Towards explainable and physically-based deep learning statistical downscaling methods
To design strategies against climate change, it is crucial to have projections generated by climate models. Unfortunately, the coarse resolution of these models poses challenges in simulating variables at the local scale. To overcome this limitation, statistical downscaling uses statistical models t...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universidad de Cantabria (UC) |
| Repositorio: | UCrea Repositorio Abierto de la Universidad de Cantabria |
| Idioma: | inglés |
| OAI Identifier: | oai:repositorio.unican.es:10902/32096 |
| Acceso en línea: | https://hdl.handle.net/10902/32096 |
| Access Level: | acceso abierto |
| Palabra clave: | Statistical downscaling Deep learning Explainable artificial intelligence Cluster computing Downscaling estadístico Aprendizaje profundo Explicabilidad Computación |
| Sumario: | To design strategies against climate change, it is crucial to have projections generated by climate models. Unfortunately, the coarse resolution of these models poses challenges in simulating variables at the local scale. To overcome this limitation, statistical downscaling uses statistical models to understand the relationship between local and large-scale variables unaffected by this coarse resolution. Although deep learning models have emerged as a promising technique to learn this relationship, they are considered black-boxes as the learned relationships are challenging to comprehend, leading to mistrust among experts. To overcome this limitation, this Thesis explores eXplainable Artificial Intelligence techniques to assess these models in terms of the learned relationships, achieving a more robust and physically-based validation of deep learning models for statistical downscaling, as well as aligning the resultant projections with some fundamental physical constraints. In addition to the application of these diagnostics to a real use-case in Iberia in the context of the Escenario-PNACC initiative, some computational aspects involved in this process are also explored. |
|---|