Investigation of the effect of genetic algorithm-based hyperparameter optimization on the yolo model for aircraft type detection from satellite images

In this thesis, a deep learning–based approach was employed for the detection of aircraft in satellite imagery, and the effects of different training strategies on model performance were investigated. The accurate detection of objects from satellite images constitutes a current research topic of significance in both civilian and military domains. For this purpose, the YOLOv5 architecture was adopted, focusing on three types of aircraft that share visually similar features but serve different functions. These classes consist of military fighter aircraft, civilian passenger aircraft, and military cargo aircraft. Unlike commonly used benchmark datasets in the literature, the dataset utilized in this study was created by manually collecting satellite images of various airports worldwide through Google Earth Pro and labeling them in YOLO format. The training process was conducted through three experimental studies. In the first stage, predefined hyperparameters were employed without modification, serving as a baseline for subsequent experiments. In the second stage, various data augmentation techniques were applied to enhance the model's generalization capability. In the final stage, genetic algorithm–based hyperparameter optimization was implemented alongside data augmentation. Since the dataset contained a limited number of samples, spatial and photometric transformations were employed to increase diversity, thereby strengthening the generalization ability of the model. Furthermore, genetic algorithm–based hyperparameter optimization contributed to a more efficient training process. The comparison of the three experimental studies revealed that the highest performance was achieved in the scenario where genetic algorithm–based hyperparameter optimization was combined with data augmentation. In this case, the mAP@0.5 increased from 75.5% to 89.9% (+14.4 points), while the mAP@0.5:0.95 improved from 45.3% to 60.2% (+14.9 points). Precision rose from 71.6% to 85.3% (+13.7 points), recall from 70.5% to 83.6% (+13.1 points), and the F1-score from 70.6% to 84.4% (+13.8 points). The findings demonstrate that, in object detection tasks conducted on datasets with limited sample sizes that challenge the learning process, the combined use of data augmentation and hyperparameter optimization provides significant improvements in model performance. Therefore, the proposed approach is regarded as a viable option for both civilian and military applications.