Generalising Across Different Crop Diseases with Deep Learning Model

Abstract

Global agriculture faces severe economic threats from plant diseases, necessitating automated diagnostic systems. However, standard deep learning models feature over-parameterized architectures that require prohibitive computational resources, limiting field deployment. This paper presents an efficient, lightweight sequential Convolutional Neural Network (CNN) optimized for rapid multi-class plant disease classification using the Kaggle New Plant Diseases Dataset. The proposed architecture streamlines feature extraction by applying a single max-pooling layer after every two 32-filter convolutional layers. To prevent overfitting, a dual dropout strategy (0.25 and 0.5) is integrated alongside a dense layer of 1,500 units. Trained over 10 epochs using the Adam optimizer (learning rate = 0.001) and varying batch sizes (32 to 512), the model demonstrated swift convergence, achieving a peak training accuracy of 98.15% and a validation accuracy of 95.87%. Evaluated across a test support of 70,295 images spanning 38 distinct crop classes, the network delivered an overall macro and weighted average of 1.00 for precision, recall, and F1-score. These results prove that an optimized, computationally economic sequential framework can match complex architectures, providing a viable solution for real-time edge deployment in precision agriculture.

Plant phenotyping relevance

植物画像から病害状態を分類する軽量CNNを開発・評価しており、病害表現型の取得・推定手法が研究の中心です。

Code and data availability

The paper trains a sequential CNN on the public Kaggle New Plant Diseases Dataset, but the article contains no author code, trained model, or supplement availability statement, and no paper-specific repository or deposit URL is given. The only public asset (the Kaggle dataset) has no URL matching the allowed list, so a

No evidence-backed public reproduction asset is currently recorded.