors are the primary driver of forage production in dryland agriculture, no-till combined with a springtime cover provided by annual legumes may improve high water use efficiency and biomass yields of rainfed sorghum-sudangrass. Analyses of data from multiple agri-environmental systems that are managed under a wider variety of tillage systems and types of cover crops are required to fully identify regionally adapted management practices.