MFC-CNN: A Deep Learning Framework for Urban Building Height Estimation From GF-7 Imagery

Estimating building heights by generating disparity maps from multi-view satellite images through stereo matching remains challenging in urban research. However, disparity maps produced by both traditional stereo matching methods and deep learning methods often contain matching errors due to insufficient texture, multi-scale complex scenes, and building occlusions, leading to underestimation of building heights. To address this issue, we proposed a new multi-scale feature connected densely network (MFC-CNN) for stereo matching to generate high-quality disparity maps from GF-7 images. Then, a spatial forward intersection algorithm was used to further generate DSM from the disparity maps. Finally, a buffer analysis method considering roof type was adopted to combine DSM with building footprints for building height extraction. The disparity maps generated by MFC-CNN on the WHU-Stereo dataset achieved EPE and D1 of 1.208 and 7.49% respectively, outperforming SGM, StereoNet, and BGA-Net models. The accuracy of DSM data generated by the spatial forward intersection algorithm also reached the best, with RMSE, ME, SD, and MAE values of 1.58, 0.11, 1.62, and 1.36 respectively. Finally, building heights in Bao'an District of Shenzhen, Jinan, Beijing, and Shanghai were extracted. In the experimental results of a total of 94,671 samples in the four cities, the MAE and RMSE were 2.86 and 3.96, and the correlation coefficient R2 between the extracted building heights and the reference heights was 0.87. These results indicate that the building height extraction results of this study are highly consistent with the reference results, and also proves that this method exhibits strong stability and transferability in different regions.