CNNAttLSTM: an attention-enhanced CNN-LSTM architecture for high-precision jackfruit leaf disease classification.

Abstract

Introduction Jackfruit cultivation is highly affected by leaf diseases that reduce yield, fruit quality, and farmer income. Early diagnosis remains challenging due to the limitations of manual inspection and the lack of automated and scalable disease detection systems. Existing deep-learning approaches often suffer from limited generalization and high computational cost, restricting real-time field deployment. Methods This study proposes CNNAttLSTM, a hybrid deep-learning architecture integrating Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) units, and an attention mechanism for multi-class classification of algal leaf spot, black spot, and healthy jackfruit leaves. Each image is divided into ordered 56×56 spatial patches, treated as pseudo-temporal sequences to enable the LSTM to capture contextual dependencies across different leaf regions. Spatial features are extracted via Conv2D, MaxPooling, and GlobalAveragePooling layers; temporal modeling is performed by LSTM units; and an attention mechanism assigns adaptive weights to emphasize disease-relevant regions. Experiments were conducted on a publicly available Kaggle dataset comprising 38,019 images, using predefined training, validation, and testing splits. Results The proposed CNNAttLSTM model achieved 99% classification accuracy, outperforming the baseline CNN (86%) and CNN-LSTM (98%) models. It required only 3.7 million parameters, trained in 45 minutes on an NVIDIA Tesla T4 GPU, and achieved an inference time of 22 milliseconds per image, demonstrating high computational efficiency. The patch-based pseudo-temporal approach improved spatial-temporal feature representation, enabling the model to distinguish subtle differences between visually similar disease classes. Discussion Results show that combining spatial feature extraction with temporal modeling and attention significantly enhances robustness and classification performance in plant disease detection. The lightweight design enables real-time and edge-device deployment, addressing a major limitation of existing deep-learning techniques. The findings highlight the potential of CNNAttLSTM for scalable, efficient, and accurate agricultural disease monitoring and broader precision agriculture applications.

Plant phenotyping relevance

植物葉の病害状態を画像から分類する深層学習手法を提案・比較評価しており、病害フェノタイピングの方法開発が中心である。

Code and data availability

The paper's core phenotyping input is a publicly available Kaggle jackfruit leaf disease image dataset (38,019 images) explicitly linked in the data availability statement; no author code or model checkpoints are deposited.