Andrew Cunningham, James Baumeister, Daniel Ablett, Chris Thomas, Tom Raimondo
Problem Statement
VineSight is a SAAFE CRC Kickstart project developing a tablet-based augmented reality tool that helps non-expert vineyard workers flag possible disease markers in the field, geo-located and recorded for an expert agronomist to confirm later. The starting problem, drawn from our industry partner Dowie Doole Winery in McLaren Vale, is that disease monitoring is a bought service costing roughly $27,000 a season across 27 blocks — a cost that does not vary with what is found, and so the first thing cut when budgets tighten. Less monitoring means later detection, which means whole-block spraying rather than targeted treatment, which raises chemical costs and squeezes the monitoring budget further. Our aim is to break that loop by making the casual workers who already walk every row into a useful first line of observation, without asking them to diagnose anything.
Progress to Date
Progress to date falls in three areas. Stakeholder interviews with our industry partner shaped the design substantially: workers and growers preferred a phone or tablet over a head-worn display, powdery mildew — the region’s main concern — turns out to be smelled before it is seen, and the partner proposed the comparative field trial design himself. On the detection side, a four-class leaf classifier (black rot, esca, leaf blight, healthy) built on public image data reached approximately 91% in offline testing after fine-tuning and a weighted ensemble, up from around 72%; we expect lower accuracy in real vineyard conditions and are collecting powdery mildew imagery. A working tablet demonstrator now runs the classifier on the live camera feed with on-screen guidance prompting the worker to look closer, turn leaves, and cover vines not yet scanned. Remaining work over the final months is the comparative field trial at Dowie Doole, extension to visually distinctive diseases such as downy mildew and botrytis, and scoping multi-modal sensing for the diseases a camera alone will always find late.