With increasing development of precision-agriculture tools there’s greater need to develop data-analysis solutions that can guide management decisions in real time. A new study at the University of Illinois offers an approach to efficiently and accurately process precision-agriculture data.
We're trying to change how agronomic research is conducted. Instead of establishing a small field plot, running statistics and publishing the means, we’re trying to involve farmers more directly. We‘re running experiments with farmers' machinery in their fields. We can detect site-specific responses to different inputs and see whether there's a response in different parts of the field.
We developed methodology using deep learning to generate yield predictions. It incorporates information from different topographic variables, soil conductivity, and nitrogen- and seed-rate treatments applied throughout nine Midwestern corn fields.
We worked with 2017 and 2018 data from the University of Illinois' Data Intensive Farm Management project. In that project seeds and nitrogen fertilizer were applied at varying rates across 226 fields in the Midwest as well as Brazil, Argentina and South Africa. To predict yield on-ground measurements were paired with high-resolution satellite images from PlanetLab.
Fields were digitally divided into 5-meter – 16-foot – squares. Data on soil, elevation, nitrogen-application rate and seed rate were fed into the computer for each square. The goal was learning how the factors interact to predict yield in each square.
We used machine learning known as a convolutional neural network. Some types of machine learning start with patterns and ask the computer to fit new bits of data into those existing patterns. Convolutional neural networks are blind to existing patterns. Instead they take bits of data and learn the patterns that organize them, similar to the way humans organize new information through neural networks in the brain. The convolutional neural network process, which predicted yield with great accuracy, also was compared to other machine-learning algorithms and traditional statistical techniques.
We don't really know what’s causing differences in yield response to inputs across a field. Sometimes people think a certain spot should respond very strongly to nitrogen and it doesn't or vice versa. The convolutional neural network can detect hidden patterns that may be causing a response. When we compared several methods, we found that the network worked very well to explain yield variation.
Using machine learning to untangle data from precision agriculture is still relatively new. Our experiment merely grazes the tip of the iceberg in terms of a convolutional neural network’s potential applications. Eventually we could use it to develop optimum recommendations for a given combination of inputs and site constraints.
Nicolas Martin is an assistant professor in crop sciences at the University of Illinois. He is a co-author of the study.