Autonomous Embedded-Vision System for Multistage Detection of Phytopathogenic Fungi in Potato and Tomato Crops UsingConvolutional Neural Networks

Abstract

Abstract Current phytopathological diagnostic systems rely on manual inspections or laboratory analyses, which delay early detection and limit in-field responsiveness. Phytopathogenic fungal diseases pose a persistent threat to food security, directly affecting the productivity of essential crops such as potato ( Solanum tuberosum ) and tomato ( Solanum lycopersicum ) [1]–[5]. Among these diseases, Phytophthora infestans , the causal agent of late blight, is characterized by its high virulence and rapid spread, capable of generating significant losses in short periods when detection occurs too late [3], [4]. To address this issue, a computer vision and deep learning–based system for multistage detection of fungal infections in potato and tomato crops is proposed. The system comprises a convolutional neural network optimized for edge processing and a mobile robotic platform equipped with a manipulator arm for localized treatment application. The developed model was deployed on a Raspberry Pi 4 connected to a 12-MP Raspberry Pi Camera Module 3 NoIR, responsible for acquiring RGB images in the field. The proposed network was compared with reference architectures—ResNet-50, VGG16, MobileNetV2, and Inception-v3—within a four-stage detection pipeline: crop identification, health-state classification, infection diagnosis, and foliar severity estimation. A dataset of 18,200 images obtained from publicly accessible online sources, under diverse lighting and background conditions, was used, partitioned into 70% for training, 20% for validation, and 10% for testing. Preliminary results show an average accuracy in the range of 0.90–0.92, with inference latencies below 60 ms per image, ensuring smooth performance on the Raspberry Pi 4 without requiring cloud connectivity. Additionally, the network demonstrated higher sensitivity to visual variations compared to the baseline models.

Code and data availability

The paper's CNN training data are two publicly available third-party plant-disease image datasets (Mendeley Data and Kaggle Plant Village) with explicit URLs in the Data Availability Statement. The authors' field-test images and experimental records are only available on request, and no analysis code or trained model/с