|
Wind-tracking algorithms produce Atmospheric Motion Vectors (AMVs) by tracking water vapor across spatial-temporal fields. Thorough error characterization of wind-track algorithms, otherwise known as uncertainty quantification, is critical in properly assimilating their produced AMVs into forecast models. Uncertainty quantification has two key quantities of interest: accuracy— the systematic difference between a measurement and the true value, and precision— a measure of variability of the measurement. Traditional techniques for uncertainty quantification through machine learning have focused on characterizing accuracy but often struggle when estimating precision. By pairing a random forest algorithm with unsupervised parametric clustering (using a Gaussian Mixture Model), we propose a machine learning based method of building uncertainty models characterizing both accuracy and precision using limited experimental data. In particular, we develop a Gaussian Mixture Model to cluster the principle quantities of interest in our training dataset— water vapor, measured AMVs, and true wind speed— into discrete regimes each with a distinct precision and accuracy. Concurrently, we train a random forest to predict true wind speed given the outputs of a wind-tracking algorithm, which works to model some of the extreme error in the algorithm. Combining these, we build a model which can place a retrieved AMV into a distinct regime with a characterized accuracy and precision.
|
