Using machine learning to model uncertainty for water vapor atmospheric motion vectors

Abstract

Abstract. Wind-tracking algorithms produce atmospheric motion vectors (AMVs) by tracking clouds or water vapor across spatial–temporal fields. Thorough error characterization of wind-tracking algorithms is critical in properly assimilating AMVs into weather forecast models and climate reanalysis datasets. Uncertainty modeling should yield estimates of two key quantities of interest: bias, the systematic difference between a measurement and the true value, and standard error, a measure of variability of the measurement. The current process of specification of the errors in inverse modeling is often cursory and commonly consists of a mixture of model fidelity, expert knowledge, and need for expediency. The method presented in this paper supplements existing approaches to error specification by providing an error characterization module that is purely data-driven. Our proposed error characterization method combines the flexibility of machine learning (random forest) with the robust error estimates of unsupervised parametric clustering (using a Gaussian mixture model). Traditional techniques for uncertainty modeling through machine learning have focused on characterizing bias but often struggle when estimating standard error. In contrast, model-based approaches such as k-means or Gaussian mixture modeling can provide reasonable estimates of both bias and standard error, but they are often limited in complexity due to reliance on linear or Gaussian assumptions. In this paper, a methodology is developed and applied to characterize error in tracked wind using a high-resolution global model simulation, and it is shown to provide accurate and useful error features of the tracked wind.

Prediction of Atmospheric Noise Temperature at the Deep Space Network With Machine Learning

“…Real-time error characterization is provided by a UQ methodology similar to that in Teixeira et al (2021), wherein a clustering algorithm is used to identify regimes of prediction error. Teixeira et al (2021) showed that error characterization of a geophysical prediction of atmospheric motion vectors benefitted from separating those predictions into different geophysical regimes because the error characteristics of the predictions can vary significantly between the regimes.…”

“…Real-time error characterization is provided by a UQ methodology similar to that in Teixeira et al (2021), wherein a clustering algorithm is used to identify regimes of prediction error. Teixeira et al (2021) showed that error characterization of a geophysical prediction of atmospheric motion vectors benefitted from separating those predictions into different geophysical regimes because the error characteristics of the predictions can vary significantly between the regimes. The cursory analysis (Figure 5) shows that the bias and RMSE of the RF prediction is significantly greater for predictions of T atm > 16 K. We enhance upon this simple observation by analyzing the relationship between the predictor variables in the ML forecast and the ML-forecast results.…”

“…As in Teixeira et al (2021), we validated our trained model using the Continuous Rank Probability Score (CRPS), a scoring rule for probabilistic forecasts. CRPS is as a function of a cumulative distribution function F and an observation x as follows:…”

“…Real‐time error characterization is provided by a UQ methodology similar to that in Teixeira et al. (2021), wherein a clustering algorithm is used to identify regimes of prediction error. Teixeira et al.…”

Uncertainty of Atmospheric Motion Vectors by Sampling Tempered Posterior Distributions

“…These indicators are based on comparing changes in AMV estimates between sequential time steps and neighboring pixels, as well as differences with model predictions. Other approaches build statistical model using linear regression against radiosonde values to correct AMV observation error [25] or rely on machine learning techniques and training data generated by independent NWP simulations [43]. On the other hand, very few optic-flow methods propose uncertainty estimates in the computer vision literature.…”