A new AI powered system, EdgePlantNet, has been developed by researchers from the Indian Institute of Technology (IIT) Patna, Indian Institute of Technology (IIT) Bombay and the Rajiv Gandhi Institute of Petroleum Technology (RGIPT), Amethi. The EdgePlantNet is designed as a lightweight, edge aware cyber physical tool to detect plant diseases in real time using image based analysis. Built specifically to run on affordable edge devices the tool is intended to serve as a practical solution in the field of smart agriculture, addressing the challenges of detecting crop diseases at the individual plant level without the need for cloud based infrastructure.
India, being the second largest country in terms of agricultural output and the largest in terms of crop cover, faces continuous threats to food security from plant diseases. These diseases can reduce crop yields significantly, with global crop losses due to plant diseases exceeding 20% annually. Farmers often resort to broad applications of fertilisers and pesticides to manage diseases, but such approaches can cause environmental harm and are not always effective against specific pathogens. A targeted strategy, based on early visual identification of diseased plants, can reduce dependency on chemicals and help in early intervention.
This is where smart agriculture comes in, leveraging digital tools and data driven technologies to monitor crop health more accurately. Technologies like Cyber Physical Systems (CPS), Internet of Things (IoT) and Artificial Intelligence (AI) are increasingly being applied to enable this precision. CPS and IoT involve inter connected sensors and devices, while AI models learn from data patterns to assist with decision making. However, the computational demand of such systems is often high, with most models depending on remote servers and cloud computing. This results in latency and increased costs, limiting the applicability of these systems in rural or resource constrained settings.
Edge Computing for Smart Agriculture
Edge computing, which processes data locally rather than sending it to a remote server, is emerging as a viable alternative for real time agricultural applications. For edge computing to be effective in agricultural contexts, AI models must be compact, energy efficient and capable of functioning on devices with limited memory and processing power. Recognising this need, the research team designed EdgePlantNet to address the trade offs between speed, accuracy and model size.
Model Design and Architecture
EdgePlantNet is built on a Convolutional Neural Network (CNN) architecture, known for its capabilities in processing and interpreting visual information. The researchers enhanced the standard CNN model by integrating a multi layered perceptron based spatial attention mechanism (MLP-ATCNN). This enhancement introduces a dual channel structure within the CNN to analyse plant leaf images more effectively.
The model processes two versions of a leaf image in parallel. One is the original image, and the other is a segmented version that focuses on non green regions, areas likely to show signs of disease. This segmentation is achieved through a technique known as k-means clustering, which groups similar colours together to isolate potentially infected areas. By using this dual input approach, the model is able to consider both the overall leaf structure and the fine grained details of disease symptoms simultaneously.
This mechanism improves the model’s ability to distinguish between healthy and diseased parts of the leaf and enhances the quality of the features learned during training. The total parameter count of the model is fewer than 200,000, with the attention module itself accounting for fewer than 5,000. This compact design enables the model to operate efficiently on edge devices (such as the Raspberry Pi).
Testing and Performance on Diverse Datasets
The researchers evaluated the performance of EdgePlantNet on two publicly available image datasets. The first, PlantVillage, consists of plant leaf images taken under controlled laboratory conditions with uniform backgrounds. The second, BPLD, is a more challenging dataset with images taken in real world agricultural settings, containing natural background elements and varying lighting conditions.
EdgePlantNet achieved high accuracy across both datasets. On PlantVillage, it reached 99.2% accuracy for potato leaf disease classification and 97.1% for tomato leaf disease. On the BPLD dataset, it achieved 95.72% accuracy for black gram leaves, demonstrating its robustness even in field conditions. The model also performed well in identifying rare or few shot diseases, those for which only a limited number of training samples were available. This is particularly important in real world settings where new or uncommon diseases may emerge.
Also read: Kuva Space Launches Hyperfield-1B Satellite to Strengthen Agri and Environmental Intelligence
Comparison with Existing AI Models
The study compared EdgePlantNet against several existing CNN models, including well known architectures such as ResNet and DenseNet. While these models are known for their accuracy, they require considerable computational resources and memory, making them unsuitable for deployment on edge devices. EdgePlantNet, in contrast, delivered comparable or superior accuracy while being much faster and requiring significantly less memory.
Other lightweight models were also evaluated, but they generally performed poorly on complex background images or lacked the ability to identify less common diseases. EdgePlantNet demonstrated the best balance among model size, speed and accuracy. The system processed images at approximately 3.65 frames per second on a Raspberry Pi and required only about 4 megabytes of memory, an order of magnitude less than many of the competing models.
Potential Applications and Future Work
The implementation of EdgePlantNet on low cost, portable hardware highlights its potential as a real world agricultural tool. Farmers could use the system directly in the field to monitor crop health, enabling timely interventions and reducing unnecessary pesticide use. This kind of targeted disease detection supports higher yields and promotes more sustainable farming practices.
The research team has identified avenues for future work, including enhancing the system to detect multiple diseases within a single image and improving its performance in extreme environmental conditions. Despite these areas for improvement, EdgePlantNet demonstrates the feasibility of deploying advanced AI tools on low resource platforms for practical agricultural applications.
The system offers a functional and deployable solution for field level disease monitoring and could contribute meaningfully to the broader adoption of smart farming practices in India and similar agrarian economies.
