THESES

Classification of olive leaf diseases using deep learning techniques

Convolutional neural network models are known to be very successful in diagnosing diseases and pests on plant leaf images. In this thesis, three different leaf types belonging to olive plant were classified with convolutional neural network model. The data set consists of 3400 olive leaf images. There are 3 different classes Olive Peacock Spot disease, Aculus Olearius and healthy leaves. The effect of different optimization algorithms and iteration numbers on the classification success was observed with the help of Convolutional Neural Networks (CNN).CNN models used in the thesis were trained with Tesla GPUs in Google Colab cloud service. Anaconda (Spyder) is written in python programming language on the IDE platform. The web model of the trained model was created with the help of the Flask framework. The proposed CNN model achieved an accuracy rate of 84% with the Adam optimization algorithm over 100 iterations. Using the transfer learning method, advanced CNN models such as VGG16 and VGG19 were trained with the generated data set. According to comparative test results, VGG16 model gave the best result with 88% accuracy. The CNN model, which was proposed in the thesis study, was applied to the dataaugmentation process with ImageDataGenarator class of Keras framework. This model, which is applied dataaugmentation method, was trained in 100 iterations by using Adam optimization algorithm. As a result of the dataaugmentation process, the proposed CNN model achieved the best classification success with 94% accuracy rate.

Neşe UYSAL, 2020. Danışman: Sinan UĞUZ

In vermicompost production regaining red californian worm eggs by separating from compost with deep learning techniques

Deep learning techniques, which are being used effectively in many areas today, will become more and more important every day and will increase the competitiveness of producers by contributing significantly to the digital transformation in the agricultural sector. In this thesis work, worm eggs detected by CNN model created using SSD architecture were separated from compost and recycled to production. The data set consists of 1000 compost images containing 3809 worm eggs and a single class named “cocoon” representing the worm eggs. The success of different iteration numbers in object detection has been observed in the SSD architecture used by transferring learning over VGG16. The CNN model used in the thesis was written in the Anaconda Spyder IDE platform in Python programming language and was trained with a GeForce RTX 2080 GPU. The training of the network created with the SSD model proposed in the thesis study was carried out in 1000 epochs by doing data augmentation and transfer learning with 1000 images and 1e-5 learning rate in the data set. As a result of the training, the loss value decreased to 2.527 at the 800th epoch and to 2.494 at the 1000th epoch, and after this point, it followed a horizontal course with converging values. AP value of the single-class model increased up to 84.9% at the 1000th epoch and the best result was obtained.

Ali ÇELİK, 2020. Danışman: Sinan UĞUZ

Real-time monitoring of community mobility with deep learning based methods for smart campuses

Various problems come to the fore due to the fact that urban environments and therefore campus environments are hosting more and more dense crowds of people. The first thing to do to solve these problems is; These crowds need to be identified and analyzed. In this direction, in this thesis, it is aimed to determine and examine the student crowds in the university campus. For this purpose, object detection was made using deep learning methods in order to detect people in the images. The commonly used datasets in the literature were examined and it was seen that the CrowdHuman dataset was the most appropriate dataset. There are 15 000 images in the relevant data set and approximately 470 000 tagged human data. State-of-the-art models such as Faster RCNN, YOLO, and RetinaNet were used in the trainings. In the experiments, 78% mAP score was obtained with YOLOv4. The models trained here were run on the server, and real-time object detection was made on the images obtained from the IP cameras in the campus. With the detection of human density in the image, the density status of the relevant region is shown on the mobile application.

Ahmet Bestami KÖSE, 2023. Danışman: Sinan UĞUZ

Pest detection based on deep learning techniques and population estimation in greenhouses

Greenhouse cultivation has a significant activity area in Turkey, particularly in the Mediterranean and Aegean regions. The products produced in greenhouses incur losses due to diseases, pests, nutritional inadequacy, misused pesticides, or improper storage. Agricultural pests, can be considered one of the most significant economic and productivity losses for greenhouse products. Detection of thrips at the preliminary stages, which is endemic in greenhouses, has immense importance. Frankliniella occidentalis, a species belonging to the thrips family, was discussed in this thesis. To observe thrips population growth, the pests on the traps are counted weekly with relevant equipment. In this respect, this process requires high labour and time. In the thesis, counting pests by using deep learning-based algorithms instead of a human factor, was aimed. Within this scope, yellow and blue sticky trap images in greenhouses in Antalya were obtained. Also, collection date from the greenhouses, temperature and humidity information of the greenhouses were obtained. The collected images were divided and a dataset was developed. This was trained with 4 different state-of-the-art models. YOLOv7 with DarkNet53 backbone was the most successful model with an AP50 value of 0.889. The collected numerical data were trained with 4 different machine-learning models. In the thesis, in which time series were used as a feature, the linear regression model gave the most successful result. A trend analysis was performed using this model and the population was estimated. An interface is designed to inform the user about the pest situation in the greenhouse.

Gülhan ŞİKAROĞLU, 2023. Danışman: Sinan UĞUZ, 2. Danışman: Prof. Dr. İsmail KARACA

Detection the quality of raisin in real time by deep learning techniques

Deep learning has a very advanced usage area and diversity, as well as the growth of computational power, deep learning algorithms have provided remarkable results in the field of agriculture. Deep learning is used in agriculture analysis of crops, pests, weeds, irrigation, fertilization, weather (climatization), farmland analysis, soil analysis and autonomous machines used in agriculture. In the study, using deep learning techniques, the conformity of Sultana raisin samples in TS 3411 to the standard was determined with a real-time system. The raisin dataset consists of three classes as 1.Quality, 2.Quality and Grape Stalk and 10544 images. In the study, train were made with YOLOv5, YOLOv6, YOLOv7, YOLOX, FCOS and TOOD models. In the train, the lowest TOOD with 0.018 of the total loss value, TOOD with 84.6% for mAPs and YOLOv5 with 88.7% for mAP^(50:05:95) were obtained. However, since YOLOv6 has higher fps in real-time detection, YOLOv6 is preferred in real-time detection.

Abdullah YAĞIZ, 2023. Danışman: Sinan UĞUZ