Integration of satellite-based passive microwave brightness temperature observations and an ensemble-based land data assimilation framework to improve snow estimation in forested regions

摘要

The utilization of a machine learning algorithm-based (i.e., support vector machine [SVM]) model as the observation operator within a one-dimensional ensemble Kalman filter (EnKF) framework for the purpose of improving snow estimates across regional-scales is explored. The multifrequency, multipolarization framework employs an SVM to predict brightness temperature spectral difference (i.e., ΔTb) between 10.65 GHz, 18.7 GHz and 36.5 GHz as a function of land surface model state information. The EnKF then merges the predictions with observations obtained from the Advanced Microwave Scanning Radiometer (AMSR-E) sensor onboard the Aqua satellite. Case studies are presented for Quebec and Newfoundland, Canada, and several pixels in North America covered with evergreen needle-leaved forest colocated with taiga snow cover type. Model results with and without assimilation were compared against in-situ snow observations as well as state-of-the-art snow products. It is shown that using an atmospheric-forest-decoupling procedure prior to SVM training and prediction activities is useful in enhancing snow characterization. However, without adequate ground-based snow observations, it is still relatively difficult to draw a full conclusion as to the feasibility of the proposed assimilation framework employing the two-step decoupling procedure.