Road Extraction by Using Atrous Spatial Pyramid Pooling Integrated Encoder-Decoder Network and Structural Similarity Loss

• Published in: Remote sensing (Basel, Switzerland) Switzerland), 2019-05, Vol.11 (9), p.1015
• Main Authors: He, Hao, Yang, Dongfang, Wang, Shicheng, Wang, Shuyang, Li, Yongfei
• Format: Article
• Language: English
• Subjects: Architectural engineering; Coders; Datasets; Deep learning; Encoders-Decoders; Engineering; Feature extraction; Image processing; Image quality; Image segmentation; Machine learning; Methods; Neural networks; Pattern recognition; Quality; Remote sensing; road extraction; Roads; Roads & highways; semantic segmentation; Semantics; Similarity; structural similarity; Traffic management; Traffic planning; Urban planning
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs11091015

The technology used for road extraction from remote sensing images plays an important role in urban planning, traffic management, navigation, and other geographic applications. Although deep learning methods have greatly enhanced the development of road extractions in recent years, this technology is still in its infancy. Because the characteristics of road targets are complex, the accuracy of road extractions is still limited. In addition, the ambiguous prediction of semantic segmentation methods also makes the road extraction result blurry. In this study, we improved the performance of the road extraction network by integrating atrous spatial pyramid pooling (ASPP) with an Encoder-Decoder network. The proposed approach takes advantage of ASPP’s ability to extract multiscale features and the Encoder-Decoder network’s ability to extract detailed features. Therefore, it can achieve accurate and detailed road extraction results. For the first time, we utilized the structural similarity (SSIM) as a loss function for road extraction. Therefore, the ambiguous predictions in the extraction results can be removed, and the image quality of the extracted roads can be improved. The experimental results using the Massachusetts Road dataset show that our method achieves an F1-score of 83.5% and an SSIM of 0.893. Compared with the normal U-net, our method improves the F1-score by 2.6% and the SSIM by 0.18. Therefore, it is demonstrated that the proposed approach can extract roads from remote sensing images more effectively and clearly than the other compared methods.