Location: Tropical Crop and Commodity Protection Research
Project Number: 2040-22430-027-013-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Jul 3, 2023
End Date: Jan 2, 2025
Objective:
We propose to facilitate the diagnosis and remediation of coffee pests, diseases, and nutrient deficiencies/toxicities by producing an AI-driven app through integration of expert experience and machine vision with extensive field validation. The app resulting from our efforts will provide significant decision support to coffee growers in Hawaii and other coffee-growing regions that are seeking to maximize their production quality and yield.
Approach:
We will focus on the most serious pest, disease, and nutritional problems that can be detected via image analysis of the coffee plant leaf. A digital library comprising thousands of curated images collected from affected plants in commercial coffee farms in Hawaii will be created to document pests, diseases, and nutrient deficiencies. Diagnosis will be verified using a combination of expert knowledge, molecular testing, and soil/leaf tissue analyses. For image analysis, the key methods we will apply are deep-learning and computer vision, training a convolutional neural network and integrating information from multidimensional datasets beyond images, namely: climate, topography, geology, phytomorphology, and agronomics. We will leverage SCINet’s Ceres cluster with the machine-learning library Keras to train a working model. Keras is a highly abstracted wrapper based on several deep-learning technologies, including Tensorflow. This leads to not only a versatile model, but a substantially portable one as well. The working model can then be ingested through Amazon Web Service’s (AWS) SageMaker, a machine-learning service which can deploy the Keras model. From this, we can create an endpoint which can interface with a mobile app (see Figure 1). Put simply, the model we train on SCINet will be portable, and capable of being run on various web services or potentially even on local IoT endpoints.
