Convolutional neural network-based detection and severity classification of rice blast disease from leaf images

Introduction

A trained plant pathologist can identify rice blast symptoms in about ten seconds per leaf but a single rice paddy contains millions of leaves, and most farms in sub-Saharan Africa don't have a pathologist within 50 kilometres. This gap between what is technically knowable and what is practically detectable costs rice farmers an estimated 10-30% of global production annually [1] . Magnaporthe oryzae attacks leaves, nodes, and panicles, producing characteristic diamondshaped lesions that expand rapidly under humid conditions [2] . Computer vision offers a way to bridge this gap. CNNs trained on labelled leaf images can classify diseases with accuracy approaching that of human experts [3] . Transfer learning where a network pre-trained on millions of general images is fine-tuned on crop-specific data reduces the amount of training data needed and speeds convergence [4] . Several CNN architectures have been tested for rice diseases, but most used controlled laboratory images rather than fieldcaptured photos with variable lighting, background, and camera angles [5] . This research trained a VGG-16-based CNN on 4,218 field-captured images from Ugandan rice paddies, comparing its performance against SVM and random forest classifiers for both binary detection (healthy vs infected) and five-class severity grading [6,7] .

Research area description

Images were collected from rice fields at the National Crops Resources Research Institute (NaCRRI), Namulonge, Uganda (0.53°N, 32.62°E, 1,140 m), during the first and second rainy seasons of 2021. The station grows both upland and lowland rice varieties under conditions highly favourable for blast development: mean temperature 21-27 °C, relative humidity 65-90%, and bimodal rainfall totalling 1,270 mm per year [8] .

Disease environment characterization

Blast pressure at the collection site was monitored using trap rows of the susceptible variety Komboka planted at field borders. Lesion counts on trap rows averaged 14.8 lesions per leaf by 45 days after transplanting in season one and 18.3 in season two, confirming high natural inoculum. Environmental data loggers (HOBO MX2301, Onset Computer Corp., USA) recorded leaf wetness duration averaging 9.4 hours per day well above the 6-hour threshold for infection [2,9] .

Material and Methods Material

A total of 4,218 RGB leaf images were captured using smartphone cameras (Samsung Galaxy A52, 64 MP) from six rice varieties in naturally infected fields at Namulonge. Images were taken at 20-30 cm distance from individual leaves under ambient field lighting between 09:00 and 15:00 h. A panel of two plant pathologists labelled each image into five classes: healthy (n = 926), mild blast (1-10% leaf area affected; n = 847), moderate (11-25%; n = 814), severe (26-50%; n = 792), and very severe (>50%; n = 839). Ethics approval: Protocol MAK-CAES/2021-023, approved 4 March 2021.

Methods

Images were resized to 224×224 pixels and augmented (rotation, flipping, brightness adjustment) to expand the training set to 12,654 images. Data were split 70:15:15 into training, validation, and test sets. A VGG-16 network pre-trained on ImageNet was fine-tuned by replacing the final fully connected layers with a 5-class softmax classifier. Training ran for 50 epochs with a learning rate of 0.0001, batch size 32, and Adam optimizer using TensorFlow 2.8 [4] . SVM (RBF kernel) and random forest (500 trees) classifiers were trained on features extracted from the VGG-16 penultimate layer for comparison. Performance was evaluated using accuracy, precision, recall, F1score, and AUC-ROC on the held-out test set [10] .

Results

Comprehensive interpretation

The CNN outperformed both traditional classifiers across all metrics (Table 1, Figure 1). Accuracy dropped with increasing severity, from 97.8% for healthy to 88.3% for very severe (Table 2, Figure 2). The most common misclassification occurred between adjacent severity classes (e.g. moderate-severe), reflecting the continuous nature of disease progression. The CNN's 6.7% accuracy advantage over SVM widened to 19.6% for very severe class, where visual complexity is highest. The detection pipeline (Figure 3) processes images end-to-end in under 0.013 seconds each.

Discussion

The 94.3% accuracy matches or exceeds results from controlledenvironment studies using similar architectures [3,5] while using field-captured images with natural variability. The VGG-16 architecture's deep feature extraction proved effective even with a moderate dataset size of 4,218 original images, thanks to transfer learning from ImageNet weights [4] . The accuracy drop for very severe cases (88.3%) reflects the visual similarity between heavily damaged and necrotic tissue from other causes a known challenge in plant disease image classification [11] .

For practical deployment, the CNN's 78 images per second throughput means a farmer could scan a sample of 50 leaves in under a second using a smartphone app. The key question is whether 88-94% accuracy is sufficient to trigger management decisions. Given that even a rough severity estimate outperforms the common practice of no assessment at all, these error rates seem acceptable for field use [12,13] .

Limitations

The model was trained on images from a single location with specific rice varieties and blast races. Performance on different varieties, under different lighting conditions, or with cooccurring diseases (brown spot, bacterial leaf blight) was not tested [14] . Field deployment would require a lightweight model optimized for mobile devices, and the current VGG-16 architecture may be too large for low-end smartphones [15] .

Conclusion

References

1. Dean R, Van Kan JAL, Pretorius ZA, et al. The top 10 fungal pathogens in molecular plant pathology. Mol Plant Pathol. 2012;13(4):414-430.
2. Ou SH. Rice Diseases. 2nd ed. Commonwealth Mycological Institute; 1985.
3. Mohanty SP, Hughes DP, Salathe M. Using deep learning for image-based plant disease detection. Front Plant Sci. 2016;7:1419.
4. Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556; 2014.
5. Sethy PK, Barpanda NK, Rath AK, Behera SK. Deep feature based rice leaf disease identification using support vector machine. Comput Electron Agric. 2020;175:105527.
6. Ramcharan A, Baranowski K, McCloskey P, Ahmed B, Legg J, Hughes DP. Deep learning for image-based cassava disease detection. Front Plant Sci. 2017;8:1852.
7. Ferentinos KP. Deep learning models for plant disease detection and diagnosis. Comput Electron Agric. 2018;145:311-318.
8. Atuhaire A, Mugerwa S, Karungi J. Rice production systems and blast disease management in Uganda. Uganda J Agric Sci. 2019;19(1):1-14.
9. Katsantonis D, Kadoglidou K, Dramalis C, Puigdollers P. Rice blast forecasting models and their practical value. Phytopathol Mediterr. 2017;56(2):153-174.
10. Chollet F. Deep Learning with Python. Manning Publications; 2018.
11. Barbedo JGA. Factors influencing the use of deep learning for plant disease recognition. Biosyst Eng. 2018;172:84-91.
12. Mwebaze E, Owomugisha G. Machine learning for plant disease incidence and severity measurements from leaf images. IEEE Int Conf Mach Learn Appl. 2016:158-163.
13. Sladojevic S, Arsenovic M, Anderla A, Culibrk D, Stefanovic D. Deep neural networks based recognition of plant diseases by leaf image classification. Comput Intell Neurosci. 2016;2016:3289801.
14. Zhang S, Huang W, Zhang C. Three-channel convolutional neural networks for vegetable leaf disease recognition. Cogn Syst Res. 2019;53:31-41.
15. Howard AG, Zhu M, Chen B, et al. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861; 2017.
16. Picon A, Alvarez-Gila A, Seitz M, Ortiz-Barredo A, Echazarra J, Johannes A. Deep convolutional neural networks for mobile capture device-based crop disease classification in the wild. Comput Electron Agric. 2019;161:280-290.
17. Singh V, Misra AK. Detection of plant leaf diseases using image segmentation and soft computing techniques. Inf Process Agric. 2017;4(1):41-49.
18. Lu Y, Yi S, Zeng N, Liu Y, Zhang Y. Identification of rice diseases using deep convolutional neural networks. Neurocomputing. 2017;267:378-384.