الفهرس 
BASICS AND BACKGROUNDOnly 14 pages are availabe for public view
 |  | from | 102 |  |  |
| | | from | 102 | | |
Abstract
Agriculture is still a significant economic sector. Plant diseases have a big effect on plant
yield and quality. So, plant diseases must be identified and prevented at their initial stage.
Unfortunately, manual plant disease diagnosis is time-consuming and inaccurate.
In previous work, a variety of machine learning (ML) and deep learning (DL) techniques
have been employed to address the problem of identifying and classifying plant leaf
diseases. However, there are limitations in current studies, such as handcrafted features
used in machine learning techniques that are inappropriate for complex plant leaf
diseases. Also, the overfitting problem occurs when many training parameters are used.
Finally, the overall accuracy is directly proportional to the dataset size.
To address the limitations of previous studies, this work proposes two diagnosis methods
for plant leaf disease classification. The proposed approach has followed two different
directions, one for improving the classification of grape leaf diseases uses deep learning
with the concepts of transfer learning (TL) and ensemble learning (EL) based on three
famous architectures (VGG16, VGG19, and Xception). The proposed system has an
accuracy of 99.82% with 73 million training parameters.
The other direction is enhancing the classification of the famous crops (tomato, potato,
and pepper) using transfer ensemble learning (MobilenetV3 small and Resnet50). The
results of the model show that it has only 37 million training parameters, which saves
training time and computational power and reduces overfitting, with an accuracy of
99.50%. Both methods address false negatives and false positives, ensuring reliable and
accurate plant disease classification at an early stage.
The proposed methods can work with any image dataset and achieve high performance.
They are validated on the benchmark open dataset Plantvillage, demonstrating excellent
performance for both proposed diagnosis approaches.