Abstract

Building footprint is a primary dataset of an urban geographic information system (GIS) database. Therefore, it is essential to establish a robust and automated framework for large-scale building extraction. However, the characteristic of remote sensing images complicates the application of the instance segmentation method based on the Mask R-CNN model, which ought to be improved toward extracting and fusing multi-scale features. Moreover, open-source satellite image datasets with wider spatial coverage and temporal resolution than high-resolution images may exhibit different coloration and resolution. This study proposes a large-scale building extraction framework based on super-resolution (SR) and instance segmentation using a relatively lower-resolution (>0.6 m) open-sourced dataset. The framework comprises four steps: color normalization and image super-resolution, scene classification, building extraction, and scene mosaicking. We took Hyogo Prefecture, Japan (19,187 km2) as a test area and extracted 1,726,006 (29.12 km2) of the 3,301,488 buildings (32.46 km2), where the number of buildings and footprint area increased by 3.0 % and 5.0 % respectively. The result indicated that the color normalization and image super-resolution could improve the visual quality of open-source satellite images and contribute to building extraction accuracy. Moreover, the improved Mask R-CNN based on Multi-Path Vision Transformer (MPViT) backbone achieved F1 scores of 0.71, 0.70, 0.81, and 0.67 for non-built-up, rural, suburban, and urban areas, respectively, which is better than those of the baseline model and other mainstream instance segmentation approaches. This study demonstrates the potential of acquiring acceptable building footprint maps from open-source satellite images, which has significant practical implications.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call