Near Real-Time Vineyard Disease Detection and Severity Estimation with Autonomous Robot Platform Integration

Authors: LIU, ERTAI - Cornell University; GOLD, KAITLIN - Cornell University; Cadle Davidson, Lance; KANALEY, KATHLEEN - Cornell University; COMBS, DAVID - Cornell University; JIANG, YU - Cornell University

Submitted to: Journal of Field Robotics
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/15/2025
Publication Date: N/A
Citation: N/A

Interpretive Summary: Agricultural robots have potential for mapping plant diseases such as grapevine downy mildew (DM). However, many farms lack the infrastructure for these analyses, so most disease mapping technologies have failed to be applied outside of research programs. We present DMNR, a computer vision model for near real-time grape DM disease quantification from high-resolution images. We deployed this DMNR on a fully autonomous ground rover carrying the sensors and computer needed for near real-time DM detection. Our approach balances efficiency and accuracy for on-robot computing, and has similar accuracy to off-robot research models for distinguishing fungicide treatments. This system could be used for continuous in-field disease monitoring, allowing farmers to detect and respond to disease outbreaks.

Technical Abstract: Agricultural robots have demonstrated their potential in accurate disease infection mapping, allowing for precision treatments of devastating diseases such as downy mildew (DM), a disease that threatens the grape and wine industry globally and causes significant losses yearly. However, existing infrastructure limitations in rural areas and design constraints of robotic platforms often confine data processing to labs, hindering real-time field operations of these robots. We present DMNR, a semantic segmentation model tailored for near real-time disease segmentation in high-resolution images. Compared to state-of-the-art real-time models, our approach achieves an optimal balance between efficiency and accuracy on the grape DM dataset using embedded computing devices. The disease severity estimation pipeline based on this model demonstrates comparable accuracy in distinguishing fungicide treatments as offline semantic segmentation models. Furthermore, we developed a fully autonomous ground rover carrying the advanced sensing system and the proposed data processing pipeline for continuous in-field disease monitoring. The successful disease mapping throughout the season showcased its potential in advancing agricultural research and disease management.