Source-linked AI summary
Plant Diseases recognition on images using Convolutional Neural Networks: A Systematic Review
Andre S. Abade, Paulo Afonso Ferreira, Flavio de Barros Vidal
TL;DR
Plant diseases reduce crop productivity, and accurate large-scale symptom identification remains a challenge for agricultural monitoring and decision-making. This paper systematically reviews CNN-based plant-disease recognition, examining 121 studies to characterize methods, datasets, crops, pathogens, trends, and gaps. The review reports advances from traditional architectures combined with optimization and customization, while identifying dataset and pathogen coverage limitations requiring further attention.
Problem
Plant diseases can reduce crop productivity, while agricultural monitoring requires rapid and accurate diagnosis and automated solutions for large plantation areas.
Method
The paper conducts a systematic literature review of CNN-based plant-disease identification, synthesizing 121 studies across methods, datasets, crops, and pathogens.
Results
The review finds significant advances in CNN-based plant-disease prediction, with traditional architectures combined with optimization and customization presenting relevant accuracy in complex real-environment image datasets.
Takeaways & Limitations
The review identifies innovative CNN trends and research gaps, including opportunities involving multispectral and hyperspectral images for agricultural applications.
Takeaways & Limitations
Limited image datasets remain an important limitation, and few CNN approaches classify nematode-caused diseases because their symptoms resemble other diseases or are not visibly expressed in leaves and stems.
Abstract
from arXiv · showhide
Plant diseases are considered one of the main factors influencing food production and minimize losses in production, and it is essential that crop diseases have fast detection and recognition. The recent expansion of deep learning methods has found its application in plant disease detection, offering a robust tool with highly accurate results. In this context, this work presents a systematic review of the literature that aims to identify the state of the art of the use of convolutional neural networks(CNN) in the process of identification and classification of plant diseases, delimiting trends, and indicating gaps. In this sense, we present 121 papers selected in the last ten years with different approaches to treat aspects related to disease detection, characteristics of the data set, the crops and pathogens investigated. From the results of the systematic review, it is possible to understand the innovative trends regarding the use of CNNs in the identification of plant diseases and to identify the gaps that need the attention of the research community.
1 Introduction
Plant diseases threaten crop productivity, while accurate symptom identification remains difficult at scale. This review characterizes CNN-based plant-disease recognition and the challenges shaping its development.
- Motivation: Plant diseases can substantially reduce crop productivity, making continuous monitoring and rapid, accurate diagnosis important for disease management.The passages connect disease control with minimizing production losses and maintaining crop sustainability.
- Problem: Correctly identifying symptoms of major crop diseases is a major agricultural challenge.Traditional manual and mechanized practices are described as unable to cover large plantation areas or provide sufficient early information.
- Approach: Image-based computer vision and CNN systems have become advanced approaches for recognizing and classifying plant diseases.The review situates CNNs within broader computer-vision and artificial-intelligence advances.
- Challenges: Popular CNN architectures, including LeNet, AlexNet, VGGNet, GoogLeNet, InceptionV3, ResNet, and DenseNet, have increased disease-identification accuracy.The reported challenges include crop diversity, pest and disease variety, dataset characteristics, architecture choices, and optimization complexity.
- Review objective: The systematic literature review characterizes the state of the art in CNN-based plant-disease recognition and identification.It examines contributions addressing identified challenges and innovations intended to improve CNN performance.
- Organization: The paper presents the background, review protocol, SLR results, discussion, and conclusion in successive sections.This passage describes the paper’s organization rather than its substantive findings.
2 Background
The background frames plant-disease recognition as a difficult, error-prone task and introduces CNN strategies used to address it. These strategies include transfer learning, custom layers, data augmentation, and architectural innovation.
- Plant disease detection: Manual visual disease identification is subjective and can be affected by psychological and cognitive phenomena, optical illusions, and error.The paper also mentions serological, molecular, and microscopic identification methods.
- Types of plant diseases: Plant diseases are classified by causative agent as biotic, originating from living organisms, or abiotic, produced by non-living ecological circumstances.Listed biotic causes include fungi, bacteria, viruses, pests, and nematodes.
- Plant disease detection system: CNN-based image and machine-learning systems are presented as versatile tools for assimilating heterogeneous data and producing predictions of complex and uncertain phenomena.The paper associates these automated solutions with improved accuracy in plant-disease diagnosis.
- CNN architectures: New CNN architectures include innovations such as parameter optimization, regularization, structural reformulation, and newly designed blocks.The review categorizes architecture proposals as new architectures when they report such innovations.
- Transfer learning: Transfer learning reuses a model from a specific task as the starting point for a related target-domain model.Its objective is to borrow labeled or extracted knowledge from related fields to improve performance in the area of interest.
- Transfer learning: The two described transfer-learning approaches are using a base network as a fixed feature extractor and fine-tuning the base network.Fixed extraction trains a new classifier from generated features; fine-tuning also modifies earlier-layer weights through backpropagation.
- CNN layers: Traditional CNNs generally use input, intermediate, and output layers, with frameworks providing standardized layers for established models.The passage presents this as the typical construction of traditional CNN architectures.
- Custom layers: Custom layers are needed when problem-specific peculiarities require operation at a lower abstraction level than standard framework layers.For example, a ResNet application may add an extra layer to normalize classifier output.
3 Planning Review
The review used a planned, auditable systematic-review process to examine how CNNs address plant-disease identification challenges. It searched five scientific databases, screened studies with predefined quality criteria, and retained 121 studies from 2,221 retrieved records.
- The review asked how CNNs innovate and overcome challenges in plant-disease identification, supported by six secondary questions on architectures, datasets, crops, frameworks, algorithms, and diseases.
- The protocol excluded gray literature and assumed relevant findings would generally be described or referenced in published scientific papers.
- The search covered five indexed databases and publications from 2010 to 2019 using a broad string combining plant-disease and machine-learning terms.
- The quality assessment scored five criteria as yes = 1, partially = 0.5, or not = 0, excluding studies with a quality index below 3.
- PRISMA-guided identification, screening, extraction, and selection used duplicate removal, independent reviewers, and full-text assessment; screening agreement exceeded 80% Kappa.
- 2,221 studies were retrieved, 91 duplicates and 682 nonqualifying reviews were removed, and 121 studies remained for synthesis.
4 Results
The 121-study synthesis mapped CNN approaches across architectures, datasets, crops, frameworks, algorithms, and pathogen groups. It found broad use of customized or benchmark datasets, a predominance of controlled-environment images, and continued interest in new architectures.
- SQ1: Seven practical approach groups were identified: customizing layers, transfer learning, new architecture, fine-tuning and hyperparameters, image segmentation, data augmentation, and unsupervised learning.
- SQ1: 25 of 121 studies proposed a new architecture, indicating that new CNN architectures were an active research trend.
- SQ2: Customized datasets appeared in 66 studies, while PlantVillage appeared in 45 studies in full or partial form.
- SQ3: Tomato was the most investigated single crop, with 18 studies, while diverse-crop approaches were especially associated with PlantVillage.
- SQ4: Six main frameworks were identified, with TensorFlow the most used framework in 38 of the 121 studies.
- SQ5: Thirty CNN algorithms were counted, with AlexNet and VGG emphasized among the predominant classic architectures; 19 studies proposed new architectures and 15 customized classic algorithms.
- SQ6: The synthesis analyzed 54 datasets spanning 142 crop types and grouped causative pathogens into virus, bacteria, fungi, algae, plague, nematodes, and abiotic categories.
5 Discussion
The review finds diverse CNN innovations for plant-disease recognition, alongside persistent dataset, crop, and pathogen coverage gaps. Controlled-environment data remain common, while real-world conditions and underrepresented diseases constrain generalization.
- 5.1 Predominant Approaches: CNN studies explored structural reformulations, architecture blending, and purpose-built models for plant-disease recognition.Reported examples include modified LeNet, AlexNet, and ResNet; YOLO-DenseNet and Inception-ResNet combinations; and dedicated hyperspectral architectures.
- 5.1 Predominant Approaches: 96.25% accuracy was reported for early Tomato spotted wilt virus detection before visible symptoms using hyperspectral images and OR-AC-GAN.The approach integrates plant segmentation, spectrum classification, and image classification.
- 5.2 Characteristic of the Data Sets: 65.28% of reviewed approaches used controlled-environment image sets, with PlantVillage appearing in 37.19% of summarized studies.Some studies combined controlled-environment and real-growing-condition images to reduce dataset limitations.
- 5.2 Characteristic of the Data Sets: Models trained exclusively on controlled-environment images showed reduced accuracy on real-scene images with challenging illumination, backgrounds, resolutions, poses, and orientations.One reported case found an approximately 30%–40% reduction in global accuracy under real crop conditions.
- 5.3 CNN Algorithms and Frameworks: About 85% of approaches used transfer learning, fine-tuning, or hyperparameters, while traditional CNN architectures remained predominant.TensorFlow was the most-used framework, followed by Keras.
- 5.4 Types Crops and Disease Causing Pathogens: Tomatoes, apples, corn, rice, cucumbers, wheat, grapes, potatoes, and bananas concentrated the most investigations, while studies primarily addressing grains and cereals were lacking.Only studies S7, S81, and S84 were identified for soybeans.
- 5.4 Types Crops and Disease Causing Pathogens: Fungal diseases dominated the literature, whereas virus and bacterial studies lacked consistent quantitative coverage and nematode-focused approaches were scarce.The review also identified studies of algal and abiotic-factor diseases.
6 Conclusion
The review finds substantial progress in CNN-based plant disease detection, including accurate traditional architectures and growing architectural innovation. It also identifies dataset representativeness and neglected crops and pathogens as continuing gaps.
- CNN-based systems reconcile phytopathology expertise with automated extraction of symptomatic features for plant disease identification and classification.
- Dataset diversity and insufficient labeled samples remain bottlenecks because real-world scenarios complicate semantic cataloging into representative training sets.
- Traditional CNN architectures combined with optimization and customization methods achieve relevant accuracy despite the complexity of images captured in real crop environments.
- New CNN architectures are increasingly proposed for plant disease identification, although approaches using traditional architectures can still achieve higher accuracy.
- Grains and cereals are often overlooked, while relatively few methods identify or classify diseases caused by nematodes.
- Combining CNN architectures with multispectral and hyperspectral images is presented as a direction for agricultural solutions that may minimize damage from biotic or abiotic agents.