Identification of Apple Leaf Diseases by Improved Deep Convolutional Neural Networks With an Attention Mechanism

Peng Wang, Tong Niu, Yanru Mao, Zhao Zhang, Bin Liu, Dongjian HeFrontiers in Plant Science 2021 · DOI

This research, published in Frontiers in Plant Science, introduces a specialized deep learning architecture designed to identify apple leaf diseases in complex field environments. By combining the efficiency of EfficientNet with a novel "Coordinate Attention" mechanism, the authors achieve state-of-the-art accuracy while maintaining a model size suitable for potential mobile deployment.

The Problem: Environmental Complexity and Fine-Grained Similarity

Apple production is a cornerstone of the agricultural industry, but leaf diseases can devastatingly impact yield and quality. Traditional identification relies on manual inspection, which is slow and subjective. While standard Convolutional Neural Networks (CNNs) have improved automated diagnosis, they face hurdles in actual orchards due to varying lighting, messy backgrounds, and the fact that different diseases (or pests like Apple leaf mites and Apple litura moths) can look remarkably similar.

Example of apple leaf disease symptoms in a field environment.

Methods: The CA-ENet Architecture

The researchers developed CA-ENet (Coordination Attention EfficientNet), built on three pillars:

  • EfficientNet-B4 Backbone: Chosen via Neural Architecture Search (NAS) for its optimized balance of network depth, width, and image resolution.
  • Coordinate Attention Block (CAB): This core innovation decomposes global pooling into two 1D feature encodings (horizontal and vertical). This allows the model to capture long-range dependencies while preserving precise positional information—learning both what the disease looks like and where it is on the leaf.
  • Efficiency Optimizations: The use of depth-wise separable convolutions and h-swish activation functions reduces parameter counts and facilitates deployment on mobile hardware.

Datasets and Robustness

The study utilized a massive dataset called ALDID (Apple Leaf Disease Identification Dataset), containing 81,700 images. These were sourced from field environments at Northwest A&F University and the public PlantVillage dataset. To ensure robustness, the authors used image augmentation (blurring, brightness adjustments) and created a Model Robustness Test Dataset (MRTD) with added noise to simulate the "messy" data typical of real-world orchard photography.

Key Insights and Results

A major finding of the paper is that spatial awareness is critical for agricultural pathology. By embedding coordinate information into the attention mechanism, the model avoids "attention drift"—a common problem where a model looks at the background or healthy tissue instead of the lesion.

Localized lesions on an apple leaf, requiring precise spatial attention for identification.

CA-ENet achieved a 98.92% accuracy and an F1-score of 0.988, significantly outperforming several heavier models:

  • ResNet-152: 93.75% accuracy
  • DenseNet-264: 94.90% accuracy
  • ResNeXt-101: 95.67% accuracy

Conclusion

While some fine-grained confusion remains between highly similar symptoms, CA-ENet proves that smarter architectural choices, like coordinate attention, are more effective than simply increasing network depth. The authors plan to integrate this model into ground-based mobile inspection platforms to provide real-time, automated orchard monitoring.

Summary updated

Abstract

The accurate identification of apple leaf diseases is of great significance for controlling the spread of diseases and ensuring the healthy and stable development of the apple industry. In order to improve detection accuracy and efficiency, a deep learning model, which is called the Coordination Attention EfficientNet (CA-ENet), is proposed to identify different apple diseases. First, a coordinate attention block is integrated into the EfficientNet-B4 network, which embedded the spatial location information of the feature by channel attention to ensure that the model can learn both the channel and spatial location information of important features. Then, a depth-wise separable convolution is applied to the convolution module to reduce the number of parameters, and the h-swish activation function is introduced to achieve the fast and easy to quantify the process. Afterward, 5,170 images are collected in the field environment at the apple planting base of the Northwest A&F University, while 3,000 images are acquired from the PlantVillage public data set. Also, image augmentation techniques are used to generate an Apple Leaf Disease Identification Data set (ALDID), which contains 81,700 images. The experimental results show that the accuracy of the CA-ENet is 98.92% on the ALDID, and the average F1-score reaches .988, which is better than those of common models such as the ResNet-152, DenseNet-264, and ResNeXt-101. The generated test dataset is used to test the anti-interference ability of the model. The results show that the proposed method can achieve competitive performance on the apple disease identification task.
BibTeX
@inproceedings{wang2021identification,
  author = {Peng Wang and Tong Niu and Yanru Mao and Zhao Zhang and Bin Liu and Dongjian He},
  title = {{Identification of Apple Leaf Diseases by Improved Deep Convolutional Neural Networks With an Attention Mechanism}},
  booktitle = {Frontiers in Plant Science},
  year = {2021},
  doi = {10.3389/fpls.2021.723294},
}
Figure