.png){kind=small}
Estimating high-resolution building volumes from multi-source satellite data using deep models
Mr LIU Rui
Abstract:
Fine-grained urban building volumes are indispensable in various studies of urban energy consumption, thermal environment, and climate, while there is an urgent need to move beyond extracting building volumes on a local scale to a larger scale in a fine-grained manner, as cities have been expanded dramatically and the trend is likely to continue in the future. Satellites can provide diverse imagery with high spatial resolution at a low cost, with data processing being relatively easier. Moreover, it is not affected by policy and sovereignty in terms of access within different urban areas. However, when resorting to high-resolution satellite images to extract building volumes, if considered from a data perspective, unimodal data may be insufficient due to highly differentiated spectrums of homogeneous land covers and semantic confusions within heterogeneous land covers. In view of methodology, model transfer in the context of spatiotemporal differences among cities remains challenging. Optical and synthetic aperture radar (SAR) are considered complementary modalities due to different imaging mechanisms which are able to obtain different information about the land surface. Despite the fruitful fusion of optical and SAR images is of great significance to building volumes extraction, it is still where many puzzles prevail, as the adverse impacts will increase geometrically with the number of modalities. However, existing studies are either based on unimodal data or naively fusing multimodal data, the fusion strategy is far from satisfactory. Additionally, current methods for extracting building volumes are mostly based on supervised machine learning or deep learning models, which limits its application due to the limited transferability and large spatiotemporal variation among urban regions. Worse still, in the era of big earth data, this disadvantage will be infinitely magnified, as these methods are labor-intensive and time-consuming. To fill above mentioned research gaps, this study aims 1) to improve building footprint extraction by fusing optical and SAR satellite images, 2) to retrieve building height from multimodal satellite data, and 3) to develop the model scalability method to achieve large scale mapping with limited labeled samples on a large spatial scale.