Decision support tools for farm management by integrating domain knowledge with machine learning and optimization techniques

Abstract

By 2050, the global population is expected to reach 10 billion, while arable land remains unchanged. Meanwhile, food prices have been rising in recent years, highlighting the need for innovative solutions to ensure an affordable food supply. Inefficient farming systems may either reduce productivity or increase production costs. This dissertation proposes tools to aid farming decisions by utilizing historical data, machine learning models, and optimization techniques. Inadequate farming strategies result in significant food wastage, with approximately 30-40% of produced food being wasted after harvest yearly. In the second chapter, a planting scheduling tool was developed to minimize this wastage. Random planting causes uneven weekly harvests, sometimes exceeding storage capacity, leading to waste. The tool predicts Growing Degree Units (GDUs) to determine harvest readiness and uses this information in an optimization model to create a planting schedule and determine optimal storage capacity. Applied to the Syngenta Crop Challenge 2021, the GDU prediction model achieved a relative root mean square error (RRMSE) of 7-8%, and the scheduling model ensured uniform weekly harvests with no food wastage. A key component of precision farming is the accurate prediction of crop yields and understanding the interactions among various factors. This insight aids in making informed decisions and ensuring a sustainable food supply. The third chapter of the dissertation developed a crop yield prediction model for the US corn belt at the county level. This model combined APSIM, a crop simulation model, with a machine learning model to create an optimized weighted ensemble model (RRMSE 9%) using weather, soil, and management information. The concept of yield prediction was further extended in the fifth chapter to the field level by incorporating genotype data. A multimodal CNN-DNN model was designed using genotype, environment, and management interaction. Further analyses were conducted to understand how these factors interact, leading to recommendations for increasing productivity. Applying fertilizer is essential in crop production; insufficient amounts lower productivity, while excessive amounts raise costs. The production cost needs to be optimized to ensure food remains affordable and profitability is sustained. In the fourth chapter, we developed a tool to recommend the economic optimum nitrogen rate (EONR). This tool used APSIM simulations to train a machine learning (ML) model to predict yield responses for different nitrogen rates. The quadratic plateau model fits these predictions into a closed form of yield response, which was used in an optimization model to recommend EONR (192KgN/ha ± 48KgN/ha) for various scenarios. In brief, this dissertation presents a planting schedule tool to reduce food waste, yield prediction models considering G × E × M interactions for county and field levels, and an optimal fertilizer tool using machine learning, neural networks, and optimization methods. These tools aim to enhance farming decisions and ensure global food security.