Abschlussarbeit

Beschreibung:

Complete, high resolution, climate data is necessary for climate models to analyze meteorological phenomena and predict trends and disasters. To improve datasets with missing values, researchers have started to apply image inpainting techniques. This investigation focuses on precipitation data that is often naturally non-linear and highly localized. For such use cases, denoising diffusion probabilistic models (DDPMs) are introduced as a promising, generative & non-deterministic inpainting technique. An experiment setup to compare DDPMs in relation to other models is proposed, alongside modifications to the DDPMs that will be tested to showcase and analyze their performance in inpainting missing climate data.