Source-linked AI summary
Identification and Recognition of Rice Diseases and Pests Using Convolutional Neural Networks
Chowdhury Rafeed Rahman, Preetom Saha Arko, Mohammed Eunus Ali, Mohammad Ashik Iqbal Khan, Sajid Hasan Apon, Farzana Nowrin, Abu Wasif
TL;DR
Rice disease and pest detection is difficult despite its importance for food security and yield, especially when mobile deployment must work offline in remote areas. The paper evaluates and fine-tunes CNNs, then proposes two-stage training for a lightweight model; fine-tuning performs best, while Simple CNN achieves high accuracy comparable to larger architectures.
Problem
Timely rice disease and pest detection is needed because diseases cause 10-15% average yield loss in Bangladesh, while remote farmers may lack reliable internet for online tools.
Method
The paper tests VGG16, InceptionV3, MobileNetv2, NasNet Mobile, and SqueezeNet using multiple training strategies, and proposes a lightweight CNN with two-stage training for mobile detection.
Results
Fine-tuning from pretrained ImageNet weights performs best across the tested architectures, while Simple CNN achieves high accuracy comparable to other state-of-the-art CNNs despite using few parameters.
Takeaways & Limitations
The proposed lightweight CNN is presented as an effective tool for rice disease and pest detection on mobile devices in remote environments.
Takeaways & Limitations
Two-stage training requires manually dividing the dataset into symptom classes, making large datasets labor intensive and vulnerable to missed or misinterpreted symptom variations.
Abstract
from arXiv · showhide
An accurate and timely detection of diseases and pests in rice plants can help farmers in applying timely treatment on the plants and thereby can reduce the economic losses substantially. Recent developments in deep learning based convolutional neural networks (CNN) have greatly improved the image classification accuracy. Being motivated by the success of CNNs in image classification, deep learning based approaches have been developed in this paper for detecting diseases and pests from rice plant images. The contribution of this paper is two fold: (i) State-of-the-art large scale architectures such as VGG16 and InceptionV3 have been adopted and fine tuned for detecting and recognizing rice diseases and pests. Experimental results show the effectiveness of these models with real datasets. (ii) Since large scale architectures are not suitable for mobile devices, a two-stage small CNN architecture has been proposed, and compared with the state-of-the-art memory efficient CNN architectures such as MobileNet, NasNet Mobile and SqueezeNet. Experimental results show that the proposed architecture can achieve the desired accuracy of 93.3\% with a significantly reduced model size (e.g., 99\% less size compared to that of VGG16).
1 Introduction
The paper targets timely rice disease and pest recognition for practical agricultural use, addressing both recognition accuracy and deployment constraints on mobile devices. It evaluates large CNNs and proposes a lightweight alternative using two-stage training.
- 10–15% average yield loss is associated with 10 major rice diseases in Bangladesh, motivating timely automated detection on readily available mobile devices.
- Prior plant-disease studies emphasized accurate recognition with CNNs but largely overlooked training-method modification and the parameter burden of high-performing models for mobile deployment.
- VGG16 and InceptionV3 were evaluated using fine tuning, transfer learning, and training from scratch; fine tuning performed best in both architectures.
- The paper proposes two-stage training and a small Simple CNN that achieves high accuracy despite having few parameters, outperforming MobileNet, NasNet Mobile, and SqueezeNet.
- The real-life dataset collection accounts for symptom variation from weather, soil, attack stage, plant part, and heterogeneous image backgrounds.
- The stated contributions combine testing large CNN frameworks on real-world rice images with a lightweight two-stage CNN intended for mobile-based detection in remote environments.
2 Materials and Methods
The study builds a heterogeneous rice-image dataset and evaluates large and memory-efficient CNN architectures under several training strategies. It also develops a small Simple CNN and uses augmentation and cross-validation to assess performance.
- 2.1 Data Collection: 1,426 rice disease and pest images were collected from BRRI paddy fields over seven months across seasons, weather conditions, heterogeneous backgrounds, and four camera types.
- 2.1 Data Collection: Nine classes comprise five diseases, three pests, and one healthy-plant class, with images covering symptoms on leaves, stems, and grains.
- 2.1 Data Collection: The dataset includes intra-class symptom variation, including distinct early- and late-stage appearances for Hispa and Brown Plant Hopper pests.
- 2.2 Experimental Setup: VGG16, InceptionV3, MobileNetv2, NasNet Mobile, and SqueezeNet were compared with the proposed Simple CNN across multiple training methods.
- 2.2 Experimental Setup: Fine tuning uses pretrained ImageNet convolutional weights while retraining dense layers, whereas transfer learning freezes convolutional weights and trains only dense layers.
- 2.2 Experimental Setup: Performance was measured with 10-fold cross-validation accuracy and standard deviation using categorical crossentropy for multiclass classification.
- 2.2 Experimental Setup: Each original image generated 10 augmented images using geometric and intensity transformations, including rotations, flips, distortion, shearing, skewing, and intensity changes.
- 2.2 Experimental Setup: The Simple CNN contains 0.8 million parameters compared with 138 million in VGG16, while two-stage training is presented as supporting high accuracy with a small model.
3 Results and Discussion
Across the evaluated CNNs, fine tuning produced the strongest and most precise results, while Simple CNN combined low parameter count with high accuracy. Remaining analysis examined misclassification patterns and learned feature representations, alongside a manual data-partitioning limitation.
- Training comparison: Fine tuning gave the best accuracy for all evaluated architectures, whereas transfer learning performed worst and training from scratch remained unsatisfactory.For the three small models, baseline accuracy was below 80%, and transfer learning fell below 50% for SqueezeNet.
- Model efficiency: Simple CNN achieved high accuracy with far fewer parameters than VGG16, supporting two-stage training as an effective memory-efficient approach.Simple CNN used 0.8 million parameters compared with VGG16’s 138 million parameters.
- Limitations: Two-stage training requires manually dividing the dataset into symptom classes, which is labor intensive and may miss or split intra-class variations.The paper proposes high-dimensional clustering as a possible way to automate identification of intra-class variations.
- Error analysis: The confusion matrix showed that 4.3% of False Smut images were misclassified, the highest error among the reported classes.The paper associates this error with symptoms occupying a small image region against heterogeneous backgrounds.
- Feature representations: Simple CNN’s first convolution layer retained regional input features, while its last layer produced less visually interpretable representations related to image class.The first layer output was 222 × 222 × 16, while the last convolution layer output was 10×10×64.
4 Conclusion
The paper contributes a real-life rice disease and pest dataset, evaluates CNN training strategies, and introduces two-stage training for a small Simple CNN. It also identifies future extensions for more comprehensive detection in heterogeneous backgrounds.
- The dataset contains 1426 rice disease and pest images collected in real-life scenarios across eight classes.
- Three training methods were evaluated across large and small CNN architectures targeted toward mobile applications.
- Fine tuning from pre-trained ImageNet weights produced the best result for all five evaluated architectures.
- Two-stage training enabled the proposed Simple CNN architecture to perform well in real-life scenarios.
- Future work could combine image data with location, weather, and soil information and apply segmentation or object detection for heterogeneous backgrounds.