A Machine Learning Method for Building Height Estimation Based on Sentinel-2 Bi-Temporal Images

Abstract

Building height information is essential for many applications such as urban planning and population density estimation. The building shadow length varies according to seasons, which is shown as different digital number values in multi-temporal images. Thus, the bi-temporal satellite remote sensing images of Sentinel-2 are used to estimate the buildings height in this study. An area of 15 km × 15 km in Beijing, China is taken as the study area. By preprocessing the data, the remaining pixels are split into two parts: 70% as the training data set and the rest as the testing data set. Then, one classification model and three regression models are proposed with using Random Forest (RF) method. Based on the testing data, it shows that the accuracy rate of the classification model has reached 98.4% with the kappa coefficient of 0.93. And the regression models' root-mean-square error (RMSE) is 0.61 floor for 1–6 floors group, 0.41 floor for 7–12 floor group, and 0.98 floor for above 12 floor group. The final RMSE is 1.62 floor with RF models. In general, this study shows the feasibility of using satellite mid-resolution optical image to estimate the building height and provides an important reference for regional building height estimation in the future.