A Bayesian machine learning approach to geostationary infrared precipitation retrievals

Abstract

This project uses geostationary satellite data to retrieve precipitation rates at surface level. It is achieved through the use of quantile regression neural networks (QRNN) calibrated against rain rates from the Global Precipitation Measurement (GPM) Core Observatory satellite. The area of exploration is located over the Amazon rainforest. The main difficulty of this problem is that geostationary data is not directly related to rain as it only perceives the cloud top temperatures. It does, however, have a high temporal and spatial resolution which makes it interesting for applications in remote areas of the Earth where groundbased radar equipment is unavailable. The result of the project is mainly a comparison between different neural network architectures such as multi-layer perceptron (MLP) and convolutional neural networks (CNN), but there is also a minor comparison to an adapted version of a Hydroestimator (HE) that is currently in use by the National Institute for Space Research (INPE) in Brazil. The best performing configuration, with regards to the loss function, in this study was a CNN. It performed significantly better than the adapted HE for a test conducted over two days in March. An unsuccessful attempt to improve the results using time-series was also conducted. Furthermore, a U-net architecture was also tested on rain rate data that has been resolution-enhanced through interpolation.