Multi‐Vine Disease Prediction in a Field Test Spanning Whole Growth Season at Wine‐Industrial Site (McLaren Vale)

Abstract

ABSTRACT Against the backdrop of increasing climate variability and rising disease pressure, vineyard disease monitoring that relies solely on manual scouting often suffers from low efficiency, subjectivity, and delayed response. Vineyard disease management involves complex climate–pathogen–crop interactions, where fixed spray calendars fail to capture these interactions. This study was conducted in a commercial, vineyard in McLaren Vale, Australia, and implemented a digital twin (DT) supported by multi‐source monitoring across the 2025 growing season, from budburst (2 September 2025; later than usual) through flowering (12 November 2025; colder‐than‐average spring). A total of five image acquisition events were conducted using fixed cameras and drone surveys. In addition, continuous data from soil sensors and weather stations (temperature, humidity, and rainfall) were also recorded, providing ongoing environmental information. Soil sensing indicated near‐neutral conditions (pH ≈ 7), while soil moisture at 20 cm depth varied between 30% and 50%. The prediction model adopts a convolutional neural network (CNN) to perform multi‐class classification across four states: black rot, black measles (Esca‐related symptoms), leaf blight, and healthy. The results show that the model achieves strong classification performance on the test set (accuracy = 0.99). Error‐based evaluation of the model outputs further indicates stable predictions (mean squared error [MSE] = 0.0004, root mean squared error [RMSE] = 0.0197, coefficient of determination [ R 2 ] = 0.998). To verify the field deployment capability, we deployed the grape‐leaf disease prediction model and applied it to automatically process in‐field leaf images, producing a predicted disease label and confidence score for each leaf/leaf group. Across four drone surveys, the tile‐based CNN aggregation revealed a clear shift in dominant risk signals: Esca severity rose steadily from about 17% in the first survey to about 62% in the final survey, whereas black rot severity stayed low, starting near 0% and briefly peaking around 14% before dropping to roughly 5% thereafter. Leaf blight severity was highest at the start at approximately 47%, then decreased to around 19% and subsequently fluctuated between roughly 21% and 31%. Overall, the findings provide reliable evidence for block‐level disease early warning, inspection prioritisation, and spray decision‐making, helping to reduce unnecessary inputs, lower environmental burdens, and improve the resilience and sustainability of vineyard production systems.

Plant phenotyping relevance

ブドウ葉の画像から病害状態・重症度をCNNで推定し、固定カメラとドローンによる圃場展開および性能評価まで行っており、植物表現型取得・抽出法が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.