Using a 10-Year Radar Archive for Nowcasting Precipitation Growth and Decay: A Probabilistic Machine Learning Approach

Abstract

Machine learning algorithms are trained on a 10-yr archive of composite weather radar images in the Swiss Alps to nowcast precipitation growth and decay in the next few hours in moving coordinates (Lagrangian frame). The hypothesis of this study is that growth and decay is more predictable in mountainous regions, which represent a potential source of practical predictability by machine learning methods. In this paper, artificial neural networks (ANN) are employed to learn the complex nonlinear dependence relating the growth and decay to the input predictors, which are geographical location, mesoscale motion vectors, freezing level height, and time of the day. The average long-term growth and decay patterns are effectively reproduced by the ANN, which allows exploring their climatology for any combination of predictors. Due to the low intrinsic predictability of growth and decay, its prediction in real time is more challenging, but is substantially improved when adding persistence information to the predictors, more precisely the growth and decay and precipitation intensity in the immediate past. The improvement is considerable in mountainous regions, where, depending on flow direction, the root-mean-square error of ANN predictions can be 20%–30% lower compared with persistence. Because large uncertainty is associated with precipitation forecasting, deterministic machine learning predictions should be coupled with a model for the predictive uncertainty. Therefore, we consider a probabilistic perspective by estimating prediction intervals based on a combination of quantile decision trees and ANNs. The probabilistic framework is an attempt to address the problem of conditional bias, which often characterizes deterministic machine learning predictions obtained by error minimization.

Pysteps: an open-source Python library for probabilistic precipitation nowcasting (v1.0)

“…Individual pysteps modules can also serve different purposes. For example, the optical flow modules can be used to study precipitation growth and decay in moving coordinates (e.g., Foresti et al, 2018Foresti et al, , 2019Zeder et al, 2018), to correct radar field accumulations accounting for advection (e.g., Wang et al, 2015;Lukach et al, 2017), to synchronize the individual radar elevation scans (e.g., Tabary, 2007) or to separate the location error of NWP precipitation forecasts (Marzban and Sandgathe, 2010).…”

“…Despite recent advances in numerical weather prediction (e.g., Sun et al, 2014), extrapolationbased nowcasting remains the primary approach at such space-and timescales, typically outperforming NWP forecasts in the first 2-5 h, depending on the weather situation, domain and NWP characteristics (e.g., Berenguer et al, 2012;Mandapaka et al, 2012;Simonin et al, 2017;Jacques et al, 2018). Other recent developments include machine learning methods, for which promising results have been obtained (e.g., Xingjian et al, 2015;Foresti et al, 2019), but these have not so far been deployed in operational nowcasting systems.…”

Precipitation Nowcasting with Orographic Enhanced Stacked Generalization: Improving Deep Learning Predictions on Extreme Events

“…Given the complexity of the Alpine environment in the area covered by the TAASRAD19 dataset and the direct known relationships between convective precipitation and the underlying orographical characteristics [9,[41][42][43][44], we add to the stack of the input images three layers of information, derived from the orography of the area: the elevation, the degree of orientation (aspect), and the slope percentage. The three features are computed by resampling the digital terrain model [45] of the area at the spatial resolution of the radar grid (500 m), and computing the relevant features in a GIS suite [46].…”

“…Sequences of reflectivity maps are used as input for prediction models. More formally, given a reflectivity field at time T 0 , radar-based nowcasting methods aim to extrapolate m future time steps T 1 , T 2 , ..., T m in the sequence, using as input the current and n previous observations T −n , ..., T −1 , T 0 .Traditional nowcasting models are manly based on Lagrangian echo extrapolation [7,8], with recent modification that try to infer precipitation growth and decay [9,10] or integrate with Numerical Weather Predictions to extend the time horizon of the prediction [11,12]. In the last few years, Deep Learning (DL) models based on combination of Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN) have shown substantial improvement over nowcasting methods based on Lagrangian extrapolations for quantitative precipitation forecasting (QPF) [13].…”

“…In fact, the main challenge faced by nowcasting methods is the progressive accumulation of uncertainty: DL architectures deal with uncertainty by smoothing prediction over time, using the intrinsic averaging effect of loss functions such as Mean Squared Error (MSE) and Mean Absolute Error (MAE), commonly used as loss functions to train DL architectures in regression problems [22]. This smoothing problem can be seen as Conditional Bias (CB): the minimization of MSE leads to models where peak values are systematically underestimated and compensated by overestimation in weak rain-rates [9,23]. Moreover, the minimization of these two errors is at odds [24]: measures taken to remove CB lead to an increase in MSE, and vice versa, the minimization of MSE results in a higher CB, manifested in an underestimation of high and extreme rain rates.While not addressing the problem directly, some DL approaches try to cope with CB by introducing weighted loss functions [15], by integrating loss functions used in computer vision [25], or by optimizing for specific rain regimes [26].…”