Feedback: Machine Vision

This page focuses on the AI/ML and machine vision stream from registrant feedback: specific tools for potato storage, disease diagnostics, optical sorting, deep learning and hyperspectral imaging, robotics, seed cutting, and which programs are proven in ag. For the business and workflow stream, see Business Processes.

One question from registrants

“What specific AI tools can be used for potato storage?” Options range from IoT sensors plus dashboards to AI-assisted simulation and code (e.g. exploring setpoints before buying equipment). See the Case Study: Storage and AI/ML & Machine Vision page; cost tiers are covered in Session 3 Q&A: Cost & investment.

What you asked about

Registrants are already using or planning deep learning for image-based classification, hyperspectral imaging, and optical sorters. You asked about: AI tools for potato storage; how to use AI in plant disease diagnostics; which AI programs have been proven in agriculture; real-world applications you can incorporate today; development of seed cutting technologies; and how AI might complement CRISPR/Cas gene editing. Some of you are excited about AI robotics for labour (e.g. dirt handling at harvest). Below we address what’s out there, how it works (including sensors to AI and ETL), data and vendors, and pathways with links for further reading.

What’s out there

Machine vision uses cameras and AI/ML to “see” and interpret images: defects, grades, disease, or guidance for robots. In potato, the most established applications are optical sorters on packing and processing lines—cameras and ML classify tubers by size, shape, colour, and defects in real time. Plant disease diagnostics from images (leaves, stems, tubers) are in research and early commercial use: you capture an image, an AI model suggests likely diseases, and you follow up with lab or agronomist confirmation. Storage often combines sensor data (temperature, humidity, CO₂, O₂) with ML to predict quality, sprouting, and disease risk and to recommend setpoints—see the Case Study: Storage. Hyperspectral imaging captures many wavelengths of light to detect quality or stress not visible to the eye; it’s used in research and increasingly in grading and storage. Robotics and automation—harvest aids, seed cutting, handling—rely on machine vision so equipment can “see” and react. Research applications include AI alongside CRISPR/Cas (sequence analysis, editing strategies) and seed cutting development (automation + vision for cut quality). The AI/ML & Machine Vision page (Session 2) goes into detail on each of these.

“Which AI programs have been proven in agriculture?” Optical sorters are long-standing; ML and vision have improved their accuracy and speed. Yield prediction and crop models, disease/pest detection from images, and sensor-driven storage management are in use and expanding. “Proven” depends on crop, scale, and region—the Resources section on Potato + Data (Presia, Scale AI Precision Harvest, and open-access papers) gives you a window into what’s being deployed and studied in potato and Canadian ag.

How it works: sensors, data, ETL, and AI models

For vision-based applications (sorting, disease ID), the pipeline is: cameras capture images → images are preprocessed (resize, normalize, sometimes label) → an AI/ML model (e.g. a convolutional neural network trained on many labelled examples) produces a prediction (grade, defect, disease class) → that output is used for sorting decisions, alerts, or reports. For sensor-based applications (e.g. storage), the flow is: sensors (temp, humidity, CO₂) collect readings → data is extracted and moved into a system → it’s transformed (cleaned, aligned to timestamps, units) and loaded into a database or analytics platform—that’s ETL. Then ML models run on that structured data to predict quality, recommend setpoints, or flag anomalies. Optionally, GenAI turns those results into plain-language summaries and alerts. The AI in an ETL Pipeline page explains ETL and where ML and GenAI plug in, with a potato example; the Case Study: Storage shows sensors → data → decisions in a real scenario.

You don’t have to build this yourself. Vendors offer turnkey optical sorters, disease apps, and storage monitoring packages. The trade-off is convenience versus control and potential vendor lock-in (see below).

Data requirements and vendors

Vision models need many labelled images for training (e.g. “this leaf has late blight,” “this tuber is defective”). Accuracy depends on data quality, diversity (lighting, angles, varieties, conditions), and whether the model was trained on data similar to yours. Ask vendors: What data did you train on? How do I add my own images or varieties? What’s the expected accuracy in my conditions, and how do you measure it? For sensor-driven ML (e.g. storage), you need consistent, timely sensor data; historical outcomes (e.g. grade, loss) improve predictions. Data ownership and export matter: can you get your data out in a standard format if you switch vendors?

Vendor lock-in in machine vision and sensor systems can be technical and commercial: proprietary camera setups, closed APIs, models trained on your data but owned by the vendor, and long-term support contracts. Before committing, ask: Who owns the data and the trained model? Can I export raw data and predictions? What happens to my data if I leave? For optical sorters and integrated equipment, also ask about compatibility with your existing line and upgrade paths. The Resources page doesn’t endorse specific vendors but points to initiatives (e.g. Scale AI Precision Harvest, Presia) and research (e.g. MDPI deep learning in potato production) so you can see what’s in use and what questions to take to suppliers.

Pathways and answering your questions

Real-world applications you can incorporate today: If you already have grading or sorting, optical sorters with ML are the most mature. For disease, consider a scouting app or research trial that uses image-based AI as a first filter—always confirm with expert or lab. For storage, the Case Study: Storage walks through a decision process: exploring sensors and data before a big vendor commitment, and how AI can help simulate and summarize. Seed cutting and robotics: Development in seed cutting often involves automation and machine vision to guide cutters and assess quality; robotics for harvest (e.g. dirt handling) is advancing but still emerging—follow industry and research updates. AI and CRISPR/Cas: In research, AI supports sequence analysis, gene function prediction, and editing strategy optimization; this is an active area at the intersection of biotech and ML. The AI/ML & Machine Vision page has more on each of these and on “using AI to search for solutions” (finding papers, comparing tools, drafting plans).

A practical pathway: identify one problem (grading, disease scouting, or storage), see what data or images you already collect, then explore which tools and vendors could use that data. Ask about accuracy, data requirements, and integration with your equipment. Use the Resources and session pages to prepare those conversations.

AI and the Potato Industry