Monitoring urban impervious surface using multi-source remote sensing in cloud-prone areas

Abstract

Urban impervious surface (UIS) is crucial in urban ecosystem studies, impacting hydrological processes, the urban heat island effect, and environmental quality. Accurate UIS identification and mapping are essential for effective urban planning and sustainability. However, cloud contamination has emerged as a critical impediment in optical remote sensing for timely and accurate urban monitoring in numerous emergent cloudy situations, such as rainstorms and flooding responses. This limitation necessitates the incorporation of Synthetic Aperture Radar (SAR) data, which is unaffected by clouds. However, traditional optical and SAR data fusion methods predominantly assume clear-sky conditions, and the challenge of effectively merging SAR with optical data in cloudy conditions has yet to be fully addressed. In response to these challenges, this thesis offers methods to bolster UIS identification by fusing optical and SAR data, emphasizing cloud-affected environments. It quantifies the impact of cloud cover on optical data accuracy and highlights SAR's increasingly crucial role as a supportive source when optical data's capabilities are hindered by clouds. A key contribution is the development of a SAR-based Polarimetric Scattering Mixture Analysis (PSMA) method. The research delves into the often-overlooked issue of scattering confusion, which is a critical cause of misclassification in SAR-based identification. It conducts a thorough analysis of the scattering mechanisms associated with different land covers, laying the groundwork for a novel surface classification strategy based on SAR polarimetric scattering modeling. Applied to C- and L-band full-polarization SAR data across various urban settings in China, the PSMA method achieved accuracy exceeding 96%, reducing misclassification and boosting accuracies by up to 13.4% for diverse UIS types. Given the advantages of both optical and SAR data, fusion is deemed beneficial. However, prior methods have focused on cloud-free fusion, prompting this research to explore fusion under clouded conditions. A novel Weighted Cloud Dictionary Learning (WCDL) method is introduced, integrating a cloud probability weighting model with a pixel-wise cloud dictionary learning approach to alleviate cloud interference. This technique enhances overall accuracy by 6% over single-source methods and surpasses other fusion methods. Finally, the thesis presents an advanced fusion framework named SAR-Optical Dictionary Learning (SODL), tailored for situations where optical data is severely obscured by dense cloud cover. The SODL framework establishes a joint dictionary space that leverages limited optical data to refine SAR's discriminatory power. Experiments in tropical and subtropical regions of China demonstrate SODL's robustness, achieving gains in UIS detection accuracy up to 33.81% compared to other fusion methods.