Using Neural Networks for Bayesian Precipitation Retrievals from GPM Passive Microwave Observations

Abstract:

The Goddard Profiling Algorithm (GPROF) is used to produce estimates of surface
precipitation and hydrometeor profiles from the passive microwave measurements
of the GPM constellation. The algorithm applies a Bayesian retrieval method
known as Monte Carlo integration (MCI). A major drawback of MCI is that it
requires a lookup in a large retrieval database for each retrieved footprint.
Furthermore, the method does not allow for simple integration of ancillary data or spatial
information into the retrieval. In this work, we investigate using Quantile
Regression Neural Networks, a recently-proposed machine-learning approach, as an
alternative to MCI within GPROF in order to overcome these limitations.

Quantile Regression Neural Networks (QRNNs) are a type of neural network that
can be used to solve Bayesian retrieval problems. QRNNs estimate the posterior
distribution of the retrieval quantity, which can be used to derive reliable,
case-specific uncertainty estimates. QRNNs thus overcome the lack of uncertainty
estimates of traditional neural-network techniques, which so far hindered their
application for remote sensing retrievals.

This study focuses on the retrieval of surface precipitation from passive
microwave observations from the GPM constellation. Comparison of the results
obtained using QRNN and MCI shows that QRNNs provide similar accuracy as MCI for
point estimates but yield better estimates of the retrieval uncertainty and
reduce processing times by an order of magnitude. Furthermore, we find that the
predicted posterior distribution can be used to accurately classify raining and
non-raining pixels.

Our results show that QRNNs can be used as a drop-in replacement for MCI in
surface-precipitation retrievals, providing better estimates of the posterior
distribution, reducing processing time and offering greater flexibility for
introducing ancillary data into the retrieval. Although the application
presented here is limited to the GPROF algorithm, the generality of the
presented method makes it suitable for a wide range of other retrieval problems.