What type of study is this?

This is a Quantitative Study study (also classified as: Experimental Study).

October 17, 2025

Research on Potato Leaf Disease Recognition Method Based on Convolutional Neural Networks

Key Points

The CNN-based model achieved an accuracy of 99.83% for detecting potato leaf diseases.
A Mixup–CutMix augmentation technique improved the baseline accuracy by 3.3 percentage points.
Training involved a novel ant colony optimization method to enhance model performance and generalization.
Results validated on a hold-out test set showed significant improvements in both accuracy and F1-score.

Abstract

ABSTRACT The early detection of foliar diseases such as early blight and late blight is crucial for securing potato yield and quality, yet traditional manual diagnosis remains labor‐intensive, inconsistent, and unsuitable for large‐scale agricultural settings. To address this, the present study investigates a CNN‐based automatic recognition framework for potato leaf diseases using RGB imagery, with a focus on enhancing both model generalization and deployment feasibility. Rather than proposing a novel network architecture, this work systematically benchmarks three representative backbones—AlexNet, MobileNetV3‐Large, and ResNet152—under a unified training pipeline. Two methodological innovations are introduced: (1) a composite Mixup–CutMix data augmentation strategy tailored for small‐scale agricultural datasets, and (2) an ant colony optimization (ACO)‐driven training strategy that jointly tunes transfer learning, L2 regularization, and Dropout. The dataset, expanded from 3000 to 18,000 images through augmentation, supports robust training under conditions of occlusion, lighting variation, and background clutter. Experimental results show that the Mixup–CutMix augmentation improves baseline accuracy by 3.3 percentage points, while the ACO‐optimized training pipeline contributes an additional 1.2 points. ResNet152 achieved the highest performance with 99.83% accuracy and an F1‐score of 0.9982 on a hold‐out test set. Grad‐CAM visualizations confirm biologically meaningful attention to lesion areas. The best‐performing model has been deployed in a lightweight PyQt5‐based GUI, and a compressed variant runs successfully in real time on Raspberry Pi 4B. This study contributes a robust augmentation‐training pipeline, an interpretable and deployable diagnostic system, and a demonstration of feasibility on edge hardware. Future work will extend the model to additional pathogens, quantify on‐device latency and FPS, and integrate multi‐modal cues for broader crop health monitoring.

اسأل الذكاء الاصطناعي

Bookmark