Thesis defence in computer science and Engineering

Defence presentation

1. Deep Learning-Based Plant Disease Classification with Mobile Application Department Of Computer Science & Engineering Jahangirnagar University Department Of Computer Science & Engineering Jahangirnagar University Presented by Manas Sarker Akash Exam Roll: 220503 Session: 2021-2022 Supervised by Samsun Nahar Khandakar Lecturer Department of Computer Science & Engineering Jahangirnagar University

2. Introduction Objective Literature Review Dataset Design and Development Deep Learning Model Deployment Android Application development Experimental Results and Discussions Conclusion 2 Agenda

3. 3 Agricultural production is crucial for global economies , yet crop diseases pose a significant threat to productivity and food security . Traditional manual diagnosis methods are labor-intensive and subjective , leading to inaccurate pesticide applications. Early and accurate automated disease recognition is essential for taking immediate measures. Recent advancements in deep learning have improved disease identification accuracy. However, existing models lack generalization across multiple crops and diseases. To address this gap, a comprehensive dataset comprising 10 common leaf diseases of different vegetables, was curated to develop a generalized deep-learning model and then integrating it with a easy to use mobile application. Introduction

4. 4 A generalized automated deep learning-based leaf disease recognition model for multiple vegetables will certainly help the farmers to increase production efficiency by taking immediate measures against any disease. Therefore, the main goals of this project: Offering an automated system to the farmers and gardeners for quickly identifying a disease by using the symptoms from leafs with high accuracy . Project Objective

5. 4 This research project addresses the generalizability gap by making fundamental contributions, including: Using a large dataset of common leaf diseases for this project. Developing a generalized model capable of accurately detecting and monitoring common leaf diseases across various crops and vegetables. Improving agricultural efficiency by reducing human efforts and costs while boosting production. Project Contributions

6. 6 Literature Review K. Ahmed, T. R. Shahidi, S. M. Irfanul Alam , and S. Momen , “Rice leaf disease detection using machine learning techniques,” in 2019 International Conference on Sustainable Technologies for Industry 4.0 (STI), 2019. A. S. Paymode and V. B. Malode , “Transfer learning for multi-crop leaf disease image classification using convolutional neural network VGG,” Artificial Intelligence in Agriculture, vol. 6, pp. 23–33, 2022. B. Min, T. Kim, D. Shin, and D. Shin, “Data augmentation method for plant leaf disease recognition,” Appl. Sci. (Basel), vol. 13, no. 3, p. 1465, 2023. The literature review explored that A dataset is important for training and validating any recognition model. However, most datasets contain singular or limited types of leaf diseases. Many deep-learning methods have been developed. However, there’s a gap of generalized method capable of dealing with common diseases across multiple crops and vegetables. Therefore, an extensive dataset containing multiple leaf diseases of various crops and vegetables with a generalized deep-learning model is an ultimate necessity.

7. Project Overview The project addressed two main developments. Deep Learning model Generation Web and Mobile Application Development.

8. 8 Dataset Generation (1) This research project has gathered data from Kaggle . It contains a total number of 11,000 original images of 11 common disease classes along with a healthy class of 3 different crops and vegetables.

9. 9 Dataset Generation (2) 11 common leaf diseases: Septoria Leaf Spot, Target Spot, Mosaic Virus, Bacterial Spot, Leaf Mold, Late Blight, Bacterial Spot, Spider Mites, Two Spotted Spider Mite, Early Blight 3 crops and vegetables: Pepper Bell, Potato, Tomato . Data Augmentation: Random rotation, scaling, shearing, shifting, and brightness adjustments were used to augment the images.

10. Class Name Crops and Vegetable Source Number of Image Acquisition Total Number of Images Bacterial Spot Tomato Kaggle 770 1612 Pepper Bell 842 Healthy Tomato Kaggle 800 2600 1000 Potato Pepper Bell 800 Late Blight Tomato Kaggle 1000 2000 Potato 1000 Leaf Mold Tomato Kaggle 373 373 Bacterial Spot Tomato Kaggle 700 700 M osaic V irus Tomato Kaggle 373 373 Target S pot Tomato Kaggle 566 566 Septoria L eaf S pot Tomato Kaggle 560 560 Spider M ites T wo S potted S pider M ite Tomato Kaggle 700 700 Early Blight Tomato Kaggle 800 1600 Potato 800 Types of classes

11. Potato: Healthy Potato: Early Blight Potato: Late Blight Tomato: Bacterial spot Tomato: Leaf Mold Pepper Bell: Bacterial spot Sample Dataset

12. 12 Deep Learning Model 30% 70% Fig: The process of leaf disease detection using Deep Learning model

13. Model Architecture VGG-16 VGG-19 DenseNet-121 ResNet-50 Inception-V4 Input Size 224*224 224*224 224*224 224*224*3 229*229 Conv. Layers 13 16 121 48 48 Max Pooling Layer 5 5 4 1 5 Dense Layer 3 - - 2 - Drop-out layer 2 - - 2 - Flatten Layer 1 - - 1 - Filter Size 3,3 3,3 3,3 3,3 1,3,5 Stride 1,2 1,2 1 2 1 Parameter 138M 138M 7M 23M 42-45M Fully Connected Layer 3 3 2 1 3 Fig: Detailed settings of the investigated deep CNNs

14. Performance Matrix Accuracy (%) = × 100 Precision (%) = × 100 Recall (%) = × 100 F1-score (%) = 2 × × 100 True Positive (TP) - the model predicts the correct sample as correct, True Negative (TN) - the model predicts the incorrect sample as incorrect, False Positive (FP) - the model predicts the incorrect sample as correct, False Negative (FN) - the model predicts the correct sample as incorrect. The class-wise TP, TN, FP, and FN are determined by using those equations.  

15. 15 Experimental Results Method Accuracy Densenet-121 97.58 Resnet-50 94.63 Inception-v4 [4] 92.33 VGG-16 [5] 90.87

16. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) (k) (l) Fig: DenseNet-121 correctly recognized the leaf diseases in (a), (b), (d), (e), (f), (g), (h), (i) and (1); ResNet-50 correctly recognized the leaf diseases in (a),(c), (d), (f) and (h) Experimental Results (2)

17. Category Accuracy Precision Recall F1 score Septoria Leaf Spot 0.97 0.85 0.97 0.91 Target Spot 1.00 0.93 0.95 0.94 Mosaic Virus 0.95 0.96 1.00 0.93 Bacterial Spot 1.00 0.98 0.96 0.97 Leaf Mold 0.97 0.95 0.93 0.94 Late Blight 0.97 0.97 0.97 0.97 Bacterial Spot 0.97 0.93 0.93 0.93 Spider Mites 0.96 1.00 0.97 0.99 Two Spotted Spider Mite 0.95 0.92 0.90 0.92 Early Blight 0.98 0.97 0.92 0.99 Healthy 1.00 0.99 0.99 1.00 Average 0.97 0.95 0.95 0.95 Experimental Results (3) Table: Accuracy, Precision, Recall, F1 score of different disease categories by Densenet-121 model

18. Experimental Results (2) Fig: confusion matrix of leaf disease detection

19. 19 Experimental Results (6)

20. Model testing using API Fig: Testing the model with Postman API

21. Web Interface (1) Fig: Web Interface for crop leaf daises detection PREDICT

22. Web Interface (2) Fig: Web Interface for crop leaf daises detection

23. Android Interface (1) Fig: Android Application Homepage

24. Android Interface (2) Fig: Showing the prediction

25. Future Work Future Work for This Project: Expansion of dataset with diverse crops and real-world images. Multilingual user interface for broader accessibility. Integration with IoT sensors for comprehensive plant health monitoring. Development of a recommendation system for treatment plans. Improved model architectures like Vision Transformers or EfficientNet. Implement explainability methods like Grad-CAM or LIME. Edge computing for offline and real-time detection. Early disease detection using spectral imaging techniques. Community and farmer engagement for collaborative learning.

26. 26 The agricultural sector confronts a formidable threat from crop diseases, which can precipitate substantial losses in food production. A comprehensive dataset comprising 11 common leaf diseases, affecting 20 widely consumable crops and vegetables, was acquired. This study addresses these challenges by conducting a thorough analysis of various transfer learning models, with the aim of accurately classifying common leaf diseases across multiple crops and vegetables. Among the pre-trained architectures, DenseNet-121 emerged as the top performer with an impressive classification accuracy of 97.58% ., outstripping ResNet-50 and others. Conclusions

27. 22 References K. Ahmed, T. R. Shahidi, S. M. Irfanul Alam , and S. Momen , “Rice leaf disease detection using machine learning techniques,” in 2019 International Conference on Sustainable Technologies for Industry 4.0 (STI), 2019. A. S. Paymode and V. B. Malode , “Transfer learning for multi-crop leaf disease image classification using convolutional neural network VGG,” Artificial Intelligence in Agriculture, vol. 6, pp. 23–33, 2022. X. Niu , M. Wang, X. Chen, S. Guo, H. Zhang, and D. He, “Image segmentation algorithm for disease detection of wheat leaves,” in Proceedings of the 2014 International Conference on Advanced Mechatronic Systems, 2014. B. Min, T. Kim, D. Shin, and D. Shin, “Data augmentation method for plant leaf disease recognition,” Appl. Sci. (Basel), vol. 13, no. 3, p. 1465, 2023. M. Yu, X. Ma, and H. Guan, “Recognition method of soybean leaf diseases using residual neural network based on transfer learning,” Ecol. Inform., vol. 76, no. 102096, p. 102096, 2023. S. Rehman et al., “A framework of deep optimal features selection for apple leaf diseases recognition,” Comput . Mater. Contin., vol. 75, no. 1, pp. 697–714, 2023. J. Zhang, Y. Rao, C. Man, Z. Jiang, and S. Li, “Identification of cucumber leaf diseases using deep learning and small sample size for agricultural Internet of Things,” Int. J. Distrib . Sens. Netw ., vol. 17, no. 4, p. 155014772110074, 2021. M. Yu, X. Ma, H. Guan, M. Liu, and T. Zhang, “A recognition method of soybean leaf diseases based on an improved deep learning model,” Front. Plant Sci., vol. 13, p. 878834, 2022. K. Tian, J. Zeng, T. Song, Z. Li, A. Evans, and J. Li, “Tomato leaf diseases recognition based on deep convolutional neural networks,” J. Agric. Eng., 2022. S. U. Rahman, F. Alam , N. Ahmad, and S. Arshad, “Image processing based system for the detection, identification and treatment of tomato leaf diseases,” Multimed . Tools Appl., vol. 82, no. 6, pp. 9431–9445, 2023.

28. Thank You 23