Classification of paddy leaf diseases using an edge-enhanced resnet-50 framework

Introduction

Rice is a staple crop across Asia and a major source of dietary energy and farm income. Its production is constrained by multiple biotic and abiotic stresses, and foliar diseases remain especially important because they directly reduce photosynthetic area and may spread rapidly under favorable field conditions (Mahlein, 2016; Patrício and Inácio, 2018; Vo-Tong Xuan, 2018) [21,25,37] . Bacterial blight, blast, brown spot, and sheath rot are among the common paddy diseases that produce visible symptoms on leaf blades and sheaths. Field diagnosis is traditionally based on visual inspection by farmers or plant-pathology specialists. This approach is practical but can be inconsistent when symptoms are subtle, when multiple diseases share similar colors or lesion shapes, or when field lighting and background conditions vary. Automated image analysis can support diagnosis by quantifying lesion color, texture, and boundary structure (Barbedo et al., 2016; Prajapati et al., 2017) [3,28] . Classical computer-vision pipelines generally rely on segmentation followed by handcrafted color, texture, and shape descriptors. Deep convolutional neural networks can instead learn hierarchical visual representations directly from images, and residual networks are particularly effective because skip connections support the optimization of deeper architectures (He et al., 2016). Edge information may complement learned features by emphasizing lesion contours and transitions between healthy and affected tissue. Canny edge detection provides a systematic procedure for smoothing, gradient estimation, non-maximum suppression, and edge linking (Canny, 1986) [7] . The objective of this study was to assess an edge-enhanced ResNet-50 workflow for five-class paddy leaf classification. The main contributions are: (i) a preprocessing sequence combining HSV-based background removal and canny edge extraction; (ii) evaluation on a field-acquired dataset containing four disease classes and healthy leaves; and (iii) class-wise assessment using precision, recall, and F1-score in addition to overall accuracy. The proposed model was also compared descriptively with a previously reported DNN-JOA baseline (Ramesh and Vydeki, 2020) [30] .

Related Work

Rice disease recognition has been investigated using both conventional machine-learning and deep-learning techniques. Earlier studies applied support vector machines, k-nearest-neighbor classifiers, radial-basis-function networks, clustering, expert systems, optimization methods, and handcrafted image descriptors to rice and other crop diseases (Pavithra et al., [2,4,15,17,18,24,26,27,29,31,34,36,38,39] . These approaches can perform well when segmentation and feature engineering are reliable, but their accuracy often depends strongly on imaging conditions and manually selected descriptors. Deep-learning studies have reported high classification performance for rice and several other crops. Ramesh and Vydeki (2020) [30] combined an optimized deep neural network with the Jaya algorithm, while Sharma et al. (2022) [32] compared multiple transfer-learning architectures for rice disease diagnosis. Lu et al. (2017) used a deep convolutional network for rice disease identification, and Shrivastava and Pradhan (2021) [33] evaluated color features within a machine-learning framework. Large multi-crop studies using AlexNet, GoogLeNet, VGG, EfficientNet, residual networks, and other convolutional architectures have further demonstrated the capacity of deep models to learn disease-related features (Mohanty et [1,6,8,9,10,11,12,13,19,22,23,35] . Reported accuracies across studies are not directly comparable because datasets differ in crop species, class definitions, image counts, background complexity, field conditions, and validation design. Controlled comparisons on the same split and under the same training protocol are therefore preferable to comparisons based only on published headline accuracy.

Table 1: Representative studies on image-based plant disease classification

Study

Crop / dataset Method Reported performance Ramesh and Vydeki (2020) [30] Paddy; own dataset DNN with Jaya optimization 97% accuracy Sharma et al. (2022) [32] Rice; own dataset InceptionResNetV2, Xception, ResNet50, MobileNet, InceptionV3 98.9% accuracy

Lu et al. (2017)

Rice; own dataset AlexNet-inspired CNN 95% accuracy Shrivastava and Pradhan (2021) [33] Rice; real-field images SVM using color features 94.65% accuracy Mohanty et al. (2016) [22] PlantVillage; 14 crops AlexNet and GoogLeNet 99.35% accuracy Ferentinos (2018) [12] PlantVillage; multiple crops AlexNetOWTBn and VGG 99.53% accuracy Chowdhury et al. (2021) [8] PlantVillage; tomato EfficientNet-B4/B7 Up to 99.95% Dai et al. (2024) [10] PlantVillage DFN-PSAN feature-fusion network 95.27% accuracy DeChant et al. (2017) [11] Maize; field imagery CNN pipeline 97% accuracy Fuentes et al. (2017) [13] Tomato; field images VGG and ResNet detector 83% reported AP/IoU performance Liu et al. (2018) [19] Apple; own dataset AlexNet-based CNN 98% accuracy Oppenheim and Shani (2017) [23] Potato; own dataset VGG-based CNN 96% accuracy Verma et al. (2020) [36] Plant leaves Capsule network 91.83% accuracy Bhatia et al. (2020b) [5] Tomato powdery mildew Hybrid SVM-logistic regression 92.73% accuracy

Materials and Methods

Dataset and Class Distribution

The dataset consisted of 650 paddy leaf images acquired in a farm environment using a high-resolution digital camera. Five classes were represented: blast (170 images), brown spot (150), bacterial blight (125), sheath rot (110), and healthy leaves (95). The class distribution was moderately imbalanced, with blast as the largest class and healthy leaves as the smallest. Representative images are shown in Figure 2.

Image Preprocessing and Background Removal

All images were resized to 300 x 450 pixels. Each RGB image was converted to HSV color space, and the saturation channel was used for foreground-background separation. An empirically selected threshold value of 90 generated a binary mask. The mask was combined with the original RGB image, and background pixels were assigned a value of zero. Canny edge detection was then applied to the processed image to emphasize lesion boundaries and local intensity transitions. The complete workflow is summarized in Figure 1, while Figures 3 and4 illustrate the background-removal and edge-extraction stages.

Color and texture characterization

Color and texture descriptors were computed from the disease region to characterize the processed images. For RGB images, the mean and standard deviation of the red, green, and blue Channels were calculated. Mean hue, saturation, and value were obtained in HSV space, and mean L, a, and b values were computed in CIELAB space. For a channel containing N pixels with intensity I_k, the mean and standard deviation were calculated as:

Texture was represented using the gray-level co-occurrence matrix (GLCM), which describes the joint occurrence of gray-level pairs at a specified spatial displacement (Haralick et al., 1973) [14] . The four descriptors used in the study were homogeneity, contrast, correlation, and energy: Homogeneity = sum(i,j) P(i,j) / [1 + (i -j)^2] (3) Contrast = sum(i,j) (i -j)^2 P(i,j) (4) Correlation = sum(i,j) [(i -mu_i)(j -mu_j) P(i,j)] / (sigma_i sigma_j)

(5) Energy = sum(i,j) P(i,j)^2 (6) The extracted descriptors were scaled to the interval [0, 1] using min-max normalization. These descriptors were used to document the visual characteristics of segmented lesions; the deep classifier operated on the processed image representation.

x' = (x -x_min) / (x_max -x_min) (7)

3.4 Edge-Enhanced ResNet-50 Classifier: ResNet-50 is a residual convolutional neural network composed of an initial convolution and pooling stage followed by four groups of residual blocks. Identity connections allow the input of a residual block to bypass one or more transformations, reducing optimization difficulty in deep networks (He et al., 2016) [16] . For this study, the network output layer was configured for five classes: blast, bacterial blight, brown spot, sheath rot, and healthy. The edgeenhanced image representation was supplied to the ResNet-50 classification pipeline implemented in MATLAB.

Dataset partitioning and evaluation design

The dataset was partitioned into 70% training, 10% validation, and 20% testing subsets, corresponding to 455, 65, and 130 images, respectively. The test set was reserved for final performance estimation. A previously reported DNN-JOA confusion matrix for the same five class labels was used as a descriptive baseline (Ramesh and Vydeki, 2020) [30] . Because the full training protocol and predictions of the baseline were not available, the comparison was not treated as a controlled significance test.

Performance Metrics

For each class, true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) were obtained from the confusion matrix. Precision, recall, F1-score, and overall accuracy were calculated as follows:

Accuracy = (TP + TN) / (TP + TN + FP + FN) (8) Precision = TP / (TP + FP) (9) Recall = TP / (TP + FN) (10) F1-score = 2 x (Precision x Recall) / (Precision + Recall) (11) Macro-averaged metrics were obtained by calculating the unweighted mean across the five classes. Weighted metrics used the number of test samples in each class as weights.

Results

DNN-JOA Baseline

The baseline confusion matrix contained 130 test observations, of which 121 were correctly classified, corresponding to an overall accuracy of 93.08%. The principal errors involved sheath rot, bacterial blight, and brown spot.

Comparative Performance

Relative to the DNN-JOA baseline, the proposed model increased overall accuracy by 4.62 percentage points. Macro F1-score increased from 93.09% to 97.62%. The largest class-wise F1 improvement occurred for bacterial blight, followed by sheath rot.

Because the baseline predictions were not available for the same individual samples, no paired significance test was performed. [32,33] and with broader crop-disease studies reporting high classification performance ( [1,6,8,9,10,11,12,13,19,22,23,35] . Nevertheless, published percentages should be interpreted in the context of each dataset. Laboratory images with uniform backgrounds are generally easier than field images, and random image-level splitting can overestimate performance if near-duplicate images from the same plant or acquisition sequence appear in multiple subsets.

The comparison with DNN-JOA suggests an improvement, but the contribution of The model should be evaluated under variable lighting, occlusion, leaf orientation, symptom severity, and mixed infections.

Reporting calibration, confidence intervals, class-wise receiver operating characteristics, inference time, model size, and failurecase visualizations would further strengthen practical evaluation. Lightweight architectures should also be considered for mobile or edge-device deployment.

Conclusion

This study presented an edge-enhanced ResNet-50 framework for classifying healthy paddy leaves and four important leaf diseases: bacterial blight, blast, brown spot, and sheath rot. The workflow combined HSV-based background removal, Canny edge extraction, and residual convolutional classification. On a 130-image test set, the model achieved 97.69% accuracy and a macro F1-score of 97.62%, outperforming the reported DNN-JOA baseline in the available comparison. The results support the potential of deep residual learning for automated paddy disease screening. However, the model should be regarded as a promising experimental classifier rather than a field-ready diagnostic system until it has been validated using controlled ablation, repeated experiments, and independent multi-location data.

References

1. Amara J, Bouaziz B, Algergawy A. A deep learning-based approach for banana leaf diseases classification. In: Datenbanksysteme für Business, Technologie und Web -Workshopband. 2017.
2. Astonkar RS, Shandilya VK. Detection and analysis of plant diseases using image processing. Int Res J Eng Technol. 2018;5(4):3191-3.
3. Barbedo JGA, Arnal LV, Koenigkan TTS. Identifying multiple plant diseases using digital image processing. Biosyst Eng. 2016;147:104-16.
4. Bhatia A, Chug A, Singh AP. Application of extreme learning machine in plant disease prediction for highly imbalanced dataset. J Stat Manag Syst. 2020;23(6):1059-68.
5. Bhatia A, Chug A, Singh AP. Hybrid SVM-LR classifier for powdery mildew disease prediction in tomato plant. In: 2020 7th International Conference on Signal Processing and Integrated Networks (SPIN). IEEE; 2020. p. 218-23.
6. Brahimi M, Boukhalfa K, Moussaoui A. Deep learning for tomato diseases: Classification and symptoms visualization. Appl Artif Intell. 2017;31(4):299-315.
7. Canny J. A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell. 1986;8(6):679-98.
8. Chowdhury MEH, Rahman T, Khandakar A, Ayari MA, Khan AU, Khan MS, et al. Automatic and reliable leaf disease detection using deep learning techniques. AgriEngineering. 2021;3(2):294-312.
9. Cruz AC, Luvisi A, De Bellis L, Ampatzidis Y. X-FIDO: An effective application for detecting olive quick decline syndrome with deep learning and data fusion. Front Plant Sci. 2017;8:1741.
10. Dai G, Tian Z, Fan J, Sunil CK, Dewi C. DFN-PSAN: Multi-level deep information feature fusion extraction network for interpretable plant disease classification. Comput Electron Agric. 2024;216:108481.
11. DeChant C, Wiesner-Hanks T, Chen S, Stewart EL, Yosinski J, Gore MA, et al. Automated identification of northern leaf blight-infected maize plants from field imagery using deep learning. Phytopathology. 2017;107(11):1426-33.
12. Ferentinos KP. Deep learning models for plant disease detection and diagnosis. Comput Electron Agric. 2018;145:311-8.
13. Fuentes A, Yoon S, Kim SC, Park DS. A robust deep-learning-based detector for real-time tomato plant diseases and pests recognition. Sensors (Basel). 2017;17(9):2022.
14. Haralick RM, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern. 1973;SMC-3(6):610-21.
15. Hassanien AE, Gaber T, Mokhtar U, Hefny H. An improved moth-flame optimization algorithm based on rough sets for tomato diseases detection. Comput Electron Agric. 2017;136:86-96.
16. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016. p. 770-8.
17. Kamal M, Mahanijah ANI, Masazhar FAR. Classification of leaf disease from image processing technique. Indones J Electr Eng Comput Sci. 2018;10(1):191-200.
18. Krishnakumar A, Narayanan A. A system for plant disease classification and severity estimation using machine learning techniques. In: Advances in Intelligent Systems and Computing. Springer; 2019.
19. Liu B, Zhang Y, He D, Li Y. Identification of apple leaf diseases based on deep convolutional neural networks. Symmetry (Basel). 2018;10(1):11.
20. Lu Y, Yi S, Zeng N, Liu Y, Zhang Y. Identification of rice diseases using deep convolutional neural networks. Neurocomputing. 2017;267:378-84.
21. Mahlein AK. Plant disease detection by imaging sensors: Parallels and specific demands for precision agriculture and plant phenotyping. Plant Dis. 2016;100(2):241-51.
22. Mohanty SP, Hughes DP, Salathé M. Using deep learning for image-based plant disease detection. Front Plant Sci. 2016;7:1419.
23. Oppenheim D, Shani G. Potato disease classification using convolution neural networks. Adv Anim Biosci. 2017;8(2):244-9.
24. Pantazi XE, Moshou D, Tamouridou AA. Automated leaf disease detection in different crop species through image features analysis and one-class classifiers. Comput Electron Agric. 2019;156:96-104.
25. Patrício D, Inácio RR. Computer vision and artificial intelligence in precision agriculture for grain crops: A systematic review. Comput Electron Agric. 2018;153:69-81.
26. Pavithra S, Priyadharshini A, Praveena V, Monika T. Paddy leaf disease detection using SVM classifier. Int J Commun Comput Technol. 2015;3(1):16-20.
27. Pinki F, Tazmim N, Khatun SMMI. Content-based paddy leaf disease recognition and remedy prediction using support vector machine. In: Proceedings of the 20th International Conference on Computer and Information Technology. 2017. p. 1-5.
28. Prajapati BH, Shah JP, Dabhi VK. Detection and classification of rice plant diseases. Intell Decis Technol. 2017;11(3):35773.
29. Ramesh S, Vydeki D. Application of machine learning in detection of blast disease in South Indian rice crops. J Phytol. 2019;11(1):31-37.
30. Ramesh S, Vydeki D. Recognition and classification of paddy leaf diseases using optimized deep neural network with Jaya algorithm. Inf Process Agric. 2020;7(2):249-60.
31. Rath AK, Meher JK. Disease detection in infected plant leaf by computational method. Arch Phytopathol Plant Prot. 2019;52(19-20):1348-58.
32. Sharma M, Kumar CJ, Deka A. Early diagnosis of rice plant disease using machine learning techniques. Arch Phytopathol Plant Prot. 2022;55(3):259-83.
33. Shrivastava VK, Pradhan MK. Rice plant disease classification using color features: A machine learning paradigm. J Plant Pathol. 2021;103:17-26.
34. Singh V, Misra AK. Detection of plant leaf diseases using image segmentation and soft computing techniques. Inf Process Agric. 2017;4(1):41-49.
35. 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.
36. Verma S, Chug A, Singh AP. Exploring capsule networks for disease classification in plants. J Stat Manag Syst. 2020;23(2):307-15.
37. Vo-Tong X. Rice production, agricultural research, and the environment. In: Vietnam's Rural Transformation. Routledge; 2018. p. 185-200.
38. Yusof M, Mohd NF, Rosli M, Othman R, Mohamed MHAA. M-DCocoa: M-agriculture expert system for diagnosing cocoa plant diseases. In: Proceedings of the International Conference on Soft Computing and Data Mining. 2018. p. 363-71.
39. Zhang S, Wang H, Huang W, You Z. Plant diseased leaf segmentation and recognition by fusion of superpixel, K-means, and PHOG. Optik. 2018;157:866-72.