Remote Sensing Based Building Height Regression

This paper presents a method for predicting building heights from monocular satellite images. Detecting building contours and heights from single images is a challenging task. We propose a method that utilizes appearance features from top perspective satellite images to detect building heights. We employ rotated object detection techniques to identify building floor contours, which serve as the basis for height detection. By using a soft attention-based CNN regression approach, we evaluate the effectiveness of building height regression using various performance metrics and discuss their strengths and weaknesses. Furthermore, we investigate the impact of augmentation techniques, specifically Cutmix, on the performance of these methods. Our results indicate that we can detect up to 70% of all buildings correctly in different cities and accurately estimate their heights with a 7 meter RMSE score. Additionally, we find that Cutmix augmentation significantly improves the performance, particularly when there are errors and shifts in the annotated images.