A convolutional neural network model and algorithm driven prototype for sustainable tilling and fertilizer optimization

Abstract

Abstract Tilling, a common agricultural practice, is being done excessively on farms leading to about 2.35 billion tons of soil erosion from US croplands annually. This causes soil erosion, soil infertility, carbon release, nutrient runoff, and fertilizer over-usage. This paper evaluates whether optimizing tillage intensity, timing, and fertilizer quantity will address these problems. A convolutional neural network based machine learning model utilizes a camera-captured field image to determine existing tilling intensity on a 7-point scale. This machine learning output, along with soil sensor and external forecast data, flows into a 10-parameter algorithm that determines optimal tilling and fertilizer levels. A fully functional tractor prototype demonstrates the above. A 30-year simulation comparing conventionally-tilled and algorithm-tilled farms showed a reduction in carbon emission by 57%, fertilizer usage by 43%, and runoff by 86% demonstrating the transformative potential of this algorithm. Additionally, a stationary prototype was deployed in 155 farms across 5 countries.