Cette publication est également disponible en :
Français
Thèse soutenue par Luiz Gustavo RASERA le 12 juin 2020, Institut des dynamiques de la surface terrestre (IDYST)
Spaceborne remote sensing has enabled near-global mapping of the Earth’s topography. However, satellite-derived digital elevation models (DEMs) are unsuited for modeling fine-scale Earth surface processes due to their limited spatial resolution.
To this day, fine-resolution DEMs remain sparsely distributed across the planet owing to the technical challenges and substantial costs for producing densely sampled data sets. Over the last decade, multispectral satellite imagery (MSI) has become widely available, providing abundant fine-resolution data for monitoring the Earth’s surface. Although rendering no elevation information, MSI has the potential to provide indirect fine-scale information about topography. Statistical downscaling enables prediction of attributes at scales finer than that of the input data. Multiple-point statistics (MPS) simulation is a powerful alternative for stochastic downscaling due to its ability to replicate complex spatial patterns and assess the uncertainty of the predictions.
Conceptually, MPS simulation methods could be employed for downscaling of coarse-resolution DEMs by extracting spatial information from available fine-resolution DEMs and MSI of better-measured data sets. The application of MPS simulation for downscaling of DEMs is compelling, but there are many issues to be addressed. Trends in elevation pose a challenge for stochastic downscaling of mountainous terrain. MPS simulation algorithms are also notably difficult to parameterize, often requiring manual parameter calibration. As a result, the integration of disparate data sources, such as DEMs and MSI, into the downscaling becomes a daunting task.
Addressing these challenges requires the development of an automated data integration approach. In this thesis, a MPS-driven data integration framework for stochastic downscaling of coarse-resolution DEMs is developed. The approach is composed of algorithms designed for three primary tasks: the statistical downscaling of data sets with trends, the automation of the downscaling process, and the integration of secondary data into the downscaling.
The first contribution of this thesis is a novel MPS-driven downscaling algorithm with inbuilt capabilities for handling data sets with trends. Terrain elevation is modeled as a spatial signal expressed as the sum of a deterministic trend and a stochastic residual component. The approach enables accurate downscaling of coarse-resolution DEMs of either flat or steep terrain.
The second contribution addresses the parametrization of the MPS-driven downscaling algorithm. An automation routine is used to infer optimal algorithm parameters by framing the parameter calibration task as an optimization problem. The framework provides an efficient alternative for automatic generation of statistically accurate fine-resolution DEMs.
The third contribution builds upon the two aforementioned developments by integrating finer-resolution MSI-sourced data as secondary information into the downscaling process. Elevation and MSI data with varying spatial resolutions are integrated based on a probabilistic framework. The approach enables to enhance the structural accuracy of the fine-resolution simulated DEMs and to reduce the inherent uncertainty associated with the downscaling.
Developments in this thesis provide an efficient, low-cost alternative for downscaling of coarse-resolution DEMs based on the integration of available finer-resolution terrain and imagery data. Future research should focus on evaluating potential applications of the downscaled DEMs for the study of Earth surface processes, the planning and design of infrastructures, and the risk assessment of natural hazards.
