Periodic Reporting for period 1 - AI4SmartCities (Artificial Intelligence for Smart Cities)
Reporting period: 2021-01-01 to 2023-12-31
We have two journal publications currently under review. One is related to 3D reconstruction in urban areas, and the other on fine-grained land use classification in urban areas.
In smart city application, 3D building reconstruction is a frequently demanded service. Typically it is done via spaceborne or aerial optical images. In AI4SmartCities, we worked on an alternative data source: synthetic aperture radar. It has the advantage of large coverage, and less weather dependent than optical images. Typical algorithm in this field known as SAR tomography is computationally expensive, rendering them impractical for real applications. We advanced the technology readiness level of SAR tomography by improving the computational efficiency by an order of magnitude using modern deep learning techniques.
In the second work, understanding the details of land cover in urban areas is crucial for various applications like urban planning, environmental monitoring, and sustainable development. However, getting this information is tough because it requires detailed labeling of data, which is time-consuming and expensive. To tackle this challenge, we come up with a solution based on weakly supervised domain adaptation (WSDA). We combine fully labeled data from one source with partially labeled data from another source to make the process easier and faster. Relying on WSDA, we are able to map land cover in fine detail for ten cities across the globe, using satellite images from different sources. This means we can understand the different types of land cover, like buildings, roads, and green spaces, even in areas where detailed labels are scarce. Our results show that this approach can make land cover mapping more efficient and cost-effective, which is important for smart cities applications.
In smart city application, 3D building reconstruction is a frequently demanded service. Typically it is done via spaceborne or aerial optical images. In AI4SmartCities, we worked on an alternative data source: synthetic aperture radar. It has the advantage of large coverage, and less weather dependent than optical images. Typical algorithm in this field known as SAR tomography is computationally expensive, rendering them impractical for real applications. We advanced the technology readiness level of SAR tomography by improving the computational efficiency by an order of magnitude using modern deep learning techniques.
In the second work, understanding the details of land cover in urban areas is crucial for various applications like urban planning, environmental monitoring, and sustainable development. However, getting this information is tough because it requires detailed labeling of data, which is time-consuming and expensive. To tackle this challenge, we come up with a solution based on weakly supervised domain adaptation (WSDA). We combine fully labeled data from one source with partially labeled data from another source to make the process easier and faster. Relying on WSDA, we are able to map land cover in fine detail for ten cities across the globe, using satellite images from different sources. This means we can understand the different types of land cover, like buildings, roads, and green spaces, even in areas where detailed labels are scarce. Our results show that this approach can make land cover mapping more efficient and cost-effective, which is important for smart cities applications.