From Artificial Intelligence to Plant Disease Surveillance: Pl@ntAgroEco’s Work Featured in the First ESV Bulletin

The National Plant Health Epidemiological Surveillance Platform (ESV) is publishing its first Methodological Bulletin dedicated to a technology that is opening up new perspectives for plant disease surveillance, particularly through artificial intelligence models and, more specifically, deep learning models applied to image analysis.

Plant diseases and other phytopathological disorders can pose significant risks to plants and have major consequences for crop yields and production quality, and therefore for food security and the sustainability of agricultural systems. Their surveillance currently relies largely on visual observation of symptoms, supplemented, when necessary and where possible, by sampling and laboratory analyses. While these approaches form a crucial foundation of plant health surveillance systems, they can be time-consuming and costly. A major challenge also lies in their scalability and large-scale deployment. The development of new methods for the faster detection of certain symptoms could therefore provide significant benefits by strengthening surveillance capacities and enabling wider deployment.

The progress made in artificial intelligence, particularly in Deep Learning and Computer Vision approaches, offers promising and innovative responses to these challenges.

By automatically analysing images of plants or leaves, some models are now able to identify visual features associated with diseases. These approaches can be applied to images captured using smartphones, drones or different types of sensors, opening the door to new approaches to field-based surveillance.

The first Methodological Bulletin of the International Health Surveillance Working Group (GT-VSI) of the ESV Platform provides an accessible introduction to these approaches. Its aim is to help non-specialist audiences better understand how these models work and the principles and considerations underlying their architectures, but above all, to understand the potential benefits and limitations of using them for plant disease surveillance. The bulletin, which we invite you to read HERE, presents three major families of models: Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and hybrid architectures combining the two approaches.

The bulletin also includes summaries of three recent scientific studies. The first focuses on the detection of Huanglongbing (HLB) in citrus trees using drone-acquired images and a CNN-based architecture. The second examines a lightweight Vision Transformer model designed to enable disease identification directly on mobile devices. The third presents a particularly interesting approach for the work being carried out as part of the Pl@ntAgroEco project. This research, in which Pl@ntNet is heavily involved, is being conducted in close collaboration with our Malaysian partners at Swinburne University’s Sarawak Campus.

This work focuses on PlantAIM, a hybrid model developed to simultaneously identify the crop plant and the disease affecting it from a single image. The originality of this approach lies in combining two types of information: local symptom features, captured in particular through CNNs, and more global information about the overall appearance of the plant, captured by Vision Transformers. The model therefore seeks to establish a link between the identity of the plant and the symptoms observed.

This study clearly illustrates the value of models capable of simultaneously taking into account the overall context of an image and the distinctive details of a symptom. The results presented show that this architecture can maintain good performance, even when only limited data are available for certain crop–disease combinations.

Beyond the performance of any particular model, this study highlights a key question for the development of surveillance support tools: how can we enable a model to learn to recognise a disease while also taking into account the plant on which it occurs? This is a particularly important question in agricultural environments, where symptoms can vary between species, but also as a result of environmental conditions or even the circumstances in which images are captured.

Of course, as the Methodological Bulletin also points out, artificial intelligence is not a complete or definitive solution to all the challenges associated with diagnosis and surveillance. Model performance still depends heavily on the quality, diversity and representativeness of the images used for training. A model that performs well on dataset A may encounter difficulties when confronted with dataset B, which may include new field conditions, different symptoms or situations that were not represented in the training data. Understanding and interpreting the decisions made by these models is also an important challenge if they are to be trusted sufficiently in operational settings.

The value of these technologies therefore lies in their ability to complement existing surveillance and diagnostic methods and practices. AI can help strengthen surveillance systems, complement human expertise and observer networks, and potentially improve both the timeliness and geographical coverage of detection. However, it remains a tool within a broader and essential ecosystem.

With this first issue, the ESV Platform provides an accessible introduction to a rapidly developing technological field. Through its educational explanations and forthcoming bulletins, it will help readers gain a better understanding of the challenges and opportunities associated with these new approaches, including the work we are developing as part of Pl@ntAgroEco.