Information fusion-based dual-polarization radar urban heavy rainfall short-term and imminent forecast deep learning method and system

Abstract

The invention belongs to the technical field of weather forecast, and discloses an information fusion dual-polarization radar urban heavy rainfall short-term and imminent forecast deep learning method and system. The invention provides a novel urban heavy rainfall short-term and imminent forecasting method based on meteorological radar data through a research thought of'space-time-space-time fusion 'on the basis of a deep learning technology in combination with a physical mechanism of heavy rainfall. According to the method, heavy rainfall is accurately forecasted by analyzing spatial and temporal distribution characteristics of heavy rainfall, extracting high-dimensional spatial characteristics, developing a super-resolution heavy rainfall spatial characteristic extraction model, combining time evolution characteristics of heavy rainfall and adopting an LSTM and spatial and temporal characteristic fusion model, and the accuracy and timeliness of urban heavy rainfall forecasting are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological forecasting, and in particular relates to an information fusion dual-polarization radar city heavy rainfall short-term forecast deep learning method and system. Background Art

[0002] With the continuous acceleration of urbanization, the frequent occurrence of heavy precipitation in urban areas has had a serious impact on urban infrastructure, traffic safety, people's lives, etc. Especially in urban areas, due to dense buildings, complex terrain, and frequent changes in airflow, the occurrence, evolution, and spatial distribution of heavy precipitation have strong local and nonlinear characteristics. Therefore, accurate and efficient short-term forecasting of heavy precipitation in cities has important social and economic significance.

[0003] At present, meteorological radar plays a vital role in heavy precipitation monitoring and forecasting.

[0004] Suman Ravuri demonstrated in a Nature paper (Ravuri, S., Lenc, K., Willson, M. et al. Skilful precipitation nowcasting using deep generative models of radar. Nature 597, 672–677 (2021).) that radar data can be used for short-term forecasts using artificial intelligence methods, and that good forecast results can be obtained. They developed the DGMR model and proved that such models are of great help in short-term forecasts.

[0005] Yuchen Zhang showed in a Nature paper (Zhang, Y., Long, M., Chen, K. et al. Skilfulnowcasting of extreme precipitation with NowcastNet. Nature 619, 526–532(2023).) that incorporating physical constraints (such as advection equations) into the radar short-term forecast model can effectively improve the radar short-term forecast effect, proving that physical constraints can be applied to related technologies.

[0006] Radar data can provide real-time observation information on the occurrence, development, spatial distribution and intensity changes of precipitation. However, the existing radar heavy precipitation short-term forecast system based on deep learning methods still faces the following technical problems:

[0007] Blurred precipitation boundaries and insufficient spatial accuracy:

[0008] In the current heavy precipitation forecast model, although the presence of precipitation can be detected, the prediction of precipitation boundaries is often not clear enough, and it is difficult to accurately identify the starting and ending boundaries of precipitation. Especially in complex urban terrain, the spatial distribution of precipitation is greatly affected by geographical factors. Existing models are difficult to accurately simulate the local characteristics of precipitation at the urban scale, resulting in insufficient accuracy of the precipitation range, affecting the reliability of the forecast results.

[0009] Nonlinear impact of heavy precipitation changes:

[0010] The occurrence and change of heavy precipitation is a highly nonlinear process, which is affected by a variety of meteorological factors, geographical conditions and atmospheric dynamics. In the existing deep learning radar forecast model, the spatiotemporal evolution characteristics of heavy precipitation have complex nonlinear dependencies, which makes it difficult for the model to fully capture these changing laws, thus affecting the accurate prediction of precipitation development trends, especially the timeliness and accuracy of short-term forecasts.

[0011] Challenges of time series feature modeling:

[0012] The temporal characteristics of heavy precipitation usually have strong dependence and periodicity, but existing models still have difficulties in modeling time series data. Although time series modeling methods such as long short-term memory networks (LSTM) have improved the processing capabilities of time series data to a certain extent, how to effectively extract and fuse the temporal evolution characteristics of precipitation is still an unresolved problem. In particular, in the construction of high-precision spatiotemporal prediction models, how to be more efficient and accurate in temporal feature extraction and spatiotemporal fusion is still a major challenge facing current technology.

[0013] Through the above analysis, the problems and defects of the prior art are as follows:

[0014] (1) Blurred precipitation boundaries and insufficient spatial accuracy: Existing models are difficult to accurately simulate the local characteristics of precipitation at the urban scale, resulting in insufficient accuracy of the precipitation range, which affects the reliability of the forecast results.

[0015] (2) Nonlinear impact of heavy precipitation changes: Since the temporal and spatial evolution characteristics of heavy precipitation have complex nonlinear dependencies, it is difficult for the model to fully capture these changing patterns, which affects the accurate prediction of precipitation development trends, especially the timeliness and accuracy of short-term forecasts.

[0016] (3) Challenges in modeling time series features: Existing models still have difficulties in modeling time series data. Although time series modeling methods such as long short-term memory networks (LSTM) have improved the processing capabilities of time series data to a certain extent, how to effectively extract and integrate the temporal evolution characteristics of precipitation remains an unresolved problem. Summary of the invention

[0017] In view of the problems existing in the prior art, the present invention provides a deep learning method for short-term forecast of heavy rainfall in cities using dual-polarization radar with information fusion.

[0018] The present invention is implemented in this way: a deep learning method for short-term forecast of heavy rainfall in cities using dual-polarization radar with information fusion includes:

[0019] Step 1, construct radar heavy precipitation dataset;

[0020] Step 2: terrain information is integrated with radar images;

[0021] Step 3, super-resolution CNN design;

[0022] The super-resolution CNN model uses a convolutional neural network to learn low-resolution images and generate high-resolution images. In the heavy rainfall forecasting task, radar images are used as input, combined with terrain information DEM maps, and feature extraction and boundary enhancement are performed through a super-resolution CNN network. Super-resolution CNN learns the mapping relationship between low-resolution radar images and high-resolution images.

[0023] Step 4, LSTM model and seasonal weights are added;

[0024] In heavy precipitation prediction, LSTM is used to process radar image sequences or other time series data;

[0025] Step 5: Joint training.

[0026] Further, the radar heavy precipitation data set is constructed:

[0027] (1) Data collection

[0028] Radar images: Raw radar reflectivity images obtained from weather radar systems, usually based on high-frequency observations over a short period of time;

[0029] Precipitation observation data: using ground meteorological stations or satellite data to collect field observation data of heavy precipitation events, including precipitation amount and precipitation intensity indicators;

[0030] (2) Data labeling

[0031] Heavy precipitation area annotation: using observation data and radar images to define heavy precipitation areas;

[0032] (3) Data preprocessing

[0033] Image slicing: divide radar images into small pieces; train network models; each image should be labeled with whether it belongs to a heavy rainfall area;

[0034] Data enhancement: Use data enhancement technology to generate more training samples;

[0035] Normalization: Radar images and precipitation intensities are normalized, usually scaling the values ​​to the range of [0, 1] or [-1, 1].

[0036] Furthermore, the terrain information is integrated with the radar image:

[0037] Fusion method design:

[0038] Fusion of input data: Combine radar images with terrain elevation maps DEM to form multi-channel input:

[0039] Radar data processing:

[0040] Get radar image: radar data is presented in the Z field of reflectivity map, indicating the intensity of the target; radar image is denoised; suppose the denoised radar image is , whose shape is ,in is the image height, is the image width, is the number of channels of the radar image;

[0041] DEM data processing:

[0042] Get the digital elevation model (DEM) map: DEM maps represent the elevation information of the surface, usually stored in matrix form, indicating the elevation value of each geographic point; DEM data needs to be aligned with the radar image, usually matched by interpolation methods to ensure that the resolution of the two is consistent; assuming that after interpolation processing, the dimension of the DEM map is ,in , represents the terrain height value of each point;

[0043] Fusion of radar map and DEM map:

[0044] Weighted fusion: The radar image and DEM image can be fused in a weighted manner, while retaining the terrain information and precipitation characteristics of the radar image; the formula for weighted fusion is:

[0045]

[0046] in, is a hyperparameter used to control the fusion ratio of radar images and DEM images; by choosing the appropriate , the influence of terrain characteristics on the spatial characteristics of heavy precipitation can be adjusted;

[0047] Splicing and fusion: Splice the radar map and DEM image in the channel dimension to generate a multi-channel input feature map; this method explicitly fuses the terrain features and precipitation features at the input layer of the model:

[0048]

[0049] Among them, the stitched image The number of channels is 2, which contains radar reflectivity and terrain information;

[0050] Assume that the size of the radar image is , the topographic map (DEM) has the same size, so we can directly concatenate the two to get a three-channel input:

[0051]

[0052] in, is the radar image, It is a terrain elevation map, the stitched image Has 3 channels.

[0053] Furthermore, the super-resolution CNN is designed as follows:

[0054] Network structure:

[0055] The super-resolution CNN model generally consists of three main parts: feature extraction (convolution layer), nonlinear mapping (activation function) and image restoration (deconvolution layer);

[0056] Input layer: The input image size is W×H×C,

[0057] Where W and H are the width and height of the image, and C is the number of channels of the input image;

[0058] Convolutional layers: Multiple convolutional layers are used to extract spatial features of the input image; each convolutional layer is followed by an activation function and batch normalization;

[0059] Upsampling layer: restore low-resolution images to high resolution using deconvolution layers or pixel rearrangement layers; deconvolution layers spatially expand the resolution of images while maintaining the spatial structure in the image;

[0060] Output layer: The output layer usually uses a convolutional layer to generate a super-resolution image; at this point, the image size has been restored to the target high resolution;

[0061] Network details and layer configuration:

[0062] Input layer: Input image of size 64×64×3; the image contains a fusion map of radar image and terrain information;

[0063] Convolutional Layer:

[0064] Layer 1: convolution kernel size 3×3, stride 1, number of convolution kernels 64, activation function uses ReLU;

[0065] Layer 2: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 128, and the activation function uses ReLU;

[0066] Layer 3: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 256, and the activation function uses ReLU;

[0067] Deconvolution layer:

[0068] Use the deconvolution layer to upsample the image; assuming the goal is to restore to high resolution, set the deconvolution kernel size to 4×4 and the step size to 2;

[0069] Layer 1: Deconvolution kernel size 4×4, stride 2, output channels 128;

[0070] Layer 2: Deconvolution kernel size 4×4, stride 2, output channels 64;

[0071] Output layer:

[0072] The output layer uses a convolution layer to restore the super-resolution image to the target image, with an output channel of 1; the output size is 128×128×1;

[0073] Solve the problem of blurred boundaries of heavy precipitation:

[0074] The boundaries of heavy rainfall areas are usually fuzzy. Super-resolution CNN can accurately reconstruct these fuzzy areas by learning detailed features.

[0075] Convolutional layer extracts features:

[0076]

[0077] Each layer of convolution operation extracts the spatial features of the input image and enhances edge details;

[0078] Deconvolution: Upsample the low-resolution image to a higher resolution through transposed convolution to restore the details in the image, especially the boundary area of ​​precipitation:

[0079] .

[0080] Furthermore, the LSTM model is added with seasonal weights:

[0081] The basic architecture of the LSTM model:

[0082] The core of the LSTM model is its internal cell structure, which contains three main gating mechanisms (input gate, forget gate, and output gate), and a memory unit; the LSTM model controls the flow of information through these gating mechanisms and maintains memory of previous inputs over a long time scale; the basic update formula of LSTM is as follows:

[0083] Input Gate: controls the impact of current input information;

[0084]

[0085] in, is the input at the current moment, is the hidden state at the previous moment, , is the weight matrix, is the bias term, is the Sigmoid activation function;

[0086] Forget Gate: determines how much memory from the previous moment is retained;

[0087]

[0088] Output Gate: controls the output of the current unit state;

[0089]

[0090] CellState: Combines the forget gate and the input gate to determine the memory update at the current moment;

[0091]

[0092] Hidden State: The final hidden state is determined by the output gate and the unit state;

[0093]

[0094] Introducing seasonal weights:

[0095] The occurrence of heavy precipitation is closely related to seasonal changes, especially in tropical regions and monsoon climate zones, where seasonal precipitation differences are very significant;

[0096] Seasonal Weight Design:

[0097] Design weight factors according to month (or season); assuming that The weight of the month can be determined through experience or statistical analysis, usually as follows:

[0098] Spring (March-May): moderate precipitation, heavy =1.0

[0099] Summer (June-August): heavy precipitation, weight =1.2

[0100] Autumn (September-November): precipitation gradually decreases, weight =1.0

[0101] Winter (December-February): low precipitation, weight =0.8

[0102] For each time step the input data In the LSTM model, the month of the moment The seasonal weights Applied to the current input data:

[0103]

[0104] in, According to the time step Corresponding month Sure;

[0105] The impact of seasonal weights during training:

[0106] During the training process of the LSTM model, seasonal weights will affect the model's attention to precipitation patterns in different seasons. In summer, the precipitation intensity is greater, so a higher weight is assigned. In winter, the precipitation is less, so a lower weight will make the model relatively less sensitive to precipitation in winter, thereby effectively reducing overfitting.

[0107] Furthermore, the joint training:

[0108] Preparation of input data:

[0109] The input data includes radar image sequences and seasonal weights, which are extracted by super-resolution CNN and then passed to LSTM for time series modeling;

[0110] Input radar image data is a multi-channel image containing terrain information and radar reflectivity information, represented as ,in is the time step Radar images at the moment, is the digital elevation model (DEM) data at the corresponding time;

[0111] CNN feature extraction: The CNN module is used to extract spatial features from the input radar image sequence, especially the spatial distribution and boundaries of heavy precipitation areas; after processing by the convolution layer and the pooling layer, the CNN output feature map , i.e., time step Spatial characteristics:

[0112]

[0113] in, It is a high-dimensional spatial feature vector that contains various patterns and information in the input radar image;

[0114] LSTM modeling: spatial features extracted by CNN As the input of LSTM, LSTM performs time series modeling on the spatial feature sequence and learns the time evolution pattern of precipitation; the input of LSTM is:

[0115]

[0116] in, is the output feature of CNN, is the LSTM hidden state at the previous moment, is the hidden state of the LSTM output at the current moment;

[0117] Integration of seasonal weights: For each moment , seasonal weights are added when inputting CNN and LSTM to reflect the seasonal variation of heavy precipitation; this seasonal weight weights the CNN input data so that the model can adjust its sensitivity to precipitation patterns according to seasonal characteristics; the specific approach is to multiply the weight by the input data:

[0118]

[0119] Then, CNN extracts features from the weighted data;

[0120] Joint training goal: The training goal is to optimize the parameters of CNN and LSTM by minimizing the loss function. The loss function needs to consider the errors of CNN feature extraction and LSTM time series modeling at the same time, so a joint loss function can be designed. Assuming there is a real label , which means that at the time step The combined loss function of CNN and LSTM is defined as:

[0121]

[0122] in:

[0123] is the prediction output of LSTM,

[0124] is the spatial feature output of CNN,

[0125] is the true precipitation value (or other target variable),

[0126] is the spatial feature of the annotation (such as the mark of the actual heavy rainfall area),

[0127] and is a weight hyperparameter used to balance the time series prediction error and spatial feature learning error;

[0128] Training process: During the training process, the weights of CNN and LSTM are shared, and the gradient backpropagation will be propagated throughout the entire network, including both CNN and LSTM. The specific steps are as follows:

[0129] Forward propagation: input radar image sequence, CNN extracts features, and LSTM performs time series prediction;

[0130] Loss calculation: Based on the true label and the predicted output Calculating the loss function ;

[0131] Back propagation: Calculate the gradient of the loss function with respect to the parameters of each layer of CNN and LSTM, and update the weights of CNN and LSTM through back propagation;

[0132] Weight update: The goal of joint training is to jointly update the parameters of CNN and LSTM by optimizing the loss function; the gradient descent optimization algorithm is used to adjust the model parameters:

[0133]

[0134]

[0135] in, and are the model parameters of CNN and LSTM respectively, is the learning rate, represents the gradient;

[0136] Design and weighting of loss function:

[0137] Weighted loss function: A weighted MSE loss function can be used to give higher weights to areas with heavy precipitation:

[0138]

[0139] in, is the weight of the heavy precipitation area. The weight can be designed according to the pixel ratio or intensity of the precipitation area. For example, if the heavy precipitation area at a certain moment occupies 30% of the image, the weight of the area Can be set to 3;

[0140] Seasonal weighting: Apply seasonal weights The loss function is incorporated to further enhance the model's adaptation to seasonal precipitation.

[0141] Another object of the present invention is to provide an information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning system comprising:

[0142] Construction module, used to construct radar heavy precipitation dataset;

[0143] Fusion module, used to fuse terrain information with radar images;

[0144] Design module for super-resolution CNN design; the super-resolution CNN model learns low-resolution images through convolutional neural networks and generates high-resolution images; in the heavy precipitation forecast task, radar images are used as input, combined with terrain information DEM maps, and feature extraction and boundary enhancement are performed through the super-resolution CNN network; the super-resolution CNN learns the mapping relationship between low-resolution radar images and high-resolution images;

[0145] Add a module for adding LSTM model and seasonal weights; in heavy precipitation prediction, LSTM is used to process radar image sequences or other time series data;

[0146] Training module, used for joint training.

[0147] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning method.

[0148] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning method.

[0149] Another object of the present invention is to provide an information data processing terminal, which is used to implement the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning system.

[0150] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0151] First, based on deep learning technology and the physical mechanism of heavy precipitation, this invention proposes a new short-term forecasting method for heavy precipitation in cities based on meteorological radar data through the research idea of ​​"space → time → spatiotemporal fusion". This method analyzes the spatiotemporal distribution characteristics of heavy precipitation, extracts high-dimensional spatial features, develops a super-resolution heavy precipitation spatial feature extraction model, and combines the time evolution characteristics of heavy precipitation. It uses LSTM and spatiotemporal feature fusion models to accurately forecast heavy precipitation and improve the accuracy and timeliness of urban heavy precipitation forecasts.

[0152] CNN-based super-resolution spatial feature extraction: The CNN network is used to extract spatial features from radar echo images, and the radar echo images are efficiently processed using technologies such as deep separable convolution, hole convolution, and multi-scale convolution, thereby improving the prediction accuracy of precipitation spatial distribution. In particular, super-resolution technology is used to improve the spatial resolution of radar echo images, solving the problem of blurred and unclear boundaries of traditional radar echo images.

[0153] Temporal feature modeling based on LSTM: The occurrence of heavy precipitation events is closely related to temporal features. The LSTM model is introduced to model the temporal evolution characteristics of heavy precipitation, especially to capture the short-term change trend of heavy precipitation. In order to improve the temporal feature learning ability of the model, a self-attention mechanism is also introduced in LSTM, so that the model can focus on the important time nodes of heavy precipitation changes and reduce the impact of noise.

[0154] Fusion and optimization of spatiotemporal features: After extracting spatial features (CNN) and temporal features (LSTM), weighted fusion technology is used to fuse the spatiotemporal features of the two to form a complete spatiotemporal feature vector. The fused feature vector can simultaneously reflect the spatial distribution information and the temporal evolution trend, providing a more accurate basis for the final heavy precipitation forecast. In addition, the present invention also combines meteorological physics knowledge to post-process and optimize the model output results to ensure that the forecast results conform to the actual meteorological laws.

[0155] Heavy precipitation forecast output: Through the fusion model of heavy precipitation spatiotemporal characteristics, the intensity, spatial distribution and occurrence time of heavy precipitation are ultimately output, and the results are corrected in combination with meteorological knowledge, thereby providing reliable scientific support for urban disaster prevention and mitigation.

[0156] Second, the technical solution of the present invention fills the technical gap in the industry at home and abroad: the present invention is the first to add terrain factors as physical constraints into the artificial intelligence energy model.

[0157] Whether the technical solution of the present invention overcomes technical prejudice: The present invention combines physical constraints and artificial intelligence to overcome the technical prejudice in the meteorological community that "artificial intelligence models are difficult to explain principles." BRIEF DESCRIPTION OF THE DRAWINGS

[0158] Figure 1 It is a flow chart of the deep learning method for short-term forecast of heavy rainfall in cities using dual polarization radar with information fusion provided by an embodiment of the present invention.

[0159] Figure 2 It is a structural block diagram of the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning system provided by an embodiment of the present invention.

[0160] Figure 3 It is a structural diagram of the method provided by an embodiment of the present invention.

[0161] Figure 4 It is a spatial feature extraction graph provided by an embodiment of the present invention.

[0162] Figure 5 1 is a diagram of current observation results and method prediction results provided by an embodiment of the present invention; wherein (a) is the current observation result, and (b) is the prediction result of the artificial intelligence model method. DETAILED DESCRIPTION

[0163] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0164] like Figure 1 As shown, an information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning method provided by an embodiment of the present invention includes the following steps:

[0165] S101, construct radar heavy precipitation dataset;

[0166] S102, fusion of terrain information and radar images;

[0167] S103, super-resolution CNN design;

[0168] The super-resolution CNN model uses a convolutional neural network to learn low-resolution images and generate high-resolution images. In the heavy rainfall forecasting task, radar images are used as input, combined with terrain information DEM maps, and feature extraction and boundary enhancement are performed through a super-resolution CNN network. Super-resolution CNN learns the mapping relationship between low-resolution radar images and high-resolution images.

[0169] S104, LSTM model and seasonal weights added;

[0170] In heavy precipitation prediction, LSTM is used to process radar image sequences or other time series data;

[0171] S105, joint training.

[0172] The deep learning method for short-term forecast of heavy rainfall in cities based on dual polarization radar based on information fusion provided in the embodiment of the present invention improves the forecast accuracy and timeliness of heavy rainfall in cities through a variety of deep learning technologies and information fusion means. The following is the working principle of the method:

[0173] Constructing radar heavy precipitation data set: First, the system constructs a heavy precipitation data set by collecting a large amount of radar data, including radar heavy precipitation echo information. This data set contains radar images of various heavy precipitation events and their corresponding precipitation intensity and time series. These data will be used for subsequent model training and testing. Through the accumulation and analysis of a large amount of data, the model can learn to extract the characteristics of heavy precipitation from radar images and then make accurate forecasts.

[0174] Fusion of terrain information and radar images: To improve the accuracy of heavy precipitation forecasts, the system fuses terrain information (such as DEM maps) with radar images. Terrain information provides the impact of urban landforms on precipitation, for example, precipitation in mountainous areas is different from that in plain areas. Combining terrain data with radar images allows the model to identify the impact of different geographical features on precipitation, thereby better predicting heavy precipitation in local areas. Through this information fusion, the model can more accurately forecast precipitation under complex terrain conditions.

[0175] Super-resolution CNN design: In the heavy precipitation forecasting task, a convolutional neural network (CNN) is used to learn low-resolution radar images and generate high-resolution images. The super-resolution CNN model enhances radar images by extracting the mapping relationship between low-resolution images and high-resolution images. Combined with terrain information (DEM map), the super-resolution CNN can not only extract features in the image, but also strengthen the image boundaries and provide a clearer definition of precipitation areas. This model can significantly improve the details and resolution of radar images and increase the accuracy of precipitation prediction.

[0176] LSTM model and seasonal weights added: Since precipitation is a meteorological process with strong temporal characteristics, the long short-term memory network (LSTM) is used to process radar image sequences or other time series data. LSTM can capture the temporal dynamic changes of precipitation and effectively consider the impact of historical precipitation information when forecasting heavy precipitation events. In addition, the system also introduces seasonal weights, which allows the model to adaptively adjust precipitation according to the climate characteristics of different seasons. Joint training of super-resolution CNN and LSTM models enables this method to more accurately predict the occurrence and changes of heavy precipitation in cities when processing spatiotemporal data.

[0177] Through the above steps, the deep learning method of the present invention combines radar data, terrain information and time series models, which can achieve high-precision and high-efficiency urban heavy rainfall forecasts, providing important support for emergency response and disaster warning.

[0178] The deep learning method of the present invention combines dual-polarization radar data with terrain information to achieve short-term forecasting of heavy precipitation in cities. The specific process is divided into steps such as data set construction, information fusion, super-resolution modeling, time series processing and joint training. In the data set construction stage (S101), precipitation data is extracted from dual-polarization radar images to form a multidimensional data set covering heavy precipitation scenes. This data set provides sufficient training samples for the model, covering precipitation characteristics under different precipitation intensities, geographical locations and meteorological conditions.

[0179] The introduction of terrain information (S102) is an innovation of this method. By fusing terrain elevation data (DEM) with radar images, the sensitivity of the model to local meteorological conditions is enhanced. Elevation information can significantly affect precipitation patterns, especially in cities and mountainous areas, where terrain differences can lead to local enhancement or weakening of precipitation. The fused image data contains the original radar reflectivity information and terrain features. This integration of multi-source information makes the model more adaptable to the complex underlying surface conditions in cities.

[0180] Next, a super-resolution convolutional neural network (CNN) model (S103) is used to process and improve the resolution of radar images. The super-resolution CNN gradually converts low-resolution radar data into high-resolution data output by learning the mapping relationship between low-resolution radar images and their high-resolution images. The convolution layer of the model extracts image features layer by layer to generate high-precision boundary information, which is important for capturing small precipitation areas and enhancing edge clarity, effectively improving image resolution and the accuracy of precipitation intensity prediction.

[0181] This method applies a long short-term memory network (LSTM) (S104) in time series processing. The LSTM network is suitable for processing time series data. By inputting the radar image sequence into the LSTM model, the model can identify and learn the dynamic evolution of precipitation. Especially in short-term forecasts, LSTM has a strong ability to capture the temporal changes of precipitation intensity. In addition, the addition of seasonal weights further improves the prediction accuracy of the model. The weight setting of different seasons enables the model to automatically adjust its judgment in different seasons, thereby enhancing its adaptability to seasonal precipitation patterns.

[0182] In the joint training stage of the model (S105), the super-resolution CNN and LSTM models are trained simultaneously so that the two can work together effectively. The spatial features extracted by the super-resolution CNN are used as the input of the LSTM. After combining the seasonal weights, the model makes a comprehensive prediction of the precipitation intensity in the time dimension. Joint training enables the model to not only improve the spatial resolution, but also use time series information to accurately predict the intensity changes of precipitation in the short term, thereby achieving efficient short-term forecasting of heavy precipitation.

[0183] In summary, the present invention forms a precipitation forecast system with high precision and timeliness through the combined application of dual-polarization radar, terrain information, super-resolution CNN and LSTM network.

[0184] The radar heavy precipitation data set provided in the embodiment of the present invention is:

[0185] (1) Data collection

[0186] Radar images: Raw radar reflectivity images obtained from weather radar systems, usually based on high-frequency observations over a short period of time;

[0187] Precipitation observation data: using ground meteorological stations or satellite data to collect field observation data of heavy precipitation events, including precipitation amount and precipitation intensity indicators;

[0188] (2) Data labeling

[0189] Heavy precipitation area annotation: using observation data and radar images to define heavy precipitation areas;

[0190] (3) Data preprocessing

[0191] Image slicing: divide radar images into small pieces; train network models; each image should be labeled with whether it belongs to a heavy rainfall area;

[0192] Data enhancement: Use data enhancement technology to generate more training samples;

[0193] Normalization: Radar images and precipitation intensities are normalized, usually scaling the values ​​to the range of [0, 1] or [-1, 1].

[0194] The terrain information provided by the embodiment of the present invention is integrated with the radar image:

[0195] Fusion method design:

[0196] Fusion of input data: Combine radar images with terrain elevation maps DEM to form multi-channel input:

[0197] Radar data processing:

[0198] Get radar image: radar data is presented in the Z field of reflectivity map, indicating the intensity of the target; radar image is denoised; suppose the denoised radar image is , whose shape is ,in is the image height, is the image width, is the number of channels of the radar image;

[0199] DEM data processing:

[0200] Get the digital elevation model (DEM) map: DEM maps represent the elevation information of the surface, usually stored in matrix form, indicating the elevation value of each geographic point; DEM data needs to be aligned with the radar image, usually matched by interpolation methods to ensure that the resolution of the two is consistent; assuming that after interpolation processing, the dimension of the DEM map is ,in , represents the terrain height value of each point;

[0201] Fusion of radar map and DEM map:

[0202] Weighted fusion: The radar image and DEM image can be fused in a weighted manner, while retaining the terrain information and precipitation characteristics of the radar image; the formula for weighted fusion is:

[0203]

[0204] in, is a hyperparameter used to control the fusion ratio of radar images and DEM images; by choosing the appropriate , the influence of terrain characteristics on the spatial characteristics of heavy precipitation can be adjusted;

[0205] Splicing and fusion: Splice the radar map and DEM image in the channel dimension to generate a multi-channel input feature map; this method explicitly fuses the terrain features and precipitation features at the input layer of the model:

[0206]

[0207] Among them, the stitched image The number of channels is 2, which contains radar reflectivity and terrain information;

[0208] Assume that the size of the radar image is , the topographic map (DEM) has the same size, so we can directly concatenate the two to get a three-channel input:

[0209]

[0210] in, is the radar image, It is a terrain elevation map, the stitched image Has 3 channels.

[0211] The super-resolution CNN design provided by the embodiment of the present invention:

[0212] Network structure:

[0213] The super-resolution CNN model generally consists of three main parts: feature extraction (convolution layer), nonlinear mapping (activation function) and image restoration (deconvolution layer);

[0214] Input layer: The input image size is W×H×C,

[0215] Where W and H are the width and height of the image, and C is the number of channels of the input image;

[0216] Convolutional layers: Multiple convolutional layers are used to extract spatial features of the input image; each convolutional layer is followed by an activation function and batch normalization;

[0217] Upsampling layer: restore low-resolution images to high resolution using deconvolution layers or pixel rearrangement layers; deconvolution layers spatially expand the resolution of images while maintaining the spatial structure in the image;

[0218] Output layer: The output layer usually uses a convolutional layer to generate a super-resolution image; at this point, the image size has been restored to the target high resolution;

[0219] Network details and layer configuration:

[0220] Input layer: Input image of size 64×64×3; the image contains a fusion map of radar image and terrain information;

[0221] Convolutional Layer:

[0222] Layer 1: convolution kernel size 3×3, stride 1, number of convolution kernels 64, activation function uses ReLU;

[0223] Layer 2: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 128, and the activation function uses ReLU;

[0224] Layer 3: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 256, and the activation function uses ReLU;

[0225] Deconvolution layer:

[0226] Use the deconvolution layer to upsample the image; assuming the goal is to restore to high resolution, set the deconvolution kernel size to 4×4 and the step size to 2;

[0227] Layer 1: Deconvolution kernel size 4×4, stride 2, output channels 128;

[0228] Layer 2: Deconvolution kernel size 4×4, stride 2, output channels 64;

[0229] Output layer:

[0230] The output layer uses a convolution layer to restore the super-resolution image to the target image, with an output channel of 1; the output size is 128×128×1;

[0231] Solve the problem of blurred boundaries of heavy precipitation:

[0232] The boundaries of heavy rainfall areas are usually fuzzy. Super-resolution CNN can accurately reconstruct these fuzzy areas by learning detailed features.

[0233] Convolutional layer extracts features:

[0234]

[0235] Each layer of convolution operation extracts the spatial features of the input image and enhances edge details;

[0236] Deconvolution: Upsample the low-resolution image to a higher resolution through transposed convolution to restore the details in the image, especially the boundary area of ​​precipitation:

[0237] .

[0238] The LSTM model provided by the embodiment of the present invention is added with seasonal weights:

[0239] The basic architecture of the LSTM model:

[0240] The core of the LSTM model is its internal cell structure, which contains three main gating mechanisms (input gate, forget gate, and output gate), and a memory unit; the LSTM model controls the flow of information through these gating mechanisms and maintains memory of previous inputs over a long time scale; the basic update formula of LSTM is as follows:

[0241] Input Gate: controls the impact of current input information;

[0242]

[0243] in, is the input at the current moment, is the hidden state at the previous moment, , is the weight matrix, is the bias term, is the Sigmoid activation function;

[0244] Forget Gate: determines how much memory from the previous moment is retained;

[0245]

[0246] Output Gate: controls the output of the current unit state;

[0247]

[0248] CellState: Combines the forget gate and the input gate to determine the memory update at the current moment;

[0249]

[0250] Hidden State: The final hidden state is determined by the output gate and the unit state;

[0251]

[0252] Introducing seasonal weights:

[0253] The occurrence of heavy precipitation is closely related to seasonal changes, especially in tropical regions and monsoon climate zones, where seasonal precipitation differences are very significant;

[0254] Seasonal Weight Design:

[0255] Design weight factors according to month (or season); assuming that The weight of the month can be determined through experience or statistical analysis, usually as follows:

[0256] Spring (March-May): moderate precipitation, heavy =1.0

[0257] Summer (June-August): heavy precipitation, weight =1.2

[0258] Autumn (September-November): precipitation gradually decreases, weight =1.0

[0259] Winter (December-February): low precipitation, weight =0.8

[0260] For each time step the input data In the LSTM model, the month of the moment The seasonal weights Applied to the current input data:

[0261]

[0262] in, According to the time step Corresponding month Sure;

[0263] The impact of seasonal weights during training:

[0264] During the training process of the LSTM model, seasonal weights will affect the model's attention to precipitation patterns in different seasons. In summer, the precipitation intensity is greater, so a higher weight is assigned. In winter, the precipitation is less, so a lower weight will make the model relatively less sensitive to precipitation in winter, thereby effectively reducing overfitting.

[0265] The joint training provided by the embodiment of the present invention:

[0266] Preparation of input data:

[0267] The input data includes radar image sequences and seasonal weights, which are extracted by super-resolution CNN and then passed to LSTM for time series modeling;

[0268] Input radar image data is a multi-channel image containing terrain information and radar reflectivity information, represented as ,in is the time step Radar images at the moment, is the digital elevation model (DEM) data at the corresponding time;

[0269] CNN feature extraction: The CNN module is used to extract spatial features from the input radar image sequence, especially the spatial distribution and boundaries of heavy precipitation areas; after processing by the convolution layer and the pooling layer, the CNN output feature map , i.e., time step Spatial characteristics:

[0270]

[0271] in, It is a high-dimensional spatial feature vector that contains various patterns and information in the input radar image;

[0272] LSTM modeling: spatial features extracted by CNN As the input of LSTM, LSTM performs time series modeling on the spatial feature sequence and learns the time evolution pattern of precipitation; the input of LSTM is:

[0273]

[0274] in, is the output feature of CNN, is the LSTM hidden state at the previous moment, is the hidden state of the LSTM output at the current moment;

[0275] Integration of seasonal weights: For each moment , seasonal weights are added when inputting CNN and LSTM to reflect the seasonal variation of heavy precipitation; this seasonal weight weights the CNN input data so that the model can adjust its sensitivity to precipitation patterns according to seasonal characteristics; the specific approach is to multiply the weight by the input data:

[0276]

[0277] Then, CNN extracts features from the weighted data;

[0278] Joint training goal: The training goal is to optimize the parameters of CNN and LSTM by minimizing the loss function. The loss function needs to consider the errors of CNN feature extraction and LSTM time series modeling at the same time, so a joint loss function can be designed. Assuming there is a real label , which means that at the time step The combined loss function of CNN and LSTM is defined as:

[0279]

[0280] in:

[0281] is the prediction output of LSTM,

[0282] is the spatial feature output of CNN,

[0283] is the true precipitation value (or other target variable),

[0284] is the spatial feature of the annotation (such as the mark of the actual heavy rainfall area),

[0285] and is a weight hyperparameter used to balance the time series prediction error and spatial feature learning error;

[0286] Training process: During the training process, the weights of CNN and LSTM are shared, and the gradient backpropagation will be propagated throughout the entire network, including both CNN and LSTM. The specific steps are as follows:

[0287] Forward propagation: input radar image sequence, CNN extracts features, and LSTM performs time series prediction;

[0288] Loss calculation: Based on the true label and the predicted output Calculating the loss function ;

[0289] Back propagation: Calculate the gradient of the loss function with respect to the parameters of each layer of CNN and LSTM, and update the weights of CNN and LSTM through back propagation;

[0290] Weight update: The goal of joint training is to jointly update the parameters of CNN and LSTM by optimizing the loss function; the gradient descent optimization algorithm is used to adjust the model parameters:

[0291]

[0292]

[0293] in, and are the model parameters of CNN and LSTM respectively, is the learning rate, represents the gradient;

[0294] Design and weighting of loss function:

[0295] Weighted loss function: A weighted MSE loss function can be used to give higher weights to areas with heavy precipitation:

[0296]

[0297] in, is the weight of the heavy precipitation area. The weight can be designed according to the pixel ratio or intensity of the precipitation area. For example, if the heavy precipitation area at a certain moment occupies 30% of the image, the weight of the area Can be set to 3;

[0298] Seasonal weighting: Apply seasonal weights The loss function is incorporated to further enhance the model's adaptation to seasonal precipitation.

[0299] like Figure 2 As shown, an information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning system provided by an embodiment of the present invention includes:

[0300] Construction module, used to construct radar heavy precipitation dataset;

[0301] Fusion module, used to fuse terrain information with radar images;

[0302] Design module for super-resolution CNN design; the super-resolution CNN model learns low-resolution images through convolutional neural networks and generates high-resolution images; in the heavy precipitation forecast task, radar images are used as input, combined with terrain information DEM maps, and feature extraction and boundary enhancement are performed through the super-resolution CNN network; the super-resolution CNN learns the mapping relationship between low-resolution radar images and high-resolution images;

[0303] Add a module for adding LSTM model and seasonal weights; in heavy precipitation prediction, LSTM is used to process radar image sequences or other time series data;

[0304] Training module, used for joint training.

[0305] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning method.

[0306] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning method.

[0307] Another object of the present invention is to provide an information data processing terminal, which is used to implement the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning system.

[0308] The present invention is specifically implemented:

[0309] like Figure 3 As shown:

[0310] 1. Build a radar heavy precipitation dataset

[0311] 1.1 Data Collection

[0312] Radar images: Raw radar reflectivity images are obtained from weather radar systems, usually based on high-frequency observations over a short period of time. These images have a certain resolution (such as 1 km or less) and can provide real-time reflectivity information in areas of heavy precipitation.

[0313] Precipitation observation data: Using ground-based meteorological stations or satellite data, collect field observation data of heavy precipitation events, including indicators such as precipitation amount and precipitation intensity. Precipitation intensity information for multiple periods can be constructed through records at different time scales, such as hours, days, and months.

[0314] 1.2 Data Annotation

[0315] Heavy precipitation area labeling: Use observation data and radar images to define heavy precipitation areas. For example, when the radar reflectivity is greater than 35dBZ, it can be defined as a heavy precipitation area. In addition, the precipitation data of the weather station can be used to label the heavy precipitation area in each radar image as a label for the supervised learning task.

[0316] 1.3 Data Preprocessing

[0317] Image slicing: Divide radar images into small blocks (such as 64×64 or 128×128 sub-images) to facilitate training of network models. Each image should be labeled whether it belongs to a heavy rainfall area.

[0318] Data enhancement: To enhance the generalization ability of the model, data enhancement techniques such as rotation, flipping, cropping, scaling, etc. can be used to generate more training samples.

[0319] Normalization: Normalize the radar images and precipitation intensity, usually scaling the values ​​to the range of [0, 1] or [-1, 1] to ensure that the model training process is more stable.

[0320] 2. Fusion of terrain information and radar images

[0321] 2.1 Fusion method design

[0322] Fusion of input data: Combine the radar image with the terrain elevation map (DEM) to form a multi-channel input. Assuming that the size of the radar image is W×H and the size of the terrain map (DEM) is the same, the two can be directly spliced ​​to obtain a three-channel input:

[0323] 2.1.1 Radar data processing

[0324] Get radar image: Usually, radar data is presented as a reflectivity map (Z field), which indicates the intensity of the target. Radar images may contain noise and need to be denoised first. Common denoising methods include median filtering or convolutional neural network (CNN) denoising. Assume that the denoised radar image is , whose shape is ,in is the image height, is the image width, is the channel number of the radar image (usually 1, indicating a single-channel radar reflectivity image).

[0325] 2.1.2 DEM data processing

[0326] Get a digital elevation model (DEM) map: DEM maps represent the elevation information of the surface, usually stored in matrix form, representing the elevation value of each geographic point. DEM data needs to be aligned with radar images, usually through interpolation methods to ensure that the resolution of the two is consistent. Assume that after interpolation, the dimension of the DEM map is ,in , indicating the terrain height value of each point.

[0327] 2.2 Fusion of radar map and DEM map

[0328] Weighted fusion: The radar image and DEM image can be fused in a weighted manner, so that the terrain information and precipitation characteristics of the radar image can be retained at the same time. The formula for weighted fusion is:

[0329]

[0330] in, is a hyperparameter used to control the fusion ratio of radar image and DEM image. , the impact of terrain characteristics on the spatial characteristics of heavy precipitation can be adjusted.

[0331] Splicing and fusion: Splice the radar map and DEM image in the channel dimension to generate a multi-channel input feature map. This method explicitly fuses terrain features and precipitation features at the input layer of the model:

[0332]

[0333] Among them, the stitched image The number of channels is 2, which includes radar reflectivity and terrain information.

[0334] Through the above integration, we can:

[0335] Enhance the spatial understanding ability of the model: Topographic information can help the model understand the relationship between heavy precipitation and terrain changes, such as the difference in precipitation between mountainous areas and plains.

[0336] Spatial context information: Topographic information has a significant impact on the distribution and morphology of heavy precipitation, especially in mountainous areas. The topography is closely related to the precipitation process and can help improve the identification and prediction of heavy precipitation.

[0337] like Figure 5 , assuming the size of the radar image is , the topographic map (DEM) has the same size, so we can directly concatenate the two to get a three-channel input:

[0338]

[0339] in, is the radar image, It is a terrain elevation map, the stitched image With 3 channels (radar intensity, DEM, elevation features).

[0340] 3. Super-resolution CNN design

[0341] like Figure 4 The core goal of the super-resolution CNN (SRCNN) model is to learn low-resolution images through convolutional neural networks and generate high-resolution images. In the heavy precipitation forecasting task, radar images are used as input, combined with terrain information (DEM images), and feature extraction and boundary enhancement are performed through the super-resolution CNN network. By learning the mapping relationship between low-resolution radar images and high-resolution images, the super-resolution CNN can enhance the spatial details in the image, especially the clarity of the heavy precipitation boundary.

[0342] 3.1 Overview of Network Structure

[0343] The super-resolution CNN model generally consists of three main parts: feature extraction (convolution layer), nonlinear mapping (activation function) and image restoration (deconvolution layer). The following is the detailed design of each part:

[0344] Input layer: The input image size is W×H×C,

[0345] Where W and H are the width and height of the image, and C is the number of channels of the input image (e.g. 3 channels: radar image + terrain elevation map + season information).

[0346] Convolutional layer (FeatureExtractionLayer): Multiple convolutional layers are used to extract the spatial features of the input image. Each convolutional layer is followed by an activation function (ReLU) and batch normalization (BatchNormalization). The role of these convolutional layers is to learn the detailed features of the radar image, especially the blurred areas at the boundaries of heavy precipitation.

[0347] Upsampling Layer: Use the Deconvolution layer or the PixelShuffle layer to restore the low-resolution image to a high-resolution one. The Deconvolution layer spatially expands the resolution of the image while maintaining the spatial structure in the image.

[0348] Specifically, the deconvolution operation is to reversely generate a higher resolution image through the convolution operation.

[0349] Output layer (ReconstructionLayer): The output layer usually uses a convolutional layer to generate a super-resolution image. At this point, the size of the image has been restored to the target high resolution.

[0350] 3.2 Network Details and Layer Configuration

[0351] Input layer: Input image of size 64×64×3 (64x64 pixels, 3 channels). The image contains a fusion map of radar image and terrain information.

[0352] Convolutional layer (feature extraction):

[0353] Layer 1: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 64, and the activation function is ReLU.

[0354] Layer 2: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 128, and the activation function is ReLU.

[0355] Layer 3: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 256, and the activation function is ReLU.

[0356] Deconvolution layer (image restoration):

[0357] Use the deconvolution layer to upsample the image. Assuming the goal is to restore to high resolution, set the deconvolution kernel size to 4×4 and the stride to 2.

[0358] Layer 1: Deconvolution kernel size 4×4, stride 2, output channels 128.

[0359] Layer 2: Deconvolution kernel size 4×4, stride 2, and 64 output channels.

[0360] Output layer:

[0361] The output layer uses a convolution layer to restore the super-resolution image to the target image, and the output channel is 1 (single-channel output, i.e. predicted radar reflectivity). The output size is 128×128×1.

[0362] 3.3 Solving the problem of blurred boundaries of heavy precipitation

[0363] The boundaries of heavy rainfall areas are usually blurred. Super-resolution CNN can accurately reconstruct these blurred areas by learning detailed features (such as edges and textures).

[0364] Convolutional layer extracts features:

[0365]

[0366] Each layer of convolution operation extracts the spatial features of the input image and enhances edge details.

[0367] Deconvolution (upsampling): The low-resolution image is upsampled to a higher resolution through transposed convolution (Deconvolution) to restore the details in the image, especially the boundary area of ​​precipitation:

[0368]

[0369] This upsampling method can improve the boundary identification accuracy of heavy precipitation areas.

[0370] 4. LSTM model with seasonal weights added

[0371] LSTM (Long Short-Term Memory) is a deep learning model widely used in time series forecasting, which can effectively capture long-term dependencies in data. In heavy precipitation forecasting, LSTM can be used to process radar image sequences or other time series data, and improve forecast accuracy by capturing the temporal pattern of precipitation. In order to further improve the effect of heavy precipitation forecasting, seasonal weights are introduced to take into account the seasonal variation characteristics of heavy precipitation.

[0372] 4.1 Basic Architecture of LSTM Model

[0373] The core of the LSTM model is its internal cell structure, which contains three main gating mechanisms (input gate, forget gate, and output gate), as well as a memory unit. The LSTM model controls the flow of information through these gating mechanisms and maintains memory of previous inputs over a long time scale. The basic update formula of LSTM is as follows:

[0374] Input Gate: Controls the influence of current input information.

[0375]

[0376] in, is the input at the current moment, is the hidden state at the previous moment, , is the weight matrix, is the bias term, is the Sigmoid activation function.

[0377] Forget Gate: Determines how much memory from the previous moment to retain.

[0378]

[0379] Output Gate: controls the output of the current unit state.

[0380]

[0381] CellState: Combined with the forget gate and the input gate, it determines the memory update at the current moment.

[0382]

[0383] Hidden State: The final hidden state is determined by the output gate and the unit state.

[0384]

[0385] 4.2 Introducing seasonal weights

[0386] The occurrence of heavy precipitation is closely related to seasonal changes, especially in tropical regions and monsoon climate zones, where seasonal precipitation differences are very significant. For example, precipitation is high in summer and low in winter. In order to improve the model's adaptability to seasonal changes, seasonal weights are introduced, that is, a weight factor is assigned to each month's precipitation data. , which enables the model to automatically focus on seasonal characteristics during training.

[0387] 4.2.1 Seasonal Weight Design

[0388] Design weight factors based on month (or season). Assume that The weight of the month can be determined through experience or statistical analysis, usually as follows:

[0389] Spring (March-May): moderate precipitation, heavy =1.0

[0390] Summer (June-August): heavy precipitation, weight =1.2

[0391] Autumn (September-November): precipitation gradually decreases, weight =1.0

[0392] Winter (December-February): low precipitation, weight =0.8

[0393] For example, for each time step of input data In the LSTM model, the month of the moment The seasonal weights Applied to the current input data:

[0394]

[0395] in, According to the time step Corresponding month Sure.

[0396] 4.2.2 Effect of seasonal weights during training

[0397] During the training process of the LSTM model, seasonal weights will affect the model's attention to precipitation patterns in different seasons. The precipitation intensity in summer is greater, so a higher weight is given to help the model learn the characteristics of heavy precipitation in summer; while the precipitation in winter is less, a lower weight will make the model less sensitive to precipitation in winter, thereby effectively reducing overfitting.

[0398] 5. Basic process of joint training

[0399] 5.1. Preparation of input data:

[0400] The input data includes radar image sequences (for example, radar reflectivity maps for multiple consecutive hours) and seasonal weights. After feature extraction by super-resolution CNN, the data is passed to LSTM for time series modeling.

[0401] Input radar image data is a multi-channel image containing terrain information and radar reflectivity information, represented as ,in is the time step Radar images at the moment, is the digital elevation model (DEM) data at the corresponding time.

[0402] 5.2. CNN feature extraction: The CNN module is used to extract spatial features from the input radar image sequence, especially the spatial distribution and boundaries of heavy precipitation areas. After processing by the convolution layer and the pooling layer, the CNN output feature map , i.e., time step Spatial characteristics:

[0403]

[0404] in, It is a high-dimensional spatial feature vector that contains various patterns and information in the input radar image (such as the boundary information of heavy rainfall areas).

[0405] 5.3.LSTM modeling: spatial features extracted by CNN As the input of LSTM, LSTM performs temporal modeling on the spatial feature sequence and learns the temporal evolution pattern of precipitation. The input of LSTM is:

[0406]

[0407] in, is the output feature of CNN, is the LSTM hidden state at the previous moment, is the hidden state of the LSTM output at the current moment.

[0408] 5.4. Integration of seasonal weights: For each moment , seasonal weights are added when inputting CNN and LSTM to reflect the seasonal variation of heavy precipitation. This seasonal weight weights the CNN input data so that the model can adjust its sensitivity to precipitation patterns according to seasonal characteristics. The specific method is to multiply the weight by the input data:

[0409]

[0410] Then, CNN performs feature extraction on the weighted data.

[0411] 5.5. Joint training objective: The training objective is to optimize the parameters of CNN and LSTM by minimizing the loss function. The loss function needs to consider the errors of CNN feature extraction and LSTM time series modeling at the same time, so a joint loss function can be designed. Assuming there is a true label , which means that at the time step The combined loss function of CNN and LSTM is defined as:

[0412]

[0413] in:

[0414] is the prediction output of LSTM,

[0415] is the spatial feature output of CNN,

[0416] is the true precipitation value (or other target variable),

[0417] is the spatial feature of the annotation (such as the mark of the actual heavy rainfall area),

[0418] and is a weight hyperparameter used to balance the time series prediction error and spatial feature learning error.

[0419] Through the joint loss function, the goal of CNN is to extract spatial features that can help LSTM prediction, while the goal of LSTM is to make accurate time series predictions based on these features.

[0420] 5.6. Training process: During the training process, the weights of CNN and LSTM are shared, and the gradient backpropagation will be propagated throughout the entire network, including both CNN and LSTM. The specific steps are as follows:

[0421] Forward propagation: input radar image sequence, CNN extracts features, and LSTM performs time series prediction.

[0422] Loss calculation: Based on the true label and the predicted output Calculating the loss function .

[0423] Backpropagation: Calculate the gradient of the loss function with respect to the parameters of each layer of CNN and LSTM, and update the weights of CNN and LSTM through backpropagation.

[0424] 5.7. Weight Update: The goal of joint training is to jointly update the parameters of CNN and LSTM by optimizing the loss function. Use a gradient descent optimization algorithm (such as Adam or SGD) to adjust the model parameters:

[0425]

[0426]

[0427] in, and are the model parameters of CNN and LSTM respectively, is the learning rate, Represents the gradient.

[0428] 5.8 Design and Weighting of Loss Function

[0429] In the prediction of heavy rainfall, heavy rainfall areas usually account for a small proportion in radar images and their boundaries are blurred. Therefore, the loss function needs to be weighted during training so that the model pays more attention to these low-proportion and high-intensity areas.

[0430] 1) Weighted loss function: A weighted MSE loss function can be used to give higher weights to areas with heavy precipitation:

[0431]

[0432] in, is the weight of the heavy precipitation area. The weight can be designed based on the pixel ratio or intensity of the precipitation area. For example, if the heavy precipitation area at a certain moment occupies 30% of the image, the weight of the area Can be set to 3.

[0433] 2) Seasonal weighting: As mentioned above, seasonal weights can be The loss function is incorporated to further enhance the model's adaptation to seasonal precipitation.

[0434] In this way, the joint training of LSTM and CNN can adaptively handle different precipitation scenarios and improve the prediction accuracy in heavy precipitation areas, especially when the precipitation boundaries are fuzzy and the proportion is low, ensuring that the model pays more attention to the key areas where precipitation occurs.

[0435] Short-term forecasts mainly focus on small and medium-scale weather, especially severe convective weather, and aim to prevent emergency local severe storms and other disastrous weather, and meet the meteorological needs of various economic sectors and social public activities. At present, there are many examples of using short-term forecasts and successfully improving the ability to warn of disastrous weather. For example, the "Baoshan City Meteorological Monitoring and Early Warning Comprehensive Service System" and "Baoshan City Severe Convective Weather Short-term Early Warning System" independently developed by Baoshan Meteorological Observatory in Yunnan Province have been fully promoted and applied in the city's five county-level meteorological stations, significantly improving the overall forecasting ability and level of the city's meteorological department;

[0436] The 1-hour cycle system of the "RMAPS-ST" short-term numerical forecast subsystem (RMAPS-ST) based on radar data developed by the Beijing Meteorological Department has officially been put into real-time business trial operation. The commissioning of this system marks an important step for Beijing in the field of short-term weather forecasting. Through high-frequency data updates and real-time processing, RMAPS-ST can provide more accurate and timely short-term weather forecasts, especially in dealing with sudden severe convective weather. This system not only improves the forecast accuracy and timeliness of the meteorological department, but also provides strong support for urban disaster prevention and mitigation, public safety, and economic and social development.

[0437] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0438] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A deep learning method for short-term forecast of heavy rainfall in cities based on dual polarization radar with information fusion, characterized in that: The following steps are involved: S101, constructing a radar heavy precipitation data set to obtain radar data of heavy precipitation scenes; S102, integrating the terrain information with the radar image to obtain a radar image containing terrain features; S103, using a super-resolution convolutional neural network to learn the low-resolution radar image to generate a high-resolution radar image; S104, processing radar image sequence data through long short-term memory network and adding seasonal weights for time series analysis; S105, jointly train the super-resolution CNN and LSTM models to obtain a comprehensive prediction model.

2. The information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning method as claimed in claim 1 is characterized in that: The radar heavy precipitation dataset is constructed as follows: (1) Data collection Radar images: Raw radar reflectivity images obtained from weather radar systems, usually based on high-frequency observations over a short period of time; Precipitation observation data: using ground meteorological stations or satellite data to collect field observation data of heavy precipitation events, including precipitation amount and precipitation intensity indicators; (2) Data labeling Heavy precipitation area annotation: using observation data and radar images to define heavy precipitation areas; (3) Data preprocessing Image slicing: divide radar images into small blocks; Train the network model; each image should be labeled whether it belongs to a heavy rainfall area; Data enhancement: Use data enhancement technology to generate more training samples; Normalization: Radar images and precipitation intensities are normalized, usually scaling the values ​​to the range of [0, 1] or [-1, 1].

3. The information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning method as claimed in claim 1 is characterized in that: The terrain information is fused with the radar image: Fusion method design: Fusion of input data: Combine radar images with terrain elevation maps DEM to form multi-channel input: Radar data processing: Get radar image: radar data is presented in the Z field of reflectivity map, indicating the intensity of the target; radar image is denoised; suppose the denoised radar image is , whose shape is ,in is the image height, is the image width, is the number of channels of the radar image; DEM data processing: Get the digital elevation model (DEM) map: DEM maps represent the elevation information of the surface, usually stored in matrix form, indicating the elevation value of each geographic point; DEM data needs to be aligned with the radar image, usually matched by interpolation methods to ensure that the resolution of the two is consistent; assuming that after interpolation processing, the dimension of the DEM map is ,in , represents the terrain height value of each point; Fusion of radar map and DEM map: Weighted fusion: The radar image and DEM image can be fused in a weighted manner, while retaining the terrain information and precipitation characteristics of the radar image; The formula for weighted fusion is: ; in, is a hyperparameter used to control the fusion ratio of radar images and DEM images; by choosing the appropriate , the influence of terrain characteristics on the spatial characteristics of heavy precipitation can be adjusted; Splicing and fusion: Splice the radar map and DEM image in the channel dimension to generate a multi-channel input feature map; this method explicitly fuses the terrain features and precipitation features at the input layer of the model: ; Among them, the stitched image The number of channels is 2, which contains radar reflectivity and terrain information; Assume that the size of the radar image is , the size of the topographic map is the same as that of the topographic map, and the two can be directly concatenated to obtain a three-channel input: ; in, is the radar image, It is a terrain elevation map, the stitched image Has 3 channels.

4. The information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning method as claimed in claim 1 is characterized in that: The super-resolution CNN design: Network structure: The super-resolution CNN model generally consists of three main parts: feature extraction, nonlinear mapping, and image restoration; Input layer: The input image size is W×H×C, Where W and H are the width and height of the image, and C is the number of channels of the input image; Convolutional layers: Multiple convolutional layers are used to extract spatial features of the input image; each convolutional layer is followed by an activation function and batch normalization; Upsampling layer: Use deconvolution layer or pixel rearrangement layer to restore low-resolution image to high resolution; Deconvolutional layers spatially expand the resolution of the image while maintaining the spatial structure in the image; Output layer: The output layer usually uses a convolutional layer to generate a super-resolution image; at this point, the image size has been restored to the target high resolution; Network details and layer configuration: Input layer: Input image of size 64×64×3; the image contains a fusion map of radar image and terrain information; Convolutional Layer: Layer 1: convolution kernel size 3×3, stride 1, number of convolution kernels 64, activation function uses ReLU; Layer 2: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 128, and the activation function uses ReLU; Layer 3: The convolution kernel size is 3×3, the stride is 1, the number of convolution kernels is 256, and the activation function uses ReLU; Deconvolution layer: Use deconvolution layers to upsample the image; Assuming the goal is to restore to high resolution, set the deconvolution kernel size to 4×4 and the step size to 2; Layer 1: Deconvolution kernel size 4×4, stride 2, output channels 128; Layer 2: Deconvolution kernel size 4×4, stride 2, output channels 64; Output layer: The output layer uses a convolution layer to restore the super-resolution image to the target image, with an output channel of 1; the output size is 128×128×1; Solve the problem of blurred boundaries of heavy precipitation: The boundaries of heavy rainfall areas are usually fuzzy. Super-resolution CNN can accurately reconstruct these fuzzy areas by learning detailed features. Convolutional layer extracts features: ; Each layer of convolution operation extracts the spatial features of the input image and enhances edge details; Deconvolution: Upsample the low-resolution image to a higher resolution through transposed convolution to restore the details in the image, especially the boundary area of ​​precipitation: 。 5. The information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning method as claimed in claim 1 is characterized in that: The LSTM model is added with seasonal weights: The basic architecture of the LSTM model: The core of the LSTM model is its internal cell structure, which contains three main gating mechanisms, namely the input gate, forget gate and output gate, and a memory unit; the LSTM model controls the flow of information through these gating mechanisms and maintains memory of previous inputs over a long time scale; the basic update formula of LSTM is as follows: Input Gate: controls the impact of current input information; ; in, is the input at the current moment, is the hidden state at the previous moment, , is the weight matrix, is the bias term, is the Sigmoid activation function; Forget Gate: determines how much memory from the previous moment is retained; ; Output Gate: controls the output of the current unit state; ; CellState: Combines the forget gate and the input gate to determine the memory update at the current moment; ; Hidden State: The final hidden state is determined by the output gate and the unit state; ; Introducing seasonal weights: The occurrence of heavy precipitation is closely related to seasonal changes, especially in tropical regions and monsoon climate zones, where seasonal precipitation differences are very significant; Seasonal Weight Design: Design weight factors based on month or season; assuming The weight of the month can be determined through experience or statistical analysis, usually as follows: Spring March-May: moderate precipitation, heavy =1.0; Summer June-August: Heavy rainfall, heavy =1.2; Autumn September-November: precipitation gradually decreases, weight =1.0; Winter December-February: Low precipitation, weight =0.8; For each time step the input data In the LSTM model, the month of the moment The seasonal weight Applied to the current input data: ; in, According to the time step Corresponding month Sure; The impact of seasonal weights during training: During the training process of the LSTM model, seasonal weights will affect the model's attention to precipitation patterns in different seasons. In summer, the precipitation intensity is greater, so a higher weight is assigned. In winter, the precipitation is less, so a lower weight will make the model relatively less sensitive to precipitation in winter, thereby effectively reducing overfitting.

6. The information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning method as claimed in claim 1 is characterized in that: The joint training: Preparation of input data: The input data includes radar image sequences and seasonal weights, which are extracted by super-resolution CNN and then passed to LSTM for time series modeling; Input radar image data is a multi-channel image containing terrain information and radar reflectivity information, represented as ,in is the time step Radar images at the moment, is the digital elevation model (DEM) data at the corresponding time; CNN feature extraction: The CNN module is used to extract spatial features from the input radar image sequence, especially the spatial distribution and boundaries of heavy precipitation areas; after processing by the convolution layer and the pooling layer, the CNN output feature map , i.e., time step Spatial characteristics: ; in, It is a high-dimensional spatial feature vector that contains various patterns and information in the input radar image; LSTM modeling: spatial features extracted by CNN As the input of LSTM, LSTM performs time series modeling on the spatial feature sequence and learns the time evolution pattern of precipitation; the input of LSTM is: ; in, is the output feature of CNN, is the LSTM hidden state at the previous moment, is the hidden state of the LSTM output at the current moment; Integration of seasonal weights: For each moment , seasonal weights are added when inputting CNN and LSTM to reflect the seasonal variation of heavy precipitation; this seasonal weight weights the CNN input data so that the model can adjust its sensitivity to precipitation patterns according to seasonal characteristics; the specific approach is to multiply the weight by the input data: ; Then, CNN extracts features from the weighted data; Joint training goal: The training goal is to optimize the parameters of CNN and LSTM by minimizing the loss function. The loss function needs to consider the errors of CNN feature extraction and LSTM time series modeling at the same time, so a joint loss function can be designed. Assuming there is a real label , which means that at the time step The combined loss function of CNN and LSTM is defined as: ; in: is the prediction output of LSTM, is the spatial feature output of CNN, is the actual precipitation value, is the spatial feature of the annotation, and is a weight hyperparameter used to balance the time series prediction error and spatial feature learning error; Training process: During the training process, the weights of CNN and LSTM are shared, and the gradient backpropagation will be propagated throughout the entire network, including both CNN and LSTM parts; the specific steps are as follows: Forward propagation: input radar image sequence, CNN extracts features, and LSTM performs time series prediction; Loss calculation: Based on the true label and the predicted output Calculating the loss function ; Back propagation: Calculate the gradient of the loss function with respect to the parameters of each layer of CNN and LSTM, and update the weights of CNN and LSTM through back propagation; Weight update: The goal of joint training is to jointly update the parameters of CNN and LSTM by optimizing the loss function; the gradient descent optimization algorithm is used to adjust the model parameters: ; ; in, and are the model parameters of CNN and LSTM respectively, is the learning rate, represents the gradient; Design and weighting of loss function: Weighted loss function: A weighted MSE loss function can be used to give higher weights to areas with heavy precipitation: ; in, is the weight of the heavy precipitation area, and the weight is designed according to the pixel ratio or intensity of the precipitation area; Seasonal weighting: Apply seasonal weights The loss function is incorporated to further enhance the model's adaptation to seasonal precipitation.

7. A dual polarization radar city heavy rainfall short-term forecast deep learning system using information fusion implementing the dual polarization radar city heavy rainfall short-term forecast deep learning method using information fusion as claimed in any one of claims 1 to 6, characterized in that: The information fusion dual polarization radar urban heavy rainfall short-term forecast deep learning system includes: Construction module, used to construct radar heavy precipitation dataset; Fusion module, used to fuse terrain information with radar images; Design module for super-resolution CNN design; the super-resolution CNN model learns low-resolution images through convolutional neural networks and generates high-resolution images; in the heavy precipitation forecast task, radar images are used as input, combined with terrain information DEM maps, and feature extraction and boundary enhancement are performed through the super-resolution CNN network; the super-resolution CNN learns the mapping relationship between low-resolution radar images and high-resolution images; Add a module for adding LSTM model and seasonal weights; in heavy precipitation prediction, LSTM is used to process radar image sequences or other time series data; Training module, used for joint training.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the deep learning method for short-term forecast of heavy rainfall in cities using dual-polarization radar with information fusion as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes the steps of the deep learning method for short-term forecast of heavy rainfall in cities using dual-polarization radar with information fusion as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the information fusion dual-polarization radar urban heavy rainfall short-term forecast deep learning system as described in claim 7.

Images