High-resolution Digital Terrain Modelling for Urban Flood Mapping with Deep Learning

Abstract

The rising demand for high-resolution (HR) terrain modelling stems from the growing need for precise 3D geospatial data across various sectors where accurate models of the Earth's surface are vital, such as flood monitoring, environmental management, urban planning, and disaster response. The advancements of remote sensing technologies, especially the Light Detection and Ranging (LiDAR) and satellite and aerial imaging, offer the capability of producing an HR digital terrain model (DTM) with unprecedented precision to support those increasingly complex analytical and operational tasks, especially in dense urban environments. This thesis tackles three significant challenges in HR terrain modelling for urban flood mapping: 1) lack of geospatial datasets for terrain modelling, 2) underutilization of DTMs in near-real-time urban flood mapping, and 3) limited availability of HR DTM. First, an ultra-large-scale airborne LiDAR point cloud dataset for ground filtering (GF) named OpenGF was built upon open-access airborne laser scanning (ALS) data from four different countries around the world. This dataset covers nine different terrain scenes with three challenging complex test sets to encourage the development of 3D deep neural networks for HR digital terrain modelling. A series of experiments were conducted on OpenGF to evaluate the performance of deep learning (DL) networks on GF. Second, a novel near-real-time multi-sensory HR urban flood mapping framework was proposed. This method features a DTM upscaling method that produces HR DTM from low-resolution (LR) DTM with a fusion approach to reconstruct urban terrain details from HR optical imagery to support urban flood mapping. Meanwhile, a near-real-time visible flood water extraction and a Geographical Information System (GIS) tool were introduced to complete the urban flood mapping workflow. Third, an advanced image-guided DTM upscaling DL network was proposed to produce HR DTM from LR DTM and HR optical imagery with multi-task learning. This network simultaneously performs guided DTM upscaling and semantic segmentation of the HR optical imagery in urban environments. The network comprises an HR image guidance subnetwork that extracts high-level semantic features, and a DTM recovery subnetwork that enhances elevation details through multimodal feature fusion. This thesis also discusses the limitations of the proposed method and provides insights for future research on HR terrain mapping for environmental monitoring.