Convolutional Neural Networks for the Automatic Identification of Plant Diseases

Justine Boulent, Samuel Foucher, Jérôme Théau, Pierre-Luc St-CharlesFrontiers in Plant Science 2019 · DOI

Abstract

Deep learning techniques, and in particular Convolutional Neural Networks (CNNs), have led to significant progress in image processing. Since 2016, many applications for the automatic identification of crop diseases have been developed. These applications could serve as a basis for the development of expertise assistance or automatic screening tools. Such tools could contribute to more sustainable agricultural practices and greater food production security. To assess the potential of these networks for such applications, we survey 19 studies that relied on CNNs to automatically identify crop diseases. We describe their profiles, their main implementation aspects and their performance. Our survey allows us to identify the major issues and shortcomings of works in this research area. We also provide guidelines to improve the use of CNNs in operational contexts as well as some directions for future research.

Introduction

Plant diseases and pests are a major threat to global food security, causing an estimated 20–40% loss in crop yields annually. This significant impact often leads to an over-reliance on chemical pesticides, which has detrimental environmental consequences. Automated identification systems offer a path toward more sustainable agriculture by enabling precise, localized treatments. However, transitioning from laboratory success to field-ready tools remains a significant challenge.

While a model can easily classify a single leaf against a plain background in a controlled setting, real-world agricultural environments are far more complex. Factors such as shifting light conditions, overlapping foliage, and "distractor" backgrounds like soil or insects create a "reality gap." This paper addresses the current state of the field, highlighting the need for standardization and more robust evaluation metrics that reflect the messy reality of the farm.

Methodology and Task Types

This study provides a systematic survey of 19 landmark research papers published between 2016 and 2018. The analysis categorizes these works based on several technical dimensions, including task type, training strategy, and model architecture.

A critical distinction in these studies is the level of granularity required for the identification task. Most research focuses on simple Classification, where a single label is assigned to an entire image. More advanced approaches include Object Detection, which identifies and locates specific lesions using bounding boxes, and Segmentation, which provides pixel-level masks of the diseased areas.

Comparison of computer vision tasks: Classification, Object Detection, and Segmentation for identifying Esca disease in grapevines.

The survey also evaluated training strategies, specifically comparing "training from scratch" against Transfer Learning. In transfer learning, models pre-trained on massive general datasets like ImageNet are fine-tuned for specific plant diseases, a method that often yields superior results with less data.

The "Reality Gap" and Data Quality

A key finding of this review is the "Performance Illusion." Many models achieving near-perfect accuracy (up to 99%) on laboratory datasets like PlantVillage see their performance plummet to as low as 31% when tested on independent, real-world images. This suggests that models often memorize specific laboratory conditions—such as a particular pot or lighting setup—rather than learning the actual symptoms of the disease. This is known as the "background trap."

To combat this, the authors emphasize that data quality and diversity are more important than model depth. Simply moving to a deeper architecture (e.g., from VGG to ResNet) provides marginal gains compared to diversifying the training set with different growth stages, varieties, and lighting conditions. Furthermore, the inclusion of "hard negatives"—images of healthy leaves, soil, or insects—is essential to prevent high false-positive rates in the field.

  • CNN Dominance: CNNs consistently outperformed traditional machine learning methods (such as SVMs or Random Forests) by margins of 3% to 29%.
  • Transfer Learning Efficacy: Fine-tuning pre-trained models generally resulted in the highest precision and significantly reduced the computational resources required for training.
  • Research Bias: There is a notable concentration of research on specific crops; 10 of the 19 studies focused on tomatoes, reflecting a strong interest in market gardening.
  • Task Distribution: 16 of the 19 studies were limited to simple classification, with only a small minority attempting the more operationally useful task of object detection.

Future Directions

To move beyond the "black box" nature of current algorithms, future research must prioritize:

  1. Independent Testing: Moving away from "held-out" validation sets toward truly independent test sets to reveal actual generalization capabilities.
  2. Early Detection: Exploring infrared or hyperspectral imaging to identify diseases before they become visible to the naked eye.
  3. Complex Scenarios: Developing multi-label models capable of identifying simultaneous diseases or nutritional deficiencies.
  4. Active Learning: Implementing strategies where models identify "confusing" images for expert annotation, making the creation of large-scale field datasets more efficient.

By addressing these shortcomings, the research community can transition CNN-based tools from academic curiosities into robust, operational assets for global agriculture.

Summary updated

Abstract

Deep learning techniques, and in particular Convolutional Neural Networks (CNNs), have led to significant progress in image processing. Since 2016, many applications for the automatic identification of crop diseases have been developed. These applications could serve as a basis for the development of expertise assistance or automatic screening tools. Such tools could contribute to more sustainable agricultural practices and greater food production security. To assess the potential of these networks for such applications, we survey 19 studies that relied on CNNs to automatically identify crop diseases. We describe their profiles, their main implementation aspects and their performance. Our survey allows us to identify the major issues and shortcomings of works in this research area. We also provide guidelines to improve the use of CNNs in operational contexts as well as some directions for future research.
BibTeX
@inproceedings{boulent2019convolutional,
  author = {Justine Boulent and Samuel Foucher and Jérôme Théau and Pierre-Luc St-Charles},
  title = {{Convolutional Neural Networks for the Automatic Identification of Plant Diseases}},
  booktitle = {Frontiers in Plant Science},
  year = {2019},
  doi = {10.3389/fpls.2019.00941},
}
Figure