Precipitation prediction over High Mountain Asia using Gaussian processes

The rivers of High Mountain Asia provide freshwater to over 1.9 billion people. However, precipitation, the main driver of river flow, is still poorly understood due to limited direct measurements in this area. Existing tools to interpolate the few available measurements or to downscale and bias-correct precipitation outputs from models have several limitations, especially over complex terrain. Future precipitation is even more challenging to predict. Improving predictions will enhance hydrological modelling, strengthen water resource management, and ultimately help communities adapt to the challenges of climate change. This thesis discusses the application of probabilistic machine learning to better understand and predict precipitation in this area. Probabilistic methods quantify our knowledge of precipitation and improve decision-making under uncertainty. This work centres around one such method, Gaussian processes. First, Gaussian processes are applied to downscale ERA5 precipitation over ungauged areas. They are used to auto-regressively combine precipitation datasets with different fidelities, i.e. resolutions and accuracies. This probabilistic multi-fidelity approach differs from previous machine learning methods for downscaling. It simultaneously estimates uncertainty distributions, captures precipitation structure and extremes, predicts at arbitrary locations, and overcomes gridding biases, all while working well with sparse data. While the first contribution addresses spatial extrapolation, the second examines temporal prediction. Here, Gaussian processes are applied to predict precipitation over a 15-year horizon from large-scale atmospheric circulation patterns. More specifically, a perfect prognosis approach is applied to predict monthly ERA5 reanalysis precipitation. Model design, including learning the non-stationary spatial distribution of precipitation, is explored. The study serves as a proof-of-concept for using features better predicted by global climate models. This approach, in turn, could be tested in a setup where the driving large-scale variables from reanalysis are replaced by the outputs of global climate models. This framework could therefore bypass the use of regional climate models to predict local precipitation. In a third study, the inaccuracy of regional climate models is tackled from a different perspective. Gaussian processes are trained as statistical surrogates of regional climate model outputs. Then, using an ensemble learning approach, a mixture-of-experts model combines the surrogates while leveraging the regional climate models' spatiotemporal biases with respect to gridded historical observations. This method improves on the traditional approach of taking an unweighted average of climate model outputs while addressing outstanding issues in multi-model ensemble research, such as accounting for model design proximity, ensemble outliers, and an infinite number of possible ensemble members. Lastly, a guideline for applying Gaussian process regression to real-world data is proposed. The framework focuses on model scalability and design, offering practical suggestions on when and how to use Gaussian processes. This framework is relevant as large high-dimensional datasets become ubiquitous, but resources for applying Gaussian processes in these settings are uncommon. This framework is illustrated with a case study of glacier elevation change over Greenland. Together, this body of work represents an improved understanding of precipitation in an understudied region and a stepping stone in applying probabilistic methods to real-world problems.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Tazi, Kenza

Advisor dc:contributor.advisor

• Turner, Richard E

• Bayesian machine learning
• Climate
• Gaussian process
• High Mountain Asia
• Hindu Kush Karakoram Himalaya
• Machine learning
• Precipitation
• Probabilistic machine learning

Licence

• https://www.repository.cam.ac.uk/bitstreams/c39198c1-a06a-4d12-be06-cccd759e26e2/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

Author Identifier

0000-0002-8169-6673

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/393893