Development of a Machine Learning Model for Partial Column Ice Water Path and Water Vapor Retrieval

Abstract:

Ice water path (IWP) and total columnar water vapor (TWV) are important factors contributing to Earth’s radiation and hydrological budgets. However, it is difficult to derive these two parameters in an accurate and efficient way from passive microwave sensors. It is even harder to determine their vertical variabilities. These complications can primarily be attributed to the microphysical properties of frozen particles and the induced radiative complexities during the retrievals.

In this work, we developed IWP and TWV retrieval algorithms based on machine learning/artificial intelligence (ML/AI) methodology. We tested the algorithms to retrieve IWP and TWV from Megha-Tropiques SAPHIR, a cross-track microwave sounding instrument with 6 channels centered around the 183.31 GHz water vapor absorption line. This expands on several prior works that limited the model output to IWP (e.g., Piyush et al. 2019, Brath et al. 2016). Collocated CloudSat radar retrieved ice water content (IWC) and ECMWF-AUX water vapor profiles are used as the reference to select and tune the models on SAPHIR data. Several candidate machine learning methods were considered, including random forest (RF), gradient boosting regression (GBR), and multilayer perceptron (MLP). The models are trained using two years of collocated data and validated against a third year of independent data. Inputs included measurements from passive sensors such as surface temperature, SAPHIR incidence angle, and the six SAPHIR channels. GBR outperformed the others based on comparisons between IWP/WV predicted by the models and those retrieved by CloudSat. Since the six channels have different penetration depths and the values from each channel are sensitive to altitude, the models are also trained to predict the vertical profiles of IWP and TWV. With the GBR models that predict IWP/WV for different altitudes, consistent agreements can be achieved between SAPHIR retrievals and CloudSat “truth”. Therefore we have demonstrated that it is feasible to extract or predict embedded information based on data from passive sensors using a machine learning approach that is difficult to obtain via physical based algorithms.