Research Information System
Thesis Type: Postgraduate
Institution Of The Thesis: Gazi University, Fen Bilimleri Enstitüsü, Turkey
Approval Date: 2023
Thesis Language: Turkish
Student: Muhammed ERDOĞAN
Supervisor: Oktay Yıldız
Open Archive Collection: AVESIS Open Access Collection
Abstract:
The protection of critically important military and civil settlements maintains its importance today as it was in the past. For this purpose, systems with various sensors are being developed. Extracting information from the data which is provided by the sensors is also important for the most efficient use of the equipment. Radar systems are frequently used for reconnaissance, surveillance and detection purposes. There are rule-based and machine learning-based methods for the classification of objects detected by radar. In machine learning-based approaches, the characteristics of the target object are learned by the model over time without the need for expert opinion. For this reason, these methods are more advantageous than rule-based methods. The studies on the target classification using the radar data with machine learning-based approaches are examined, it can be seen that the radar types are divided into four categories. These are doppler radars, SAR, passive radars and long-range air defense radars. Radars developed based on the Doppler effect are called Doppler radar. Within the scope of the study, various experimental studies are carried out on the unbalanced Doppler data set, in addition to the existing methods and approaches in the literature, it is aimed to increase the classification performance by diversifying the methods in the stages performed for classification and to evaluate the results obtained by comparing them with the studies in the literature. For this purpose, target classification is made from radar data using a public, unbalanced data set. Random forest and SVM algorithms, which are frequently preferred in the literature, and the convolutional neural network model are tested in line with various preprocessing steps and different training approaches. It is aimed to determine the most efficient preprocessing, classifier model and training approach in terms of performance. 99.98% training accuracy is reached with the proposed convolutional neural network model k-fold cross-validation approach with the data set balanced using the SMOTE data preprocessing technique. Performance metrics obtained with test data also achieved more successful results than other methods. Considering the studies carried out in the literature with the same data set, it can be said that the proposed method is at a level that can compete with other methods in the literature.
Key Words
: Radar, unmanned aerial vehicle, K-Fold, SMOTE, convolutional neural
network