Classification of buildings from VHR satellite images using ensemble of U-Net and ResNet

Abstract

The urbanization rate of India is 35.9 % by 2022 reports. In this 45.23 % of urbanization is happening in Maharashtra and it is the third most urbanized state of India after Tamil Nadu and Kerala. In metropolitan areas, the classification of land cover from satellite images has been the focus of remote sensing over the years. Due to complex architecture and a lack of labeled data, classifying buildings in metropolitan areas from very high resolution (VHR) satellite imagery is challenging. Traditional approaches for building classification include hand-crafted features and transfer learning methods. These methods often struggle with the variability in building shapes, orientation, and viewpoint, leading to low accuracy in densely populated urban areas and limited performance when dealing with high- resolution satellite images. A deep-learning based approach for semantic segmentation using U-Net with a backbone of ResNet-34 is proposed for building classification. Urban area Dataset with Images of 0.5 m resolution is prepared from SASPlanet. One hot Encoding is applied for classifying buildings. U-Net is trained with encoded data. The proposed model is evaluated on the Indian dataset, specifically, the urban areas of Nashik, Maharashtra state and the accuracy obtained for the classification dataset is 60 % and the accuracy of the building detection is about 85 %. Change detection is calculated from bi-temporal images. The GIS maps are updated to detect changes in buildings, represented by different colors to distinguish newly constructed buildings, existing structures and demolished ones.