Building Height Prediction with Instance Segmentation

Estimating the height of buildings only based on monocular satellite images is a difficult task due to limited annotated data quantity, subpar data quality, changes in building kinds, and varying angles of satellite perspective, light and shadow. Given the task of predicting a single height value for each building in a satellite image, this work proposes a method for instance segmentation based building height estimation. In a single end to end multitask model, the system predicts contours and corresponding heights of buildings from a monocular remote sensing image. Using a dataset curated for this study, we present initial results on building height classification with three classes. Additionally we use transfer learning from a public building contour dataset to improve results. Overall, an average classification accuracy of 70% is achieved for buildings in each standalone height class.