Deep Learning for Cloud Gap Filling in Urban Land Surface Temperature Data
09.07.2026Satellite-derived land surface temperature (LST) provides valuable information about urban heat patterns at high spatial resolution. However, observations from thermal infrared satellite sensors are frequently covered by clouds, resulting in substantial gaps in the available temperature fields.
Recent deep learning approaches, including convolutional neural networks and diffusion models, have demonstrated strong performance in image inpainting. Applying these methods to LST is challenging because reconstructed values must preserve both spatial structures and physically meaningful temperatures.
The aim of this thesis is to implement and evaluate deep learning models for reconstructing cloud-covered regions in Landsat LST images. The models may incorporate additional geospatial information such as land cover. Their performance will be compared with conventional gap-filling approaches (e.g. Kriging) using synthetic cloud masks.
Supervisor: Daniel Weggenmann
Requirements:
Experience programming with Python and PyTorch
(Beneficial) Interest in data-driven climate science or diffusion models
