Classification of Satellite Images With Deep Convolutional Neural Networks and Its Effect on Architecture

Abstract

Unlike traditional machine learning methods, deep learning methods that can learn from image, video, audio, and text data, especially recently with the increase in hardware power, are also increasing in success. Considering the success and benefits of deep learning methods in many different fields with increasing data, similar effects are expected in architecture. In this study, we focused on textures by going down to specifics rather than general images. In this direction, a total of 4500 satellite images belonging to cloud, desert, green areas and water bodies were classified in the model developed using deep convolutional neural networks. In the developed model, 0.97 accuracy for cloud images, 0.98 accuracy for desert images, 0.96 accuracy for green areas images and 0.98 accuracy for water bodies images were obtained in the classification of previously unused test data (675 images). Although there are similarities in the images of cloud and desert, and images of green areas and water bodies, this success in textures shows that it can be successful in detecting, analyzing, and classifying architectural materials. Successful recognition, analysis and classification of architectural materials and elements with deep convolutional neural networks will be able to facilitate the acquisition of appropriate and useful data through shape recognition among many data, especially at the information collection phase in the architectural design process. Thus, it will help to take more accurate decisions by obtaining more comprehensive data that cannot be obtained from manual data analysis. Learning the distinctive features for classification of data in deep convolutional neural networks also explains architectural design differences and similarities. This situation reveals the hidden relationship in the designs and thus can offer architects the opportunity to make creative and original designs.