Automated Building Extraction from Aerial Imagery with Mask R-CNN

Zilong Yang, Purdue University

Abstract

Buildings are one of the fundamental sources of geospatial information for urban planning, population estimation, and infrastructure management. Although building extraction research has gained considerable progress through neural network methods, the labeling of training data still requires manual operations which are time-consuming and labor-intensive. Aiming to improve this process, this thesis developed an automated building extraction method based on the boundary following technique and the Mask Regional Convolutional Neural Network (Mask R-CNN) model. First, assisted by known building footprints, a boundary following method was used to automatically best label the training image datasets. In the next step, the Mask R-CNN model was trained with the labeling results and then applied to building extraction. Experiments with datasets of urban areas of Bloomington and Indianapolis with 2016 high resolution aerial images verified the effectiveness of the proposed approach. With the help of existing building footprints, the automatic labeling process took only five seconds for a 500*500 pixel image without human interaction. A 0.951 intersection over union (IoU) between the labeled mask and the ground truth was achieved due to the high quality of the automatic labeling step. In the training process, the Resnet50 network and the feature pyramid network (FPN) were adopted for feature extraction. The region proposal network (RPN) then was trained end-to-end to create region proposals. The performance of the proposed approach was evaluated in terms of building detection and mask segmentation in the two datasets. The building detection results of 40 test tiles respectively in Bloomington and Indianapolis showed that the Mask R-CNN model achieved 0.951 and 0.968 F1-scores. In addition, 84.2% of the newly built buildings in the Indianapolis dataset were successfully detected. According to the segmentation results on these two datasets, the Mask R-CNN model achieved the mean pixel accuracy (MPA) of 92% and 88%, respectively for Bloomington and Indianapolis. It was found that the performance of the mask segmentation and contour extraction became less satisfactory as the building shapes and roofs became more complex. It is expected that the method developed in this thesis can be adapted for large-scale use under varying urban setups.

Degree

M.Sc.

Advisors

Shan, Purdue University.

Subject Area

Artificial intelligence|Computer science|Pedagogy

COinS