[Paper Review] A two-step machine learning approach for crop disease detection: an application of GAN and UAV technology

Aaditya Prasad, Nikhil Mehta+2 more|arXiv (Cornell University)|Sep 19, 2021

Smart Agriculture and AI14 references4 citations

TL;DR

This paper proposes a two-step machine learning framework that leverages GAN-generated high-fidelity images and a low-fidelity data generator to train a UAV-based crop disease detection system. It achieves 96.3% accuracy on high-fidelity classification and 75.5% confidence in low-fidelity identification, enabling efficient, scalable, and accurate disease detection using minimal labeled data.

ABSTRACT

Automated plant diagnosis is a technology that promises large increases in cost-efficiency for agriculture. However, multiple problems reduce the effectiveness of drones, including the inverse relationship between resolution and speed and the lack of adequate labeled training data. This paper presents a two-step machine learning approach that analyzes low-fidelity and high-fidelity images in sequence, preserving efficiency as well as accuracy. Two data-generators are also used to minimize class imbalance in the high-fidelity dataset and to produce low-fidelity data that is representative of UAV images. The analysis of applications and methods is conducted on a database of high-fidelity apple tree images which are corrupted with class imbalance. The application begins by generating high-fidelity data using generative networks and then uses this novel data alongside the original high-fidelity data to produce low-fidelity images. A machine-learning identifier identifies plants and labels them as potentially diseased or not. A machine learning classifier is then given the potentially diseased plant images and returns actual diagnoses for these plants. The results show an accuracy of 96.3% for the high-fidelity system and a 75.5% confidence level for our low-fidelity system. Our drone technology shows promising results in accuracy when compared to labor-based methods of diagnosis.

Motivation & Objective

• To address the challenge of limited labeled UAV imagery for crop disease detection.
• To improve detection accuracy and efficiency in large-scale agricultural fields using low- and high-fidelity image analysis.
• To reduce dependency on manual labeling by generating synthetic, representative training data at both resolution levels.
• To enable real-time, on-the-fly disease detection using UAVs with minimal data and computational overhead.
• To validate a scalable, data-efficient pipeline that outperforms traditional labor-intensive diagnosis.

Proposed method

• A GAN-based data generator creates synthetic high-fidelity images to augment the original dataset and address class imbalance.
• A low-fidelity data generator synthesizes realistic low-resolution UAV images from high-fidelity outputs to train the identifier model.
• A two-stage pipeline is implemented: first, a low-resolution identifier detects potentially diseased plants in real time; second, a high-resolution classifier diagnoses the disease.
• The identifier uses object detection to localize suspicious plants in low-fidelity UAV images, while the classifier performs multi-class disease classification on close-up images.
• The framework is trained and validated using a single small public dataset of high-fidelity apple tree images, with synthetic data generation enabling robust model training.
• Model performance is evaluated using confusion matrices, precision, recall, F1-score, and confidence intervals to assess both stages of the pipeline.

Experimental results

Research questions

• RQ1 Can a two-step machine learning pipeline using synthetic data generation improve the accuracy of UAV-based crop disease detection?
• RQ2 To what extent can GANs and a low-fidelity data generator reduce the need for large-scale, manually labeled UAV image datasets?
• RQ3 How does the performance of the identifier model on low-fidelity images affect the overall accuracy of the disease detection pipeline?
• RQ4 Can a data-efficient framework achieve competitive performance compared to state-of-the-art models using only a small initial labeled dataset?
• RQ5 What are the key bottlenecks in the current system, and how can they be addressed to improve real-world deployment?

Key findings

• The high-fidelity classifier achieved a precision of 98% on the F1-score and 96.3% accuracy, demonstrating strong performance on disease classification.
• The low-fidelity identifier achieved a confidence level of 75.5% for detecting potentially diseased plants, with a theoretical upper bound on pipeline accuracy matching this value.
• Classifier performance metrics showed high precision and recall across all classes: 0.96–0.98 for precision, 0.98–1.00 for recall, and F1-scores above 0.96.
• The confusion matrix revealed minimal misclassification, with only 1–2 false positives per class in most cases, indicating strong model generalization.
• The full pipeline accuracy was estimated at 71.8%, with the identifier being the primary bottleneck due to the complexity of object detection in low-resolution images.
• The framework demonstrated that synthetic data generation significantly reduces reliance on costly manual labeling while maintaining high model accuracy.