Screen Shot 2019-10-22 at 2.06.00 PM.png

Seed Company “A” Case Study

AI Machine Learning Models for Image Recognition in Plant Pathology

Objective

Develop, test, and demonstrate pest and pathogen identification capabilities for nine targets specified by the client. Test both existing Farmwave AI models and time to deployment and accuracy of data models for new targets.


Approach

Leverage Cloud Optimized recognition Engine (CORE) image recognition algorithms, some extant and some which were developed specifically for individual targets. Obtain imagery from selected pathologists for building AI models on severity rating based on their expertise.


Results

On multiple trials, we were able to demonstrate that our technology could learn and produce results with 99% accuracy or better on existing Farmwave models in corn. New deployment models on Cucumber Downy Mildew were 69% accurate with only 1,400 images and 48 hours of AI model development.

The data set was incomplete with little, (20%), severity classification given by SME’s prior to training. Potato Early Blight resulted in 67% accuracy with 2,800 images with zero classification set by SME’s prior to training. Given the learning obtained, a simple classification from SME on both of these models and minor retraining would yield a 95% or higher accuracy.


Key Learnings

The effort required to solve additional problems with this technology would essentially require the development of new algorithms due to the complexity, differences, and variability of inherent characteristics specific to each target. Three key segments require optimization: data, (imagery itself or proper quality and resolution), The severity classification process and scale rating (which can differentiate between organizations thus setting a standard) and the AI models and algorithms themselves.