A Novel Lightweight AlexNet Convolutional Neural Network for Tomato Leaf Disease Classification

Abstract

Tomato leaf diseases must be identified early and accurately in order to reduce output loss and advance sustainable agriculture. Deep learning models have shown encouraging results in the identification of plant diseases, but their high processing requirements and inability to adjust to field-specific limitations sometimes make it difficult to implement them in real-world applications. For the purpose of accurately and efficiently classifying tomato leaf diseases, we present a convolutional neural network based on AlexNet that is lightweight and field-aware. Our proposed model is designed with less complexity than traditional architectures and is able to perform inference faster without sacrificing accuracy, making it suitable for real-time implementation in low resource agricultural environments. The algorithm was trained and tested on a dataset of 7704 augmented images of tomato leaves from 7 different disease categories. The Lightweight AlexNet achieved accuracy comparable to VGG variants and outperformed deeper models such as ResNet (96.81%) with lower parameter overhead. It was trained from scratch using TensorFlow & Keras on 100×100 pixel inputs, achieving a training accuracy of 99.15% and a validation accuracy of 99.74%. Moreover, an effective structure of the model enables the installation on edge devices, which offers a scalable precision farming solution. Our work helps to bridge the gap between deep learning research and real-world application in agriculture, allowing the development of real-field, resource-efficient disease detection systems.

Plant phenotyping relevance

トマト葉の病害状態を画像から分類する軽量CNNを開発し、精度・計算量・エッジ実装性を評価しており、植物フェノタイピング手法が研究の中心である。

Code and data availability

The paper's field-collected tomato leaf image dataset (7,704 augmented images from Golaghat, Assam) is paper-specific but only available from the corresponding author upon reasonable request. No public code, model checkpoints, or dataset URLs are provided.

No evidence-backed public reproduction asset is currently recorded.