الكشف المؤتمت لآفة دوباس النخيل باستخدام الذكاء الاصطناعي

Abdullah Mazin Mughames AL-Mahmood

Abstract

The Dubas Palm Insect (Ommatissus Lybicus) is a highly damaging pest that affects palm trees and crops, and the last are crucial resources for human life. Early detection and localization of Dubas can significantly decrease economic losses and human effort. However, traditional insect detection methods are time-consuming and require a great deal of human effort, especially for large farms and tall palm trees. In order to overcome these challenges, this thesis proposes an intelligent system that can capture images and perform diagnostics.
The proposed work employs a drone (Mavic Air 2- 48MP) to capture images that are entered into the system to classify palms into four classes: healthy, infected by bugs, infected at the honeydew stage, and mixed (bugs and honey). The collected images are preprocessed before the classification stage by slicing, magnifying, enhancing, and Lifting Wavelet Transform (LWT) decomposing. The classification phase involves training the data using six types of pre-trained deep learning networks, including AlexNet, Xception, InceptionV3, ResNet50, Vgg19, and DenseNet121. Three different strategies are used for each pre-trained network: with transfer learning, without transfer learning, and Fine-Tuned. In the test phase, the same preprocessing stages are repeated for the input image and then the approximation rules are applied to classify the input to one of the four categories.
The system performance metrics, such as classification accuracy, and processing time are calculated based on our dataset. The dataset has been collected during two seasons at Kerbala- Aoun from eight farms and over six months. The experimental results show the efficiency of the proposed system in Dubas classification for different deep learning networks.
There are three different scenarios for image classification using various deep-learning models. The first scenario involves without pre-processing, the worst accuracy achieved by DenseNet121 is 81.65% and the best accuracy achieved by InceptionV3 is 88.24%. VGG19 spends the highest training time 10.31 hours, while AlexNet’s training time is 8.61 hours. In the second scenario, pre-processing excluding LWT is used. The best accuracy is more than 97%, which is achieved by ResNet50, while the lowest accuracy and least achieved by Xception is 88.21% with a training time is 10.21 hours. In the third scenario, pre-processing including LWT is used. The best accuracy is by ResNet50 more than 94 %. However, the training time is significantly reduced by more than 60%. The above scenarios are based on non-transfer learning, other results are based on fine-tuning offering an accuracy of 99% for InceptionV3 and preprocessing. After completing the learning of the networks, the tests meet the real-time requirements. It does not exceed 0.3 seconds for the test. This research concluded that the use of artificial intelligence techniques can be an effective solution to the problem of automatic detection of Dubas bug, and can be used in modern agricultural applications to improve farm productivity and agricultural product quality, thus reducing economic and manpower losses and environmental impact.