Four seasonal composite Sentinel-2 images for the large-scale estimation of the number of stories in each individual building

Knowledge of the number of building stories (NoS) is critical for understanding and regulating the urban development process. Existing approaches often transform building heights into numbers of stories using a specific empirical story -height coefficient, e.g., 3 meters for 1 story. However, the story heights of different buildings might differ for various reasons, such as different functional types within a city, differences in urban planning regulations among cities, or the regulations in different construction years. Based on a theoretical analysis and empirical statistics regarding the changes in vertical building information in seasonal composite images, we present a novel method for directly estimating the NoS in individual buildings from optical images. Specifically, four seasonal composite Sentinel -2 images taken within a year are utilized to estimate the NoS of each building with a modified object -detection network to make full use of vertical building information. The proposed method is called the Stories number EstimAtion from Seasonal composite images with an Object detection Network (SEASONet) method. Both theoretical analysis and empirical statistics are used to determine why seasonal composite optical images can effectively provide vertical building information. To validate the performance of the proposed method, we collect data from 61 Chinese cities with various building types, train the model with data from 47 cities (1365998 buildings) and quantitatively test the model using data from the remaining 14 cities (246 584 buildings). In addition, M3Net using ZY-3 multiview images for the pixellevel estimation of building heights is adapted for comparison. The experimental results show that SEASONet achieves lower mean absolute error (MAE) and root mean square error (RMSE) values than M3Net over all 14 cities used for testing. Ablation experiments show that the four seasonal composite images are the keys for improving the estimation of the number of stories in high-rise buildings. A comparison with the results of two state-of-the-art methods that use empirical coefficients to convert building height to story number further confirms the superiority of the proposed method, especially its effectiveness in estimating the number of stories in high-rise buildings.