Two-phase methods to segment man-made objects around reservoirs

• Ano de defesa: 2023
• Autor(a) principal: Hamidishad, Nayereh
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizagem profunda; Deep learning; High-resolution RS imagery; Image processing; Imagem RS de alta resolução; Man-made object segmentation; Object-based method; Processamento de imagem; Reservoir segmentation; Segmentação de reservatórios; Segmentação feita pelo homem
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-05062023-071021/

Reservoirs are fundamental infrastructures for the management of water resources. They reduce the effects of interseasonal and interannual streamflow fluctuations and hence facilitate water supply, hydroelectric power generation, and flood control, to name a few. There is a significant interaction between the environment and reservoirs. For example, reservoirs affect the quality of the water downstream of their dams, and human activities affect the quality of the reservoirs inflowing water and its chemical and biological processes. Construction around reservoirs is a human activity that can negatively impact the reservoirs water quality. This social issue can be detected by segmenting the man-made objects around reservoirs in the Remote Sensing (RS) images. Traditional pixel-based, Object-Based (OB), and Deep-Learning (DL) methods are three Land-Cover Mapping (LCM) approaches. We developed a new approach based on image processing techniques and the OB method for LCM of the selected regions around reservoirs. Disadvantages of the OB approach, such as the high dependency of results on the choice of parameters, led us to use DL to circumvent excessive parameter specification and tunning that are often required by OB methods. In recent years, DL has attracted considerable attention as a method for segmenting the RS imagery semantically and has achieved remarkable success. To segment man-made objects around the reservoirs utilizing an end-to-end workflow, segmenting reservoirs and detaching the Region of Interest (RoI) around them are essential. However, reservoirs are always considered in a broad class termed water bodies in RS semantic segmentation studies. Besides, man-made object semantic segmentation in the RoIaR is not explored in the literature. Moreover, man-made object segmentation in high-resolution images, especially countryside man-made object segmentation, is not extensively explored in the literature. In this research, we develop a new approach based on DL and image processing techniques for man-made object segmentation around the reservoirs. In the proposed two-phase workflow, the reservoir is initially segmented using a DL model. Then, a post-processing stage is proposed to remove errors such as floating vegetation. Next, the RoI around the Reservoir (RoIaR) is detached using the proposed image processing techniques. Finally, the man-made objects in the RoIaR are segmented by a DL model. We collected high-resolution Google Earth (GE) images of eight reservoirs in Brazil, mainly located in the countrysides, over two different available years to train the workflow models. Furthermore, we validated the prepared workflow with a test dataset not seen during training. The F1-scores of the phase-1 semantic segmentation stage, post-processing stage, and phase-2 semantic segmentation stage on the external test set are 92.54%, 94.68%, and 88.11%, respectively, which show high generalization ability of the prepared workflow.