The simplest approach to generate precipitation “forecasts” is to deploy the Eulerian persistence. For this, the most recent available observation, usually a radar composite, is used and then replicated several times for the future steps. This approach is quite accurate for very short lead times but obviously fails to provide meaningful forecasts in a quickly evolving system for timescales beyond several minutes such as the atmosphere. Thus, the related forecast quality can be considered as the minimum level for a prediction model to be useful.

Conventional precipitation nowcasting systems typically use a Lagrangian framework to predict the development of precipitation patterns. Although this framework often assumes the persistence of the precipitation features' intensity and displacement, it is still capable of outperforming mesoscale NWP models in precipitation nowcasting . The Lagrangian method applies a two-step approach where the precipitation features are first tracked and then extrapolated to future time steps . Typically, the tracking step is accomplished with the help of optical flow methods that infer the motion of patterns from consecutive images. For precipitation nowcasting, radar composite images are subject to a tracking algorithm such as cross-correlation tracking or centroid tracking techniques . The tracked objects are then applied to different extrapolation schemes, e.g., image warping , constant-vector advection , or semi-Lagrangian schemes . With this two-step approach, several operational precipitation nowcasting systems have been established over the globe in the last 3 decades, such as the Thunderstorm Identification Tracking Nowcasting TITAN;, the Storm Cell Identification and Tracking SCIT;, and the Short‐Term Ensemble Prediction System STEPS; seefor a review on operational systems.

Recently, implemented a set of advanced optical flow models into an open-source Python library called rainymotion. Two different groups of methods are part of this library from which we select the DenseRotation model that performs best in their study. The tracking algorithm of this model is based on the dense inverse search algorithm proposed in providing an estimate of the motion of each pixel based on two consecutive radar images. The extrapolation is then performed with a semi-Lagrangian advection scheme capable of representing rotational motions.

In our study, the Eulerian persistence model and the DenseRotation model are used to show how well the traditional methods can perform for the precipitation nowcasting task and how much benefit can be further obtained by using DL-based video prediction methods.

As already mentioned, the application of deep learning techniques in the meteorological community has gained momentum over the recent years. In particular, several studies have started to explore these techniques to tackle the precipitation nowcasting problem. By formulating precipitation nowcasting as a sequence prediction task, proposed a network of ConvLSTM cells which apply a convolution in the recurrent layers of the vanilla LSTM to capture spatiotemporal features in the underlying data. Their two-layer ConvLSTM network was able to outperform the Variational Methods for Echoes of Radar (ROVER) by , an operational precipitation nowcasting system based on optical flow methods with a semi-Lagrangian advection scheme. further extended the recurrent cells of gated recurrent units (GRUs) with non-local neural connections and proposed the Trajectory GRU (TrajGRU) model which enables the learning of location-variant structures of precipitation.

Besides, advanced the application of ConvLSTM networks and proposed the predictive recurrent neural network (PredRNN). They deployed a stack of recurrent layers that feature a “zigzag memory flow” and involve an explicit spatiotemporal memory state. In this way, they enable an explicit communication of abstracted spatiotemporal features between different levels of the recurrent network which yields improved precipitation predictions. While this approach already provided promising results, the PredRNN model was updated to PredRNN-v2 . The updates comprise the implementation of a “decoupling loss”, named ST-LSTM, to enhance the featuring of the spatiotemporal variations and a new, improved long-term modeling strategy. The model attains remarkable improvements when applied to multiple datasets including radar observations.

Meanwhile, other network architectures were explored in the scope of precipitation nowcasting. One of them is the fully convolutional U-Net architecture, which is a u-shaped hierarchical encoder–decoder network with skip connections. The architecture enables the abstraction of features on different spatial scales. Notably, the RainNet architecture proposed by proved to significantly outperform optical-flow-based nowcasting methods for weak precipitation events. However, their network tends to provide too smooth precipitation fields and therefore fails to provide added value for more intense precipitation events with a rain rate above 10 mm h-1. Recently, a deep generative model for the probabilistic precipitation nowcasting was proposed and showed state-of-the-art performance for the task .

All these studies demonstrate that deep neural networks have the potential to provide added value for precipitation nowcasting. In our study, we focus on further improving the predictions of strong precipitation events and therefore choose a simple ConvLSTM and the advanced PredRNN-v2 model for competing with our newly proposed model architecture.

In the following, we present our proposed CLGAN architecture in more detail. Since CLGAN aims to benefit from ConvLSTM models, the U-Net architecture, and GAN models, we first introduce its components separately to provide a deeper understanding of and reasoning for the chosen components.

In addition to the widely used remote sensing data, e.g., radar composite images, measurements from densely distributed automatic weather stations can serve as an alternative in the data-driven weather forecasting task. In this study, minute-level precipitation measurements by rain gauges of AWSs over Guizhou, China (Guizhou AWS_ML precipitation dataset), are collected for the precipitation nowcasting task. Guizhou is a mountainous and rainy province located in southwest China (see Fig. a) where mudslides happen frequently during summertime. For instance, the region was affected by a severe rainstorm in September 2020, in which some regions experienced more than 1500 mm rainfall within 20 d. Accurate precipitation nowcasting, especially for heavy precipitation, is crucial to reduce damage from these events. Hence, the Guizhou AWS_ML precipitation dataset is established for better simulation of precipitation with data-driven approaches. The AWS locations comprise 93 basic national stations and 1740 automatic weather stations (see Fig. b). Among other meteorological quantities (2 m temperature, 10 m wind, surface pressure, and relative humidity), the AWSs measure precipitation at a high observation frequency (every minute), and the data are provided between 1 January 2015 and 31 December 2019 by Guizhou Meteorological Bureau.

Several preprocessing steps are conducted for preparing the dataset of our experiment. First, the precipitation data are accumulated over 10 min, which still constitutes a reasonably high temporal resolution. To obtain a gridded dataset, the observations are then interpolated bilinearly onto a regular, spherical grid. The target grid comprises w×h=48×40 data points in zonal and meridional directions, respectively, and covers a domain from 24.625 to 29.5∘ N and 103.625 to 109.5∘ E with 0.125∘ resolution. To obtain the data needed for training our CLGAN and the baseline models (see Sects.  and ), we generate sliding sequences of 24 consecutive gridded data samples (frames) which comprise a temporal period of 240 min, and 120 min (12 frames) of each sequence serves as input to predict the next 120 min (12 frames). Since there are many periods with no or only weak precipitation, we furthermore only select sequences whose averaged precipitation rate exceeds the empirical 60 % quantile of the complete dataset. This results in 35 054 sequences for the subsampled dataset. Finally, a log transformation is applied to each sequence to make the data more Gaussian-like. The log transformation reads as follows: x′=ln⁡(x+ε)-ln⁡(ε), where ε is a small constant (here 0.01). We use the data from 2015 to 2017 for training, the data of 2018 for validating, and the data of 2019 for testing.

As pointed out in and more specifically for precipitation in , precipitation nowcasting should be evaluated in terms of application-specific scores. This is due to the unique statistical properties of precipitation rates, as well as the chaotic atmospheric processes which underpin the formation of precipitation. Additionally, we would like to emphasize that a single score alone can barely evaluate the model performance applied to high-dimensional data . Therefore, we take several evaluation metrics into account to provide a comprehensive overview.

The first considered family of evaluation metrics is established for continuous quantities in the meteorological community. The root mean square error (RMSE) measures the distance between the predicted and the observed field on a grid-point level. The correlation coefficient (CC) measures the association or the linear relationship between the two fields. A perfect correlation would result in CC=1, while CC=0 indicates no linear relationship between forecast and observation on the grid-point level.

The second set of scores is built on dichotomous events which are obtained by thresholding the gridded precipitation fields. A 2×2 contingency table is commonly used to show the frequency of “yes” and “no” forecasts and occurrences and give a joint distribution for events with a precipitation rate exceeding a given threshold tpr. According to the elements in the contingency table, a variety of categorical statistics can be computed to evaluate the dichotomous forecasts in particular aspects. Critical success index (CSI), also known as threat score, measures the fraction of hits with respect to the number of occurrences where the events are either forecasted or observed. The frequently applied equitable threat score (ETS) is a variant of the CSI and explicitly accounts for random forecasts which perform well just by chance .

However, due to the highly nonlinear and complex processes causing precipitation formation, scores acting on grid-point level are prone to penalize predictions which recover the high spatial variability but fail to match exactly the observed precipitation field. The issue leads to the double penalty problem where the model gets penalized twice, once for missing the exact placement of a precipitation event and once for shifting it spatially . To relax the requirement for exact spatial matching, the fractions skill score (FSS) is computed here as a fuzzy verification metric . Similar to the CSI and ETS, the FSS operates on dichotomous events but allows for spatial shifts by considering a local neighborhood around each grid point. Within this neighborhood, the fractional coverage of the precipitation events is calculated for both the predictions and the observations. Let f(mis) and f(ois) denote the fraction of event grid boxes within the local neighborhood of size s around the grid point i in the model prediction and observation. The fractions Brier score (FBS) is given by 4FBS=1N∑i=1N(f(mis)-f(ois))2, which quantifies the quadratic difference between the prediction and the observation for all N grid points over the domain. The final FSS is then obtained with 5FSS=1-FBS/FBSworst, where FBSworst is the sum of the squared fractions of events in the prediction and in the observation. Higher FSS values indicate better forecast, while it can be shown that a forecast becomes “useful” when FSS≥0.5 is attained for a given neighborhood scale s (typically expressed in terms of squares with an edge length of N grid points).

Nonetheless, FSS also does not capture the spatial precipitation patterns since each grid point in the neighborhood is treated equally and no check for spatial coherence is undertaken. Thus, we additionally perform an object-based diagnostic evaluation, called MODE , to focus on pattern attributes such as location, area, and shape. To obtain the desired attributes, a convolutional filter of size k is first applied over the precipitation field. Afterwards, objects are defined by applying a threshold on the precipitation rate tPr and on the object area tA. A fuzzy logic scheme is then used to merge and pair precipitation objects in the predicted and observed precipitation field. Finally, the object-based threat score OTS; is computed to verify how well the predicted precipitation patterns match the observed ones. Here, we choose the object area, the centroid location, and the object shape (aspect ratio and orientation angle) as target attributes for computing the OTS.

All the mentioned verification methods are listed in Table  with a brief description. The details can be found in the corresponding references.

To ease the comparison between the baseline models and the simplistic persistence forecast, we furthermore calculate skill scores (except for the FSS). In general, a skill score (SS) can be constructed by considering the target score Sm of the model, the score obtained with the reference forecast Sref, and the perfect score Sperf: 6SS=Sm-SrefSperf-Sref. The higher the SS is, the better the model performs against the reference score. Perfect models thereby obtain SS=1, while inferior models show up with -∞<SS<0. Note that Sperf=0 holds for the RMSE, whereas the other scores under consideration attain Sperf=1. Since the size of our dataset is not unlimited, we also apply a block bootstrapping procedure to estimate sampling uncertainty . The block bootstrapping procedure accounts for autocorrelation between the sliding sequences and thus divides the dataset into non-overlapping blocks before the resampling of the blocks with replacement is performed. Here, we set the block length to 10 h (60 frames) and perform 1000 block bootstrapping steps.

To gain further insight into the realism of our predictions, a heavy precipitation event occurring on 12 June 2019 is visualized as an example (see Fig. ) to compare the model performance with an “eyeball” analysis. Figure a shows the observed precipitation rates in millimeters per 10 min for every 20 min over the forecast period starting at 06:50 CST. It is seen that a fairly strong precipitation system moves from west to east while it further intensifies. The predictions of the different models are presented as difference plots in Fig. b–f. For the first 60 min, persistence and DenseRotation show up with the smallest discrepancies. However, for longer lead times, clear dipole structures in the difference plots indicate that the movement of the system is not captured. Thus, the Lagrangian persistence framework is inaccurate for longer lead times, and more advanced models are required to capture the long-term dependence.

While the deep learning models also show increasing differences with longer lead times, they perform better in capturing the movement and the intensification of the precipitation system (see Fig. d–f). PredRNN-v2 tends to overestimate the precipitation intensity, which causes large coherent areas of positive differences. The averaged RMSE over the study area of the PredRNN-v2 forecasts at the lead time of 2 h is around 0.37 mm. ConvLSTM and CLGAN perform even better with smaller discrepancies, with a lower RMSE of 0.33 and 0.34 mm, respectively. Consistent with the quantitative evaluation results, ConvLSTM outperforms the CLGAN model in terms of RMSE in this given case, whereas CLGAN obtains a higher CSI and ETS. The CSI of the CLGAN forecasts with a threshold of 8 mm h-1 at the 2 h lead time is around 0.11, 0.01 higher than the ConvLSTM model. The results demonstrate that the prediction by ConvLSTM has a lower bias, while CLGAN can capture more heavy precipitation grid points. Compared to ConvLSTM, the difference plot of CLGAN contains more fine structures which indicates that CLGAN can generate more details of the precipitation system.

As shown in Eq. (), the loss function used in our CLGAN model consists of two terms: the adversarial loss LGAN and the reconstruction loss L2. To assess the contribution of each loss term on the forecasts' performance, sensitivity experiments on the weight of the reconstruction loss in CLGAN were carried out. Larger weight of the adversarial loss (equals smaller weight of reconstruction loss) is equivalent to a stronger contribution made by the GAN component. Figure  presents the results of the CLGAN model with different weights λ assigned to the L2 loss. It is seen that the RMSE is increased when reducing the weight of L2 loss (Fig. a). However, the scores for dichotomous events and fuzzy verification framework reveal improvements for smaller λ. The model using a pure reconstruction loss (λ=1 in Eq. ) performs significantly worse than the model applying an adversarial loss in terms of CSI and FSS (Fig. b and c). Similar results are obtained in terms of the OTS (see Fig. d). The results of sensitivity experiments indicate that the adversarial training with the GAN component encourages the model to generate forecasts which are more similar to the ground truth data. Despite a slight increase in RMSE, a relatively stronger contribution of the GAN component helps to capture the statistical properties of the observed precipitation (on the tail, as well as their spatial attributes) which in turn improves the prediction of strong precipitation events.