Citrus Plant Disease Identification using Deep Learning with Multiple Transfer Learning Approaches
Citrus plant fruits constitute a significant part of Pakistan's agricultural fruits production. A substantial proportion of citrus fruits is destroyed every year because of different diseases. Citrus plants need to be examined manually to identify the disease prevalence. This paper proposed an...
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
The University of Lahore
2020-09-01
|
| Series: | Pakistan Journal of Engineering & Technology |
| Subjects: | |
| Online Access: | http://dev.ojs.com/index.php/pakjet/article/view/439 |
| Summary: | Citrus plant fruits constitute a significant part of Pakistan's agricultural fruits production. A substantial proportion of citrus fruits is destroyed every year because of different diseases. Citrus plants need to be examined manually to identify the disease prevalence. This paper proposed an in-depth learning approach to identify disease in citrus plants automatically. The dataset used comprised of a small number of citrus plant leaves divided into five categories; 4 of them are disease affected, and the fifth category includes healthy leaves. DenseNet 121 is used as a deep learning model to train the dataset. First, the model is prepared without a transfer-learning approach. Then the model is pre-trained on external data of plant leaves, ImageNet dataset, and a combination of external data and ImageNet data. Model without transfer learning failed to identify the diseases. Model pre-trained on external data resulted in an accuracy of 92%, AUC score of 98.8%, and F1-score of 95%. The model with combined pretraining resulted in an accuracy of 88% and F1-score of 88%. Pre-training on ImageNet data resulted in an accuracy of 82% and F1-score of 87%. |
|---|---|
| ISSN: | 2664-2042 2664-2050 |