DefoliationCAM Uses Edge AI to Assess Insect Defoliation in Grapevines

Automated assessment in the vineyard

Insect-induced defoliation can threaten grapevine health and yield, while judging the severity of damage commonly depends on manual field scouting. Because that process can be subjective, the developers of DefoliationCAM, or D-CAM, designed an automated alternative intended to operate in real time.

The proposed platform is a solar-powered edge AI system. It processes imagery near the point of capture and then transmits its canopy defoliation estimates through LoRa to a web dashboard. This architecture combines image-based assessment and wireless reporting within a single field-oriented workflow.

A two-stage imaging pipeline

D-CAM works with 4K images. The first component of its processing pipeline is YOLOX-Nano, which localizes grapevine cordons in the images. A MobileNet-based classifier then evaluates 320 × 320-pixel patches using a lightweight channel-attention mechanism.

The dataset contains 29,674 patches collected in the field. They were labeled either as background or as one of four damage levels: no to negligible, mild, moderate and severe. The system is intended to place insect-related canopy damage into these categories and make the resulting assessment available in real time.

Using compact computer-vision models is consistent with the system’s edge-oriented design, although the supplied material does not give details about the computing hardware. Likewise, no specifications are provided for the solar power subsystem, the LoRa transmission range or the dashboard’s functions beyond receiving the results.

Limits of the available information

The available abstract and summary do not report accuracy metrics, comparative benchmark results or detailed latency measurements. They also do not identify where the field images were collected, which grape varieties were represented, which insect species caused the observed damage or how long the solar-powered installation was tested.

Without those details, the system’s practical performance and its ability to generalize across vineyards cannot be assessed from the supplied information alone. The material nevertheless outlines an integrated approach that brings together image localization, patch-level severity classification, off-grid power and long-range data transmission for vineyard monitoring.

The original publisher of the research is arXiv – Viticulture & Grapevine.