ConvGeM-next: a deep learning framework for plant disease detection

Abstract

Introduction Plant diseases pose a major challenge to sustainable agriculture, particularly in regions that heavily depend on farming. Early and accurate identification of plant diseases is crucial for ensuring food production and minimizing crop losses. The rapid advancement of deep learning, particularly in convolutional neural networks (CNNs), has significantly enhanced plant disease classification performance. However, many models often struggle to generalize effectively in real-world scenarios due to challenges such as low-intensity visuals, low contrast between the background and foreground of the suspected sample, noise, and chrominance variation. Methods To address the challenges mentioned above, we introduce ConvGem-NeXt, an end-to-end deep learning architecture specifically designed for fine-grained plant disease classification, built on the ConvNeXt baseline model featuring enhanced generalization capabilities. More precisely, our method incorporates a learnable Generalized Mean pooling layer and ReLU activation in the ConvNeXt model to enhance spatial feature representation, and a custom classifier head that integrates batch normalization, ReLU activation, and dropout to mitigate overfitting and improve classification accuracy. Results We tested the presented model on two large-scale and diverse databases, PlantVillage and the PlantDoc. The model achieved 99.65% accuracy on the PlantVillage dataset and 94.69% accuracy on the real-world PlantDoc dataset, demonstrating the efficacy of our method for reliably classifying plant diseases. Discussion This work contributes to the rapidly growing field of agricultural automation by providing a reliable framework for timely disease diagnosis and supporting the enhancement of crop productivity.

Plant phenotyping relevance

植物病害を画像から分類する深層学習アーキテクチャを開発し、複数データセットで性能検証しており、植物の病害状態の取得・推定が研究の中心である。

Code and data availability

The paper's plant disease classification experiments were performed on two public image datasets, PlantVillage and PlantDoc, both explicitly linked in the data availability statement. No author code or model checkpoints are released.