Indian Researchers Unveil an AI System That Detects Disease in Grape Leaves

Researchers in India have reported a new artificial intelligence system designed to identify diseases in grape leaves from images, a tool they say could help growers detect problems earlier and respond before damage spreads through vineyards.

The study was published Monday in Scientific Reports. It describes a model called HASFT-Net, short for Deep High Attention Stage-by-Stage Forward Taylor Network, developed to classify grape leaf diseases using image analysis and several layers of computer vision processing.

According to the paper, grape leaf disease can be caused by different pathogens that lead to discoloration or deformities in leaves. The authors said many existing methods detect disease only after visible symptoms appear, which can delay treatment and reduce yields. Their goal was to improve classification performance by combining image cleanup, segmentation of affected areas and feature extraction before the final disease prediction step.

The research team included Lalitha R of Rajalakshmi Institute of Technology in Chennai, Sivasangari Ayyappan of GMR Institute of Technology in Andhra Pradesh, Saranya K of Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology in Chennai, and A. S. Malleswari of Aditya University in Andhra Pradesh. The corresponding author is Malleswari. The paper states that the authors declared no competing interests.

In the system described by the researchers, images of grape leaves were taken from an existing database rather than collected directly in vineyards for this study. The images first went through a denoising stage using what the paper calls an Attention-Guided Convolutional Neural Network, or ADNet. After that, diseased regions were segmented with a Spatial Pyramid-oriented Encoder-Decoder Cascade Convolutional Neural Network, identified in the paper as SPEDCCNN, using a Tversky loss function. The segmented output was then refined with a method called Siberian Tiger Optimization.

The model also extracted several visual features from the leaf images, including Opponent Color Local Binary Pattern, Scale Invariant Feature Transform and shape measures such as rectangularity, eccentricity and circularity. Those inputs were then used by HASFT-Net for classification. The authors said the architecture was built to combine elements from NASNet, Taylor series methods and a Deep High-order Attention Neural Network.

In the results reported by the paper, HASFT-Net reached 92.172% accuracy when trained with 90% of the learning data. The study also reported a true positive rate of 92.883%, a true negative rate of 91.993%, precision of 90.887% and an F1-score of 91.874%.

Those figures suggest the model performed well under the conditions tested, though the paper does not establish how it would perform across all vineyard environments, lighting conditions or grape varieties in commercial use. Because the work relied on a database of images, further validation in field conditions would likely be needed before growers could depend on it as a routine management tool.

That matters for the beverage sector because grapes are the raw material for wine production, and disease pressure can affect both yield and fruit quality. If systems like this prove reliable outside laboratory-style testing, they could support more precise vineyard management, help growers intervene sooner and potentially reduce crop losses. Earlier detection could also improve decisions on when and where to apply plant protection treatments, an issue with direct implications for costs and sustainability in winegrowing.

The paper was received on April 2, accepted on July 6 and published on July 27. It appears as open access under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

The study adds to a growing body of work applying machine learning to agriculture, especially in crop health monitoring through image recognition. In viticulture, that trend has drawn attention because disease outbreaks can move quickly and because visual inspection across large vineyard areas is labor-intensive. A system that can flag suspect leaves from photographs could eventually be used alongside drones, mobile phones or fixed cameras, although those applications were not tested in this paper.

For vineyard operators and wine producers, the practical question is not only whether an algorithm can classify diseased leaves in a research setting but whether it can do so early enough and consistently enough to change field decisions. This study does not answer that broader operational question on its own, but it offers another example of how AI tools are being built around one of viticulture’s most persistent challenges: spotting disease before it cuts into harvest volume or grape quality.