Classification and performance measurement of public buildings from satellite images with deep learning methods

Today, since artificial intelligence is a technology that can be applied in every subject, its use for many visual and audio data has increased considerably. Both private and public institutions have started to develop systems that work with various classification algorithms in order to go out of their existing structures and to carry out work easier and faster. Visual classifications used in studies for city and regional planning, especially in municipalities, are insufficient in terms of achieving desired results, elaboration and speed. While providing this type of building detection study through satellite data contributes to the authenticity of the data, testing the deep learning algorithm and performing it with the best performance shows that this study can be adapted to many institutions and / or sectors. In the study carried out on the image, it was ensured that the areas open to the public (hospital, school, airport, mosque and field) were defined. Building classification and detection were analyzed with the deep learning-based YOLOv3 and YOLOv5 library on 300 images, divided into 5 classes, taken from satellite images available for use. Test and training data are separated for deep learning on the data set. Images are trained with the algorithms of the YOLOv3 and YOLOv5 library designed in Python. Faster regional-convolutional neural networks were applied to control the correctly classified data rate before and after the training step. Classification results were compared. Darknet-53 open source neural network framework and real-time object detection system YOLOv3 and YOLOv5 deep learnings model were used to detect these public buildings. A qualified weight model was created by classifying these buildings. Thus, it contributed to object detection in space images in the future ecoinformatics field.