Retrieving precipitation over Brazil. A quantile regression neural networks approach

Abstract

Close and accurate monitoring of precipitation on a global scale is key to understanding our future climate as well as our current weather. Geostationary weather satellites, as opposed to other measuring methods, provide high-resolution information covering large regions. The sensors carried can not, however, measure precipitation directly but are restricted to capturing cloud top temperatures (in IR radiation). Earlier work presents a range of models that aim to relate these geostationary observations to precipitation, including simple regression as well as more elaborate machine learning techniques. In this thesis we aim at predicting a posterior distribution instead of a single precipitation value for each set of cloud top temperatures. To achieve this, we make use of Quantile Regression Neural Networks (QRNNs), a supervised machine learning approach. The two main questions asked are as follows: Can this deep learning method be used to improve upon algorithms currently in operation? and Can spatial information be used to improve the retrieval? The models are trained on GOES-16 IR data over Brazil with a precipitation product from the GPM Core Observatory. Our results on held-out test data show that it is possible to model the precipitation distribution using a QRNN. Additionally, a 20% decrease in mean squared error and a 25% decrease in mean absolute error is observed on the test data when using the spatially aware model, which illustrates the general performance improvement by utilizing the spatial information. The QRNN models also show promising results on an independent rain gauge dataset where they are compared against the currently-in-operation Hydro-Estimator. Here our most promising QRNN shows a 30% decrease in mean squared error compared to the present model.