Accurate, timely and wide-ranging radar Quantitative Precipitation Estimation (QPE) is crucial for effective water resource management and climate research. However, the insufficient consideration of surface meteorological conditions such as moisture information leads to significant biases for existing QPE methods. This paper proposes a novel quantitative precipitation estimation network based on hierarchical multi-branch convolutional blocks (MCB-Net). The network incorporates relative humidity data and radar data into a multimodal-fusion architecture, effectively extracting information from diverse data sources and correcting precipitation loss due to evaporation. Additionally, a series of MCB blocks are designed to replace the traditional convolutional operation with multi-branch convolutional units, enabling MCB-Net to capture complicated spatial and temporal features related to precipitation. The indices including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Bias Ratio (MBR) and Correlation Coefficient (CC) are adopted to evaluate the performance of the proposed method. Experimental results demonstrate that MCB-Net outperforms conventional method and other deep learning based QPE networks including Volume-to-Point CNN network, U-Net and ResNet. The visualized results illustrate that the proposed network can effectively mitigate precipitation overestimation and enhance the accuracy of precipitation estimation.

For ground-based weather radar systems, reflectivity is particularly crucial for monitoring severe convective events. However, its limited coverage poses challenges in acquiring reliable radar data, especially for oceanic and mountainous regions. In contrast, geostationary meteorological satellites offer near-global coverage and near-real-time cloud-top observations. This paper introduces a novel deep learning-based radar reflectivity emulation method to reconstruct surface radar observations from cloud-top satellite data, termed Dynamic Residual Convolution-based Network (DRC-Net), aiming to provide more accurate and reliable reflectivity data in regions lacking radar coverage. It uniquely combines dynamic convolution, which focuses attention on convolutional kernels for dynamically adjusting weights based on input, with residual convolution, effectively enhancing network’s ability to capture intricate radar echo details. Experimental results demonstrate that DRC-Net outperforms existing methods in various assessment indices, including Probability Of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI) and Heidke Skill Score (HSS). Generalization tests and case studies further illustrate its effectiveness in reconstructing radar reflectivity across various regions, particularly in mountainous and oceanic areas.

A ground-based weather radar is commonly used for observing severe convective weather. However, the limited coverage of the radar poses difficulties in obtaining reliable radar observations for oceanic and mountainous regions. An effective solution is to derive radar data from meteorological satellite observations using deep-learning methods. This study proposes a novel feature redistribution module-based convolutional neural network (FR-CNN) to retrieve radar composite reflectivity (CREF) data from geostationary satellite observations. Differing from existing skip connection (SC)-based CNNs, FR-CNN adopts a feature redistribution module (FRM) to alleviate the problem of information scarcity during network propagation. In the FRM, a parallel attention block (PAB) is introduced to preserve key feature information and improve the retrieval ability of the FR-CNN. The evaluation results show that the FR-CNN can effectively reconstruct radar reflectivity data and has a better performance than other methods like U-Net in terms of assessment indices including the probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI).

Accurate and real-time estimation of surface precipitation is crucial for decision-making during severe weather events and for water resource management. Polarimetric weather radar serves as the primary operational tool employed for quantitative precipitation estimation (QPE). However, the conventional parametric radar QPE algorithms overlook the dynamic spatiotemporal characteristics of precipitation. In addition, challenges such as radar beam attenuation and imbalanced distribution of precipitation data further compromise the estimation accuracy. This article develops a 3-D star neural network (StarNet) for polarimetric radar QPEs that integrate physical height prior knowledge and employ a reweighted loss function. To better cope with the dynamic characteristics of precipitation patterns, 3-D convolution is introduced within StarNet to effectively capture the spatiotemporal features between successive radar volume scanning data. In particular, multidimensional polarimetric radar observations are utilized as inputs, and surface gauge measurements are employed as training labels. The feasibility and performance of the StarNet model are demonstrated and quantified using U.S. Weather Surveillance Radar-1988 Doppler (WSR-88D) observations collected near Melbourne, Florida. The experimental results show that the StarNet model enhances the prediction accuracy of moderate to heavy precipitation events and improves the estimation performance over long distances, with a mean absolute error (MAE) of 1.55 mm, a root mean square error (RMSE) of 2.63 mm, a normalized standard error (NSE) of 25%, a correlation coefficient (CC) of 0.92, and a BIAS of 0.94 for hourly rainfall estimates. The results suggest that StarNet is able to effectively map the connection between polarimetric radar observations and surface rainfall.

Real-time and accurate precipitation estimation is critical for environmental protection and water resources management. Compared to traditional methods, i.e., radar reflectivity (Z) and rainfall rate (R) relations, relying on local raindrop size distributions, the deep learning model can fit the functional relationship between radar observations and rainfall rate measurements. However, the black-box nature of deep learning models makes it difficult to explain the physical mechanisms behind their results. To address this problem, this study proposes DQPENet, a deep learning model for polarimetric radar QPE utilizing dense blocks. We employ a permutation test to understand the relative importance of different radar data input variables. Additionally, we propose a regression importance value (RIV) method for the precipitation estimation task to visualize feature importance regions. Our experimental results show that radar reflectivity and specific differential phase at the lowest elevation angle are the two most important observables for the model’s precipitation estimation. Furthermore, we find that radar data closer to the rain gauge are more influential on the model’s results, indicating that the deep learning model is able to capture the underlying physical mechanism of atmospheric data.

Research on nowcasting through dual-polarization weather radar data using deep learning approach is rare but worth exploring. This paper lightens a previous work, the MCT (Multivariate Channel Transformer) model, which leads to the design of the MSF (Multivariate Swin Fusion) model. The commonalities between the two are as follows: on one hand, both fuses several dual-polarization observables including reflectivity (Z), specific differential phase (Kdp), and differential reflectivity (Zdr) to more comprehensively consider meteorological particle features; on the other hand, they introduces the attention mechanism to more fully fuse multi-frame, multi-variate, and multi-scale features. In the experimental evaluation, this study first selects observation data from KMLB radar in FL, USA, and uses traditional optical flow method, deep learning TrajGRU method, etc. as controls. The results show that both MCT and MSF perform better than the control, and the 60min forecast scores of both are 8.78/9.31 for RMSE and 0.46/0.18/0.07 for CSI (20/35/45dBZ), and this conclusion is verified by case study. Further, the role of the attention mechanism is verified by ablation experiments.

The task of nowcasting by deep learning using multivariate, rather than just reflectivity, is limited by poor interpretability. The previous experiment designed MCT (Multivariate Channel Transformer), a deep learning model capable of nowcasting with dual-polarization radar data. Four analytical methods are designed to further explore the contribution of polarization parameters: (i) Case studies of different meteorological processes. (ii) A permutation test ranking the significance of each variable. (iii) Visualization of the feature maps obtained by forward propagation of the input data. (iv) Data downscaling of polarimetric radar data. The results show that the polarization parameters serve as a guide to predict the location and shape of strong reflectivity, as well as the energy retention of strong echoes at 40-50 dBZ. The contributions of Zdrand Kdpare more evident in the prediction results after 30 min, and the importance of Kdpexceeds that of Zdrin case of strong convective weather.

Precipitation is the primary driver of the regional/global hydrological cycle. Accurate precipitation estimation is essential for disaster management and climate change studies. Geostationary meteorological satellites provide a wealth of temporal and spatial information for quantitative precipitation estimation (QPE), but how to extract useful precipitation-related information from raw satellite data has always been a challenge. As a result, there is still a gap between the current satellite precipitation products and ground-based precipitation observations. Recent advancements in deep learning (DL) techniques now allow the development of more accurate QPE algorithms. This letter proposes a PREcipitation estimation algorithm based on Multi-task learning (PREM) using satellite observations extracted from the Geostationary Operational Environmental Satellite-R (GOES-R) series. The key design points of PREM are as follows: 1) the detection (rain/no-rain separation) and rainfall rate estimation are integrated in PREM; 2) infrared (IR) radiance data are the main input to PREM, and ground-based radar precipitation products are used as the target in the training phase; and 3) in addition, PREM integrates lightning data (called PREM-L) to investigate its impact on heavy precipitation estimation. Independent verification shows that PREM performs better than the operational GOES-R products, and PREM-L outperforms PREM when precipitation is heavy.

Accurate estimation of surface precipitation with high spatial and temporal resolution is critical for decision making regarding severe weather and water resources management. Polarimetric weather radar is the main operational instrument used for quantitative precipitation estimation (QPE). However, conventional parametric radar QPE algorithms such as the radar reflectivity (Z) and rain rate (R) relations can not fully represent clouds and precipitation dynamics due to their dependency on local raindrop size distributions and the inherent parameterization errors. This article develops four deep learning (DL) models for polarimetric radar QPE (i.e., RQPENetD1, RQPENetD2, RQPENetV, RQPENetR) using different core building blocks. In particular, multi-dimensional polarimetric radar observations are utilized as input and surface gauge measurements are used as training labels. The feasibility and performance of these DL models are demonstrated and quantified using U.S. Weather Surveillance Radar - 1988 Doppler (WSR-88D) observations near Melbourne, Florida. The experimental results show that the dense blocks-based models (i.e., RQPENetD1and RQPENetD2) have better performance than residual blocks, RepVGG blocks-based models (i.e., RQPENetRand RQPENetV) and five traditionalZ-Rrelations. RQPENetD1has the best quantitative performance scores, with a mean absolute error (MAE) of 1.58 mm, root mean squared error (RMSE) of 2.68 mm, normalized standard error (NSE) of 26%, and correlation of 0.92 for hourly rainfall estimates using independent rain gauge data as references. These results suggest that deep learning performs well in mapping the connection between polarimetric radar observations aloft and surface rainfall.

The short-term convective storm forecasting (i.e., nowcasting) mainly relies on weather radar, which can resolve the 3-D structure of convective storms. With the rapid development of numerical models, modern models can produce 3-D reanalysis data, which gives atmospheric background information of convective storms. Current deep learning nowcasting models use only 2-D radar images for nowcasting and often require massive historical data for training. But, it may not be operationally feasible to collect long-term radar data to train a new model. Hence, how to establish a nowcasting model using only a small dataset has become an important issue. In addition, the existing models do not effectively use the state-of-the-art model reanalysis data, which is a shortcoming of these models. To tackle these problems, this article develops a pixelwise convolutional-recurrent neural network (Pixel-CRN) for precipitation nowcasting. It has three key designs: 1) through a concise pixelwise sampling and oversampling technique, Pixel-CRN can be trained using only a small dataset; 2) in spatial learning, Pixel-CRN embeds a spatial convolution subnet into the recurrent unit, which can input raw 3-D radar and model reanalysis data; thus, valuable atmospheric background information can be learned to assist nowcasting; and 3) in spatiotemporal learning, according to the information bottleneck principle, Pixel-CRN builds a heterogeneous encoder–decoder structure to squeeze multichannel 3-D input data into latent space and recurrently generates 30- and 60-min nowcasts. Compared with the existing deep learning nowcasting methods, the experimental results show that the Pixel-CRN can provide skillful results with a rather small training dataset.

The correct distinction between precipitation echo (PE) and non-precipitation echo (NPE) of weather radar is one of the key steps in the quantitative precipitation estimation (QPE). Based on the high spatiotemporal resolution data of the Himawari-8 satellite and the precipitation data of the rain gauge, this paper proposes a fuzzy logic algorithm (FLA) to remove NPE in weather radar data through statistically analyzing the probability density distribution of the satellite temperature of black body (TBB) according to different precipitation scenarios. Two methods, i.e. the FLA and the threshold method (TM), are compared and analyzed by using the data in August 2019. The critical success index (CSI) score of the FLA is 0.056 higher than that of the TM (0.796 vs. 0.74), and the probability of detection (POD) score is 0.058 higher than that of the TM (0.802 vs. 0.744). The results show that the proposed method can effectively improve the quality of weather radar data.

Convective storm is one of the major catastrophic weather events. In recent years, deep learning algorithms fused with various meteorological data are emerging as mainstream techniques in related research areas. However, the black-box nature of deep learning often makes such algorithms hard to interpret from physical point of view. In this study, three interpretation methods are developed to interpret a deep learning nowcasting model, i.e. permutation test, saliency maps and Grad-CAM (Gradient-weighted Class Activation Mapping). After ranking the importance of the eight variables selected in this study by permutation test, it is concluded that the radar reflectivity is the most important factor for the occurrence of convective storm. Based on the results of saliency maps and Grad-CAM, the areas with high perturbation temperature, strong wind and low humidity are also important for short term prediction of convective storms.

Detection of convective initiation (CI) is an important step of early warning of strong convective weather. This study proposes a machine learning approach for CI detection using 18 interesting fields extracted from weather radar, Himawari-8 Advanced Himawari Imager (AHI), and the variational Doppler radar analysis system (VDRAS). The collected features are used to train the machine learning model for CI detection. Therein, the support vector machine (SVM) is used to identify the CI and non-CI. It is concluded that a better result can be achieved by using multiple-source data compared to using satellite data only. The atmospheric boundary layer thermal dynamic information retrieved by VDRAS is proved to be useful for CI detection.

Radar quantitative precipitation estimation (QPE) is one of important applications of Doppler weather radar. The use of the rain gauge data to correct and improve the accuracy of radar QPE is important for weather forecasting. Based on observation data of rain gauges, this study proposes two machine learning methods to correct radar QPE product. The Support Vector Regression (SVR) and Random Forest (RF) models are used as the regression model. The rain gauge data is used as the ground truth. The radar QPE data is divided into three intervals: 0–10 mm/h, 10–30 mm/h, and> 30 mm/h. One traditional method, i.e., the PDF method, is selected for comparison. The experimental results show that the machine learning methods achieve better performance than the PDF method. The scatter plot shows that SVR effectively mitigate the overestimation problem which is a main problem of the original radar QPE product.

Deep learning methods have been applied to weather forecasting and achieved superior performance than traditional methods. Can we use postprocessing or bias correction methods to further improve the forecasting results? In this study, a frequency-matching correction method is proposed to correct the nowcasting results provided by a deep learning nowcasting method, i.e., TrajGRU. The frequency distribution curve is calculated first using the forecast and radar observation. Then, the correction matrix can be obtained based on this distribution curve. Finally, the forecast is corrected using the correction matrix. The experimental results show that the deep learning model also has bias which needs to be corrected. The proposed method achieves better forecast performance after correction. When using 50 dBZ as the threshold, more obvious forecast improvement can be observed.

In this study, we propose a deep learning nowcasting method called multivariate channel transformer U-net (MCT If-net). It incorporates the channel transformer into the basic U-net structure. Three dual-polarization radar variables, i.e., reflectivity ($Z$), specific differential phase ($K_{dp}$), and differential reflectivity ($Z_{dr}$), are used as inputs to the proposed model. Compared to traditional U-net, MCT U-net has two advantages: 1) We use channel transformer to replace the skip-connection operation to better capture global information. Due to the inability to fuse the semantic features at different scale, a simple skip connection operation cannot capture the global multi-scale context, and this study uses the self-attention mechanism in Transformer to tackle this problem. 2) MCT U-net extracts multi-scale features through down-sampling each variable and then concatenates and sends them to a channel transformer. The combination of multiple variables in higher semantic level helps the model learn each individual modality better. The experimental results show that MCT U-net improves nowcasting performance compared with traditional methods.

Precipitation nowcasting is an important operational service for protecting public property losses and people’s safety. Short-Term Ensemble Prediction System (STEPS) is a probabilistic nowcasting system which has been widely used in the research community (commonly referred as PySTEPS). This study investigates the predictability of PySTEPS during different precipitation systems, i.e., convective and stratiform events. In particular, two study domains, namely, Dallas-Fort Worth (DFW) area in northern Texas and San Francisco Bay Area in northern California, are selected to represent these two typical precipitation patterns, respectively. The experimental nowcasting results show that PySTEPS works well in both areas, especially during stratiform rainfall events in the Bay Area. In addition, PySTEPS exhibits different performance for different precipitation patterns. For convective cases in the DFW area, PySTEPS tends to underestimate rain rate for high-intensity precipitation regions. For stratiform cases in the Bay Area, PySTEPS can predict the precipitation intensity more accurately. With the increase of nowcasting lead time, the qualitative evaluation scores (POD - Probability of Detection, and CSI - Critical Success Index) of PySTEPS decrease slowly during stratiform events compared with convective events, which is also in line with the quantitative evaluation results.

Convective precipitation nowcasting remains challenging due to the fast change in convective weather. Radar images are the most important data source in nowcasting research area. This study proposes a radar data-based U-Net model for precipitation nowcasting. The nowcasting problem is first transformed into an image-to-image translation problem in deep learning under the U-Net architecture, which is based on convolutional neural networks (CNNs). The input of the model is five consecutive radar images; the output is the predicted radar reflectivity image. The model consists of three operations: upsampling, downsampling, and skip connection. Three methods, U-Net, TREC, and TrajGRU, are used for comparison in the experiments. The experimental results show that both deep learning methods outperform the TREC method, and the CNN-based U-Net can achieve almost the same performance as TrajGRU which is a recurrent neural network (RNN)-based model. With the advantages that U-Net is simple, efficient, easy to understand, and customize, this result shows the great potential of CNN-based models in addressing time-series applications.

Deep learning is emerging as a powerful tool in scientific applications, such as radar-based convective storm nowcasting. However, it is still a challenge to extend the application of a well-trained deep learning nowcasting model, which demands to incorporate the learned knowledge at a certain location to other locations characterized by different precipitation features. This article designs a transfer learning framework to tackle this problem. A convolutional neural network (CNN)-based nowcasting method is utilized as the benchmark, based on which two transfer learning models are constructed through fine-tune and maximum mean discrepancy (MMD) minimization. The base CNN model is trained using radar data in the source study domain near Beijing, China, whereas the transferred models are applied to the target domain near Guangzhou, China, with only a small amount of data in the target area. The influence of a varying number of target data samples on the nowcasting performance is quantified. The experimental results demonstrate that the deep transfer learning models can improve the nowcasting skills.

Convective precipitation nowcasting remains challenging due to the fast change of convective weather. Radar images are the most important data source in nowcasting research area. This study proposes a radar data-based U-Net model for precipitation nowcasting. The input of the model is five consecutive radar images; the output is 30-min prediction of radar reflectivity image. The model consists of three parts: upsampling, downsampling and skip-connection. Two models, U-Net and TrajGRU, are used for comparison in the experiments. Different from U-Net, which is a CNN-based (Convolution Neural Network) model, TrajGRU is an RNN-based (Recurrent Neural Network) model, which is good at time-series processing and has been widely used in precipitation research community. The experimental results show that the CNN-based U-Net can achieve almost the same performance as TrajGRU. This result shows the great potential of CNN-based models in handling time-series applications.

Compared to the legacy GOES Imager, the GOES-16 Advanced Baseline Imager (ABI) provides measurements at much higher resolution in both spatial and temporal dimensions, which benefits many applications such as rainfall rate estimation. However, it is challenging to quantify the water droplets in the cloud and precipitation intensity with only the Infrared (IR) radiation and converted brightness temperature information. Comparison between ground radar rainfall estimates and the operational GOES-16 rain rate products indicates that uncertainty associated with GOES-16 rain rate estimates is significant. This paper proposes an innovative deep learning system for precipitation estimation using measurements from GOES-16/ ABI. Cloud-top brightness temperature information observed by channels 8/10/11/14/15 of the ABI are used as input to this deep learning framework. The rainfall estimates from a ground radar network are used as target labels in the training phase. Independent verification shows that the deep learning-based rainfall system can estimate precipitation time, intensity, and amount very well, especially in heavy rain regions.

Satellite remote sensing observations at high spatial and temporal resolutions have advantages for precipitation retrievals. However, it is still challenging to produce reliable precipitation estimates over complex terrain regions. This paper proposes a deep learning framework to correct the bias associated with satellite revivals of orographic precipitation, focusing on the influence of terrain on precipitation estimation. Using precipitation product derived by the NOAA Climate Prediction Center morphing technique (i.e., CMORPH-CRT) in the western United States as an example, the deep learning model is trained and tested, and the ground-based Stage IV rainfall estimates are used as target labels in the training phase. Independent validation results show that the deep learning-based bias correction model can significantly enhance the performance of satellite precipitation products over complex terrain.

Nowcasting of precipitation associated with landfalling hurricanes refers to the hurricane intensity analysis and forecast for the next few hours. It plays an important role in early weather warning system and real-time decision making for disaster management and risk reduction. However, the skill of traditional nowcasting techniques decreases rapidly in the first hour. In addition, numerical weather prediction (NWP) models have deficiency to provide accurate warning of heavy precipitation on the scale needed for hurricane nowcasting. Therefore, this paper proposes a deep learning-based system for precipitation nowcasting for landfalling hurricanes using multi-radar observations near the coast. The reflectivity imageries from previous 1-hr observations are used as input to the deep learning-based nowcasting model and the actual observations from “future” sequences are used as targets in the training stage. The performance and evaluation results of the nowcasting products with lead time up to 3 hours show that this deep learning system performs well in hurricane nowcasting.

Convective storm nowcasting has long been an important issue and has attracted substantial interest. 3D radar images and 3D re-analysis data contain spatiotemporal information of the convective processes. This paper proposes a multi-channel 3D convolutional recurrent neural network (3D-CRN) for convective storm nowcasting, which aims to learn spatiotemporal information directly from these 3D radar and reanalysis data. 3D-CRN is composed of two sub-networks: three multi-channel 3D convolutional networks are used as the front-end spatial sub-networks, and the convLSTM encoder-decoder is constructed as a back-end temporal sub-network. By combing two subnets into one unified network, 3D-CRN can be jointly trained effectively. In order to give forecasts at different lead time simultaneously, we construct a many-to-many encoder-decoder structure to avoid tedious need to train several models respectively. Experimental results show the effectiveness of the proposed method.

The goal of convective storm nowcasting is local prediction of severe and imminent convective storms. Here, we consider the convective storm nowcasting problem from the perspective of machine learning. First, we use a pixel-wise sampling method to construct spatiotemporal features for nowcasting, and flexibly adjust the proportions of positive and negative samples in the training set to mitigate class-imbalance issues. Second, we employ a concise two-stream convolutional neural network to extract spatial and temporal cues for nowcasting. This simplifies the network structure, reduces the training time requirement, and improves classification accuracy. The two-stream network used both radar and satellite data. In the resulting two-stream, fused convolutional neural network, some of the parameters are entered into a single-stream convolutional neural network, but it can learn the features of many data. Further, considering the relevance of classification and regression tasks, we develop a multi-task learning strategy that predicts the labels used in such tasks. We integrate two-stream multi-task learning into a single convolutional neural network. Given the compact architecture, this network is more efficient and easier to optimize than existing recurrent neural networks.

Precipitation is a critical element in global atmospheric circulation and ecosystem. However, it is challenging to monitor the global or even continental scale precipitation features using ground-based rain gauges due to the spatial coverage limitations. The Geostationary Operational Environmental Satellite-R (GOES-R) series provide new opportunities for continuous observation of precipitation at large scales. This paper presents a detailed cross-comparison of quantitative precipitation estimates (QPE) between GOES-R (GOES-16) and the ground radar-based rainfall product derived from the National Ocean and Atmospheric Administration's (NOAA) multi-radar multi-sensor (MRMS) system over the continental United States (CONUS). The precipitation detectability of GOES-R is investigated using MRMS products as references. In addition, uncertainties associated with the GOES-R precipitation estimates are quantified in different regions over the CONUS.

Precipitation nowcasting is one of the most important applications of weather radars. Essentially, the nowcasting performance is determined by the radar data quality and the implemented nowcasting model. It is well known that the dual-polarization radar can observe more detailed microphysical information about precipitation and significantly enhance the radar data quality compared to traditional single-polarization radars. In this study, we extend an ensemble-based probabilistic precipitation nowcasting scheme for high-resolution polarimetric radar applications focusing on fast-evolving convective precipitation events. In particular, the nowcasting scheme blends an extrapolation technique with a probabilistic nowcasting method that is adapted to work on polarimetric radar-derived precipitation estimates. Case studies based on the S-band Weather Surveillance Radar - 1988 Doppler (WSR-88D) in Fort Worth, Texas (i.e., KFWS radar) show that the nowcasting products agree well with the real observations. However, the accuracy of short-term rainfall nowcasting is greatly influenced by the motion field, and it is still a challenge to predict the initiation of new precipitating clouds.

Convective storms are one of the severe weather hazards found during the warm season. Doppler weather radar is the only operational instrument that can frequently sample the detailed structure of convective storm which has a small spatial scale and short lifetime. For the challenging task of short-term convective storm forecasting (i.e., nowcasting), 3-D radar images contain information about the processes in convective storm. However, effectively extracting such information from multisource raw data has been problematic due to a lack of methodology and computation limitations. Recent advancements in deep learning techniques and graphics processing units (GPUs) now make it possible. This article investigates the feasibility and performance of an end-to-end deep learning nowcasting method. The nowcasting problem was transformed into a classification problem first, and then, a deep learning method that uses a convolutional neural network (CNN) was presented to make predictions. On the first layer of CNN, a cross-channel 3-D convolution was proposed to fuse 3-D raw data. The CNN method eliminates the handcrafted feature engineering, i.e., the process of using domain knowledge of the data to manually design features. Operationally produced historical data of the Beijing-Tianjin-Hebei region in China was used to train the nowcasting system and evaluate its performance; 3 737 332 samples were collected in the training data set. The experimental results show that the deep learning method improves nowcasting skills compared with traditional machine learning methods.

very short-term weather forecasting or nowcasting has attracted substantial attention in various fields. Existing methods can nowcast storm advection based on radar data. Due to the limitations of the radar observations, it is still challenging to nowcast storm initiation and growth. However, as the real-time re-analysis meteorological data can now provide valuable atmospheric boundary layer thermal dynamic information, which is essential to predict storm initiation and growth. It is of great importance to leverage these re-analysis data.This paper describes our first attempt to nowcast storm initiation, growth, and advection simultaneously under the framework of convolutional neural network using the very large multi-source meteorological data. To this end, we construct a multi-channel 3D-cube successive convolution network which leveraging both raw 3D radar and re-analysis data directly without any handcraft feature engineering. These data are formulated as multi-channel 3D cubes, to be fed into our network, which are convolved by cross-channel 3D convolutions. By stacking successive convolutional layers without pooling, we build an end-to-end trainable model for nowcasting. Experimental results show that deep learning methods achieve better performance than traditional extrapolation methods. The qualitative analyses of our approach show encouraging results of nowcasting of storm initiation, growth, and advection.