Menu Close
 

Optimization of Nutrient Management using Convolutional Neural Networks and Transfer Learning

Project Team

Abstract

This proposal seeks to expand the domains of Data and Automation impacting People in Agriculture through decision-making processes. It extends preliminary research from PIs where results showed that the convolutional neural network (CNN) based learning methodology got up to 65% reduction on the test dataset RMSE value when compared to multiple linear regression, and up to 38% when compared to a shallow fully connected neural network. Optimizing crop nutrient management is of great importance for increasing food production and reducing environmental impact. It poses a challenging problem since crop yield response depends on many environmental and soil properties. Moreover, the spatial structure at different scales of such properties is known to have a significant impact on the resulting yield. This proposal encompasses an optimization algorithm based on a predictive CNN model for yield response to nutrient management and then using this response determines optimum rates based on the constraints. In this proposal, the PIs propose to investigate and develop an algorithm to apply deep learning methods to extract features that are relevant across different fields and use a technique called Transfer Learning to optimize prescription maps for fields where on-farm experiments had not been conducted. An additional objective is to investigate optimization algorithms based on models, where the manageable variables (i.e. nitrogen and seed rates) are relevant in the context of other seasonal environmental factors.