Radar precipitation echo time interpolation spatial super-resolution method based on DConvLSTM

Abstract

The invention provides a radar precipitation echo time interpolation spatial super-resolution method based on DConvLSTM. The radar precipitation echo time interpolation spatial super-resolution method based on DConvLSTM comprises the following steps: constructing a radar precipitation echo time interpolation spatial super-resolution deep neural network model based on DConvLSTM; training, adjusting and optimizing the constructed deep neural network model by using the preprocessed X-band radar networking data; inputting the preprocessed S-band radar echo data into the trained and optimized deep neural network model; and S-band radar precipitation echo time interpolation spatial super-resolution data are output through the deep neural network model. According to the method, the characteristics of radar data of different wave bands can be effectively utilized, collaborative fusion of multi-source data is achieved, the finally generated radar echo image has higher time resolution and spatial resolution and can be used for refined disastrous weather forecast, the limitation of original radar observation data is remarkably improved, and the method is suitable for being used for forecasting disastrous weather. And the practicability and the prediction precision of the data are improved.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and specifically relates to a spatial super-resolution method for radar precipitation echo time interpolation based on DConvLSTM. Background Technology

[0002] With the continuous development of meteorological observation technology, high spatiotemporal resolution radar observation data plays a crucial role in the monitoring, forecasting, early warning, and mechanism research of severe weather. However, while current S-band radars have the advantages of wide coverage and reliable detection data, their spatiotemporal resolution is relatively low, making it difficult to meet the needs of refined meteorological analysis and severe weather forecasting. X-band radars, although possessing high spatiotemporal resolution, have a smaller observation range and are susceptible to rain attenuation, potentially leading to blindness in extreme cases. Therefore, they limit the ability to accurately monitor and forecast severe weather. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a spatial super-resolution method for radar precipitation echo temporal interpolation based on DConvLSTM. By constructing a deep neural network model for spatial super-resolution of radar precipitation echo temporal interpolation based on DConvLSTM, high-quality temporal and spatial super-resolution processing is performed on S-band dual-polarization radar echoes, providing more refined data support for meteorological monitoring and forecasting, and overcoming the deficiencies of existing radar systems in terms of resolution and coverage.

[0004] To achieve the above objectives, this invention discloses a spatial super-resolution method for radar precipitation echo temporal interpolation based on DConvLSTM, which includes:

[0005] Construct a spatial super-resolution deep neural network model for radar precipitation echo time interpolation based on DConvLSTM;

[0006] A deep neural network model was trained and optimized using preprocessed X-band radar network data.

[0007] Input the preprocessed S-band radar echo data into the trained and optimized deep neural network model;

[0008] The deep neural network model outputs spatial super-resolution data of S-band radar precipitation echo time interpolation.

[0009] As an optional implementation, the construction of a radar precipitation echo time interpolation spatial super-resolution deep neural network model based on DConvLSTM includes:

[0010] Set the input layer of the network model, where the input radar echo data sequence consists of two adjacent time frames;

[0011] Multi-scale data features are extracted from two adjacent time frames of the input to obtain feature map data viewL1 for parallel processing channel L1, feature map data viewL2 for parallel processing channel L2, and feature map data viewL3 for parallel processing channel L3.

[0012] For the feature map data viewL1, viewL2 and viewL3 of two adjacent time frames in parallel processing channels L1, L2 and L3, we perform spatial pyramiding and spatial alignment of time series features to enhance the ability to capture spatial deformation features.

[0013] Interpolating the time series of radar echo images using a bidirectional deformable convolutional long short-term memory network generates an interpolated frame for a new time between two time points.

[0014] High spatial resolution radar echo images are generated by improving image spatial resolution through feature integration and upsampling.

[0015] Set up the output layer of the network model to output radar echo images with doubled spatial resolution and doubled temporal resolution.

[0016] As an optional implementation, the step of extracting multi-scale data features from two adjacent input time frames to obtain the feature map data viewL1 for parallel processing channel L1 includes:

[0017] The radar echo data sequence features of two adjacent input time frames are stacked 10 times by a residual module without batch normalization to obtain the feature map data viewL1 of the parallel processing channel L1.

[0018] As an optional implementation, obtaining the feature map data viewL2 for the parallel processing channel L2 includes:

[0019] The obtained feature map data viewL1 is processed through a convolutional layer (conv) and a LeakyReLU activation layer to obtain the feature map data viewL2 for parallel processing channel L2.

[0020] As an optional implementation, obtaining the feature map data viewL3 for the parallel processing channel L3 includes:

[0021] The obtained feature map data viewL2 is processed again through a convolutional Conv and Leaky ReLU activation layer to obtain the feature map data viewL3 for the parallel processing channel L3.

[0022] As an optional implementation, the process of performing spatial pyramiding and spatial alignment of time-series features on the feature map data viewsL1, viewL2, and viewL3 of two adjacent time frames across the three scales of parallel processing channels L1, L2, and L3, thereby enhancing the ability to capture spatial deformation features, includes:

[0023] The feature map data of two adjacent time frames of the three scales L1, L2 and L3 in parallel processing channels, namely viewL1, viewL2 and viewL3, are fed into the pyramid, cascade and deformable convolution module PCD. The PCD module then fuses these features into the frame rate of the intermediate time, performs spatial alignment of time series features, and enhances the ability to capture spatial deformation features.

[0024] As an optional implementation, the step of improving image spatial resolution through feature integration and upsampling to generate a high spatial resolution radar echo image includes:

[0025] The feature data processed by the bidirectional deformable convolutional LSTM first undergoes data shape adjustment, and then the features are fused with the LeakyReLU activation function through two convolutional layer Conv operations to obtain a reconstructed feature map. Then, it enters the upsampling stage to generate a high-resolution radar echo image.

[0026] As an optional implementation, the deep neural network model trained and optimized using preprocessed X-band radar network data includes:

[0027] Acquire raw data for X-band radar network;

[0028] Preprocessing the raw data of the X-band radar network yields X-band radar network echo image data with different temporal and spatial resolutions.

[0029] Pairs of X-band radar network echo image data with different temporal and spatial resolutions were input to train and optimize a radar precipitation echo temporal interpolation spatial super-resolution deep neural network model based on DConvLSTM.

[0030] As an optional implementation, the step of inputting preprocessed S-band radar echo data into the trained and optimized deep neural network model includes:

[0031] The raw S-band radar data is acquired, preprocessed, and a radar combined reflectivity product CREF with the same latitude and longitude range as the X-band radar network data is generated. The S-band radar combined reflectivity product CREF is then input into the trained and optimized deep neural network model.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The model of this invention divides the radar echo into three different feature maps and inputs them into a pyramid, cascading and deformable (PCD) module. These features are fused into the frame rate of the intermediate time. Then, these three temporal features are fed into a bidirectional deformable convolutional long short-term memory network (BiDConvLSTM) to learn the features in the preceding and following time sequences. Finally, the output is processed through three convolutional layers to achieve temporal and spatial refinement.

[0034] 2. This invention performs multi-source data fusion, using X-band phased array radar echo data to train and optimize the model and S-band dual-polarization radar echo data usage model to carry out high-quality processing in a dual-band collaborative manner. It fully utilizes the advantages of the two types of radar in terms of spatiotemporal resolution, coverage, and data reliability, and overcomes the limitations of existing radar system echo data in terms of resolution and coverage. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the network architecture of a radar precipitation echo time interpolation spatial super-resolution deep neural network model based on DConvLSTM, according to an embodiment of the present invention.

[0036] Figure 2 This is a graph showing the proportion of the number of pixels of precipitation echo to the total number of pixels in a data sample of an embodiment of the present invention.

[0037] Figure 3 A distribution diagram of the proportion of precipitation echoes before they are clipped, based on a 30% threshold setting according to an embodiment of the present invention.

[0038] Figure 4 A distribution map of the proportion of precipitation echoes after the precipitation echoes are clipped, based on a 5% threshold setting of one embodiment of the present invention.

[0039] Figure 5 This is a schematic diagram of radar precipitation echo image cropping for an expanded sample according to an embodiment of the present invention.

[0040] Figure 6 This is a meteorological radar observation echo image at a corresponding moment according to an embodiment of the present invention;

[0041] Figure 7 This is a bicubic interpolation super-resolution image obtained by optical flow method according to one embodiment of the present invention;

[0042] Figure 8This is a spatial super-resolution image of radar precipitation echo time interpolation based on DConvLSTM, according to an embodiment of the present invention.

[0043] In the picture:

[0044] 1-Echo image cropped from the top left vertex of the radar precipitation echo image;

[0045] 2-Echo image cropped from the upper right vertex of the radar precipitation echo image;

[0046] 3-Echo image cropped from the lower left vertex of the radar precipitation echo image;

[0047] 4-Echo image cropped from the lower right vertex of the radar precipitation echo image. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-8 Further explanation is provided in the examples and embodiments.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] As an embodiment of the present invention, the present invention provides a spatial super-resolution method for radar precipitation echo time interpolation based on DConvLSTM, which includes the following steps:

[0051] S100. Construct a spatial super-resolution deep neural network model for radar precipitation echo time interpolation based on DConvLSTM.

[0052] S200, a deep neural network model was trained and optimized using preprocessed X-band radar network data;

[0053] S300: Input the preprocessed S-band radar echo data into the trained and optimized deep neural network model;

[0054] S400 outputs S-band radar precipitation echo time interpolation spatial super-resolution data through a deep neural network model.

[0055] As an optional implementation, the above-mentioned S100, which involves constructing a radar precipitation echo time interpolation spatial super-resolution deep neural network model based on DConvLSTM, is shown in the attached figure. Figure 1 As shown, it includes:

[0056] S101. Set the input layer of the network model. The input radar echo data sequence consists of two adjacent time frames.

[0057] S102. Extract multi-scale data features from two adjacent time frames of input to obtain feature map data viewL1 of parallel processing channel L1, feature map data viewL2 of parallel processing channel L2, and feature map data viewL3 of parallel processing channel L3.

[0058] S103. Perform spatial pyramiding and spatial alignment of time series features on two adjacent time frame feature map data viewL1, viewL2 and viewL3 of the three scales of parallel processing channels L1, L2 and L3, and enhance the ability to capture spatial deformation features.

[0059] S104. Interpolate the time series of radar echo images using a bidirectional deformable convolutional long short-term memory network to generate an interpolated frame for a new time between two time points.

[0060] S105. Improve image spatial resolution through feature integration and upsampling to generate high spatial resolution radar echo images;

[0061] S106. Set the output layer of the network model to output radar echo images with doubled spatial resolution and doubled temporal resolution.

[0062] As an optional implementation, in step S101 above, the input layer of the network model is set, and the input radar echo data sequence consists of two adjacent time frames, including:

[0063] As attached Figure 1 As shown, the input radar echo data sequence consists of two adjacent time frames (Fi and Fi+1). The data dimensions are: Number of frames (nframe): refers to the number of frames that may be contained in each time step. Batch size (batch): refers to the number of samples input into the network at one time. Image height (height) and width (width): refer to the two-dimensional spatial dimensions of each radar echo frame.

[0064] Suppose the shape of the input data is (batch, nframe, height, width), where nframe = 2, indicating that there are two adjacent time frames.

[0065] As an optional implementation, S102 above involves extracting multi-scale data features from two adjacent input time frames to obtain feature map data viewL1 for parallel processing channel L1, feature map data viewL2 for parallel processing channel L2, and feature map data viewL3 for parallel processing channel L3, including:

[0066] As an optional implementation, the above-described method for obtaining the feature map data viewL1 of the parallel processing channel L1 includes:

[0067] The radar echo data sequence features from two adjacent time frames (Fi and Fi+1) are stacked 10 times using a residual module (ResidualBlock_noBN) without batch normalization to generate the feature map data viewL1 for parallel processing channel L1, from which high-order spatial features are extracted. Feature extraction is then performed using the residual network structure module ResidualBlock_noBN, with a total of 10 stacking operations. It is important to note that this module does not include batch normalization; this design helps preserve more image detail information. The output is the feature map data viewL1 for parallel processing channel L1.

[0068] As an optional implementation, the feature map data viewL2 obtained by parallel processing channel L2 includes:

[0069] The obtained feature map data viewL1 is processed through a convolutional layer conv and a LeakyReLU activation layer to obtain the feature map data viewL2 of the parallel processing channel L2, which further extracts high-order spatial features and refines the expression of spatial scale features.

[0070] Convolutional layers (conv) extract spatial features from an input radar feature map by applying a series of learnable convolutional kernels (or filters). Each kernel focuses on detecting specific features in the image, such as edges, corners, or more complex shapes. By stacking multiple convolutional layers, the network can learn more complex feature representations. The result of the convolution operation is a feature map, which represents the spatial distribution of detected features in the input image. As the network depth increases, the feature map becomes increasingly abstract, capturing higher-level image content.

[0071] The LeakyReLU activation function introduces non-linearity, enabling the network to learn complex mappings. Especially after convolutional operations, it helps the model extract more useful spatial features from radar echo data.

[0072] LeakyReLU (Leaky Rectified Linear Unit) is a variant of ReLU (Rectified Linear Unit) that produces a small gradient (typically 0.01 or other small values) when the input is less than zero, instead of outputting zero like ReLU. The advantage of LeakyReLU is that, compared to ReLU, it helps the network avoid the "dead ReLU" problem during learning, where some neurons may never be activated. It allows the propagation of negative activation values, which helps the network better capture and propagate gradients during training, thus learning richer features.

[0073] Spatial features are extracted by combining Conv and LeakyReLU, including:

[0074] Feature extraction: Spatial features in the radar feature map are gradually extracted and abstracted through the Conv convolutional layer of the CNN.

[0075] Nonlinear activation: Applying the LeakyReLU activation function after each convolutional layer increases the nonlinearity of the network, enabling it to learn more complex features and maintain the flow of gradients during backpropagation.

[0076] Feature enhancement: LeakyReLU helps enhance the network's sensitivity to weak or negative features in radar images, which is especially important for identifying subtle features such as precipitation and storm structure.

[0077] In this way, combining CNN and LeakyReLU can effectively extract more spatial features from radar feature maps, which is very beneficial for improving the accuracy and reliability of radar image analysis.

[0078] The generated feature map data viewL1 is processed through a convolutional layer conv and a LeakyReLU activation layer to obtain the feature map data viewL2 of the parallel processing channel L2, which is then used to further extract higher-order spatial features and refine the expression of spatial scale features.

[0079] As an optional implementation, the above-described method for obtaining the feature map data viewL3 for parallel processing channel L3 includes:

[0080] The obtained feature map data viewL2 is passed through a convolutional Conv and LeakyReLU activation layer again to obtain the feature map data viewL3 of the parallel processing channel L3, and then the expression of spatial scale features is further refined.

[0081] Similar to the above, convolutional Conv and LeakyReLU activation functions are used to further extract higher-order spatial features, generating the feature map data viewL3 for parallel processing channel L3.

[0082] This invention divides radar echoes into three feature maps as subsequent inputs, performs multi-scale feature extraction and fusion, which can effectively enhance the model's ability to perceive spatial details.

[0083] As an optional implementation, S103 above, which performs spatial pyramiding and spatial alignment of time series features on the feature map data viewL1, viewL2, and viewL3 of two adjacent time frames at the three scales of parallel processing channels L1, L2, and L3, and enhances the ability to capture spatial deformation features, includes:

[0084] The feature map data from three scales (L1, L2, and L3) of the parallel processing channels (two adjacent time frames (Fi and Fi+1)) are jointly fed into the Pyramid, Cascading, and Deformable Module (PCD). The PCD module then fuses these features into an intermediate time frame rate for spatial alignment of the time-series features (PCD Align) to ensure accurate spatial matching between consecutive frames. The spatially aligned data is then input into a cascaded structure and a deformable convolution to enhance the ability to capture spatially deformable features.

[0085] This module performs spatial pyramiding on the aforementioned three feature maps (viewL1, viewL2, and viewL3) (multi-scale feature extraction, where viewL1 is the 1x size, viewL2 is the 1 / 2x size, and viewL3 is the 1 / 4x size). It also combines cascaded structures and deformable convolutions to enhance the ability to capture spatial deformation features. This structure effectively captures feature information at different scales, helping the network better adapt to changes in the spatial details of radar echoes and improving spatial detail and perception capabilities. The PCD module combines pyramids, cascaded structures, and deformable convolutions to fuse feature map data from different scales. This fusion helps the network better understand and predict the motion and changes of objects when processing time-series data.

[0086] As an optional implementation, S104 above, which involves interpolating the time series of the radar echo image using a bidirectional deformable convolutional long short-term memory network to generate an interpolated frame for a new time between two time points, includes:

[0087] The two-frame sequence of feature maps formed after processing by the PCD module is fed into the bidirectional deformable convolutional long short-term memory network BiDConvLSTM to interpolate the time series of the radar echo image. The function of the bidirectional deformable convolutional long short-term memory network BiDConvLSTM is to interpolate the time series of the radar echo image, that is, to use the feature information between two adjacent time points (Fi and Fi+1) to generate an interpolated frame of a new time point between the two time points.

[0088] BiDConvLSTM combines the long short-term memory of traditional LSTM with the spatial adaptability of deformable convolution, enabling it to efficiently capture temporal features while better perceiving local changes and irregular deformations in spatial information. The bidirectional structure means the model can simultaneously reference image features from two time points, thus more accurately deriving interpolation results. By simultaneously referencing feature frames before and after the current time (T) (T-1 and T+1), it more precisely captures the dynamic information of radar echo evolution over time, achieving efficient interpolation and prediction in the temporal dimension and obtaining deep feature representations containing temporal information.

[0089] A bidirectional deformable convolutional long short-term memory network (BiDConvLSTM) is used to perform temporal modeling of the fused features, capturing the dynamic evolution characteristics of radar echoes over time. The specific process includes:

[0090] I. Forward Process:

[0091] Forward input feature extraction: First, perform a two-dimensional convolution operation on each frame of the radar echo image sequence to extract low-level spatial feature information and form an initial feature map, which lays the foundation for subsequent processing.

[0092] Forward deformable convolution processing: For the initial feature map of each frame, deformable convolution operation is further introduced to achieve adaptive adjustment of the receptive field of traditional regular convolution, thereby capturing non-rigid motion features and local structural changes in radar images more precisely.

[0093] Forward LSTM module processing: Temporal features enhanced by deformable convolutions are input into the forward LSTM unit. The LSTM unit sequentially calculates the input gate, forget gate, output gate, and cell state of the current frame, and updates the hidden state sequence by combining the current input features with historical memory information. Through the forward time propagation mechanism, long-term temporal dependency information is extracted step by step from the first frame to the last frame to obtain the forward hidden state sequence.

[0094] II. Reverse process:

[0095] Reverse input feature extraction and deformable convolution processing: Starting from the last frame of the input sequence, feature extraction is performed on the image frames in reverse chronological order. The process is the same as the forward process, applying two-dimensional convolution and deformable convolution in sequence to obtain the reversed feature map.

[0096] The inverse LSTM module processes the data by inputting the inverse feature map into the inverse LSTM unit. The LSTM structure utilizes the current input and the hidden and cell states from the previous time step to compute the input gate, forget gate, output gate, and updated cell state, thereby achieving feature modeling in the inverse temporal dimension. This mechanism effectively captures inverse temporal dependencies in image sequences.

[0097] Backpropagation computation: Starting from the last frame, recursively advance frame by frame to obtain the sequence of backward hidden states, thus achieving complete modeling in the time dimension.

[0098] III. Two-way fusion and interpolation generation:

[0099] Feature concatenation and fusion: The forward and backward hidden state sequences are concatenated along the feature channel dimension to form a bidirectional temporal feature set. This concatenation operation combines global spatiotemporal context information from both the forward and backward directions, which helps improve the prediction capability of intermediate time frames.

[0100] Fusion Layer Output and Interpolation Reconstruction: The stitched feature sequence is input into the fusion module, which includes structures such as attention mechanisms, channel weighting, and residual connections. This module further extracts key fusion information and reconstructs the fused features using deformable convolution, outputting an interpolated image frame with high spatiotemporal resolution. This module can interpolate between any two adjacent time frames (e.g., T-1 and T) in a radar echo image sequence to generate a predicted image at a new time (e.g., T-0.5), achieving temporal completion and enhancement.

[0101] IV. Overall Advantages Description:

[0102] The bidirectional deformable convolutional LSTM network structure proposed in this invention can extract global spatiotemporal dependency features from radar echo image sequences, overcoming the shortcomings of traditional models in nonlinear motion modeling and long-term dependency capture. Through bidirectional feature learning, dynamic convolutional receptive field, attention-guided fusion, and high-quality interpolation reconstruction mechanisms, this method can effectively improve the resolution of radar images in both temporal and spatial dimensions, providing a more accurate image foundation for weather forecasting and severe convection analysis.

[0103] As an optional implementation, the above-described S105, which improves the spatial resolution of the image through feature integration and upsampling to generate a high spatial resolution radar echo image, includes:

[0104] Further feature integration and upsampling stage: After processing by bidirectional deformable convolutional LSTM, the feature data enters the upsampling stage. First, the data shape is adjusted (view), and then the features are fused with the LeakyReLU activation function through two upsampling Conv operations to obtain a reconstructed feature map of size (B,T,64,2H,2W). Then, the upsampling (Up) stage is entered, where upsampling (such as PixelShuffle or deconvolution) enhances the image spatial resolution, thereby generating a higher resolution (e.g., twice the original resolution) radar echo image.

[0105] In a deep learning model, a PixelShuffle layer is inserted. The stride parameter controls the upsampling ratio. To double the spatial resolution (from 1000 meters to 500 meters), the stride parameter needs to be set to 2. For deconvolution operations, if the kernel size is k×k and the goal is to double the resolution (i.e., double the height and width), the stride is typically set to 2. This way, each convolution operation doubles the image's spatial resolution, thus achieving a spatial resolution improvement.

[0106] As an optional implementation, S106 above, setting the output layer of the network model to output radar echo images with doubled spatial resolution and doubled temporal resolution, includes:

[0107] Configure the output layer of the network model: output high spatial resolution (resolution doubled, i.e., 1000 meters becomes 500 meters) interpolated radar echo images. The temporal interpolation function improves the original low temporal resolution radar observation data (e.g., the original time interval of S-band radar is 6 minutes) to higher temporal resolution data (e.g., 3 minutes) to meet the needs of refined analysis and forecasting.

[0108] The model of this invention mainly divides the radar echo into three feature maps and inputs them into a pyramid, cascading and deformable (PCD) module. These features are fused into the frame rate of the intermediate time. Then, these three temporal features are fed into a bidirectional deformable convolutional long short-term memory network (BiDConvLSTM) to learn the features in the preceding and following time sequences. Finally, the output is processed through three convolutional layers to achieve temporal and spatial refinement.

[0109] As an optional implementation, the aforementioned S200, which uses preprocessed X-band radar network data to train and optimize a deep neural network model, includes:

[0110] S201. Obtain raw data for X-band radar networking;

[0111] Guangdong Province has now built an X-band phased array radar network (a total of 47 X-band radars). X-band radars have excellent spatiotemporal resolution, with a spatial resolution of 30 meters and a temporal resolution of 1 minute or 1.5 minutes. However, they have disadvantages such as a small observation range and susceptibility to rain attenuation. Most importantly, the observation attenuation is severe for heavy rain and above, and in extreme cases, it may lead to blindness.

[0112] The original radar data of the X-band phased array radar network, which overlaps with the observation coverage area of ​​the S-band radar, is acquired. In this embodiment, historical data of the X-band phased array radar network from November 17, 2020 to November 28, 2024 is selected. Based on the observation time (every 1.5 minutes or 1 minute) and the X-band radar network data path, data for corresponding time intervals is retrieved according to the principle of time proximity, and time normalization is performed, such as normalizing 00:05 seconds to 00 minutes and 00:55 seconds to 01 minutes. This ensures that both 1.5-minute and 1-minute data can provide data characteristics and are not wasted.

[0113] S202. Preprocess the raw data of the X-band radar network to obtain X-band radar network echo image data with different temporal and spatial resolutions, including:

[0114] S2021. Determine whether the acquired X-band radar network data contains precipitation.

[0115] Determining whether precipitation has occurred based on the acquired raw radar data from the X-band phased array radar network, including:

[0116] If the maximum basic reflectivity of the radar raw data is less than 20 dBZ, then there will be no precipitation at that time.

[0117] If the maximum basic reflectivity of the original radar data is ≥20dBZ, select 17*17 grids (1KM / 60m≈17) centered on the grid points with a basic reflectivity ≥20dBZ. The number of valid values ​​in this area should be >50%, where a basic reflectivity >0dBZ is considered a valid value. Calculate the average basic reflectivity (avg) of this small area (17*17 grid area). If at least one small area has an avg ≥10dBZ, then precipitation is considered to have occurred at that time; otherwise, there is no precipitation at that time.

[0118] S2022. Downsampling of X-band radar network data with precipitation;

[0119] Downsampling of X-band radar network data with precipitation: For X-band radar network data with precipitation, downsampling is performed based on the current data spatial resolution of 30 meters and the target spatial resolution (here, the target spatial resolutions are 960 meters and 480 meters respectively), representing both high and low spatiotemporal target resolutions. Downsampling includes the following methods:

[0120] 1. Read X-band radar network data;

[0121] 2. Based on the resolution multiplier, select the grid point correspondence, calculate the correspondence between n current grid points and 1 target grid point, and take the maximum value of the n current grid points as the target grid point value;

[0122] For example:

[0123] 1) 16 grids with a resolution of 30 meters correspond to 1 grid with a resolution of 480 meters;

[0124] 2) Select the maximum value of the 16 grids as the value of 1 grid corresponding to 480 meters, and the same applies to 960 meters.

[0125] 3) Downsampling in time involves rounding the original data from 1 to 1.5 minutes into 3-minute intervals based on the principle of time proximity. Obtaining 6-minute interval data involves sampling data at 3-minute intervals.

[0126] S2023. Construct initial matrices of nan based on the specified latitude and longitude range (in this embodiment, the area covered by the overlapping observations of the S-band radar and X-band radar mosaic (a total of 47 units) in the Pearl River Delta region centered on Guangzhou) and the target resolutions of 960 meters for 6 minutes and 480 meters for 3 minutes.

[0127] S2024. Fill the corresponding positions of the initial matrix with the downsampled X-band networking data;

[0128] S2025 outputs X-band radar network data with different rainfall coverage rates, totaling 9600+ rainfall time periods with different rainfall coverage rates. There are two sets of data with target resolutions of 960 meters and 480 meters, and each set of data contains 500,000+ rainfall radar images.

[0129] S2026. Screen and process the rainfall radar image data to ensure the validity and consistency of the model input data;

[0130] Filtering and processing rainfall radar image data involves deleting data with weak echoes and uniformly cropping data with strong echoes, so that the radar echo (rainfall) data input into the network is more evenly distributed, forming the dataset for model training.

[0131] The specific processing steps include:

[0132] This radar precipitation echo dataset includes over 9000 rainfall time periods, with over 500,000 radar precipitation echo images per group. To better explore the data distribution of radar precipitation echoes, a graph showing the proportion of precipitation echo pixels to the total number of pixels in the data sample, divided into 5% intervals, was plotted (see attached). Figure 2 As shown in the figure, it can be seen that the proportion of radar precipitation echoes in most samples is less than 5%.

[0133] During training, if the traditional method of randomly cropping radar precipitation echo images into 64×64 pixel images is used as model input data, there is a high probability that the data samples with a small number of precipitation echo pixels will not contain any precipitation echoes. Such image inputs lacking effective precipitation echoes will make it difficult to capture the key features of precipitation, thus affecting the training effect and model performance.

[0134] To address this issue, this invention controls the proportion of precipitation echo pixels in the total number of pixels entering the network sample by setting two thresholds. The first threshold is set before precipitation echoes are cropped (the distribution of precipitation echo proportions when the threshold is 30% is shown in the appendix). Figure 3 The second threshold is set after the precipitation echo is clipped (when the threshold is 5%, the distribution of precipitation echo proportions is shown in the appendix). Figure 4 However, after this screening, the number of data samples was significantly reduced, leaving only 2000+ samples. To expand the sample size, this invention uses the four vertices of the radar precipitation echo image as starting points. For precipitation echoes with spatial resolutions of 960 meters and 480 meters, each echo image is cropped into four identical echo images with a small number of overlapping areas of the same size, respectively, at dimensions of 256×256 pixels and 512×512 pixels. This ensures that the resulting images cover as much precipitation feature information as possible. Figure 5 As shown in the figure, the blue-bordered echo image 1 is cropped from the top left vertex of the radar precipitation echo image; the purple-bordered echo image 2 is cropped from the top right vertex of the radar precipitation echo image; the purple-bordered echo image 3 is cropped from the bottom left vertex of the radar precipitation echo image; and the red-bordered echo image 4 is cropped from the bottom right vertex of the radar precipitation echo image. After multiple experiments, this invention selected echo images representing 30% of the total number of pixels (precipitation echo pixels account for 30% of the total number of pixels) to form the training dataset, with each group containing 960 meters and 480 meters containing more than 14,000 data samples.

[0135] For a specific example, a 480×480 precipitation echo image, divided into 256×256 sub-images, yields the following four images, along with the coordinates of their four vertices:

[0136] First image (top left): Top left vertex coordinates: (0,0); Bottom right vertex coordinates: (255,255);

[0137] The second image (top right corner): Top left vertex coordinates: (224, 0); Bottom right vertex coordinates: (479, 255);

[0138] The third image (bottom left): Top left vertex coordinates: (0, 224); Bottom right vertex coordinates: (255, 479);

[0139] The fourth image (bottom right corner): top left corner vertex coordinates: (224, 224); bottom right corner vertex coordinates: (479, 479).

[0140] Each 256×256 image is cropped from one corner of the original 480×480 image, with a cropped size of 256 pixels × 256 pixels. Since 480 - 256 = 224, the top-left corner coordinates of the second and fourth images are (224, 0) and (224, 224), respectively. Similarly, the top-left corner coordinates of the third and fourth images are (0, 224) and (224, 224), respectively. The bottom-right corner coordinates of each image are obtained by adding 255 to the top-left corner coordinates (because pixel indexing starts from 0).

[0141] For a 960×960 precipitation echo image, dividing it into 512×512 pixels yields the following four images, along with the coordinates of their four vertices:

[0142] First image (top left): Top left vertex coordinates: (0,0); Bottom right vertex coordinates: (511,511);

[0143] The second image (top right corner): Top left vertex coordinates: (448, 0); Bottom right vertex coordinates: (959, 511);

[0144] The third image (bottom left): Top left vertex coordinates: (0, 448); Bottom right vertex coordinates: (511, 959);

[0145] The fourth image (bottom right corner): top left corner vertex coordinates: (448, 448); bottom right corner vertex coordinates: (959, 959).

[0146] Each 512×512 image is cropped from a corner of the original 960×960 image. Since 960-512=448, the top-left corner coordinates of the second and fourth images are (448,0) and (448,448), respectively. Similarly, the top-left corner coordinates of the third and fourth images are (0,448) and (448,448), respectively. The bottom-right corner coordinates of each image are obtained by adding 511 to the top-left corner coordinates (because pixel indices start from 0).

[0147] By segmenting a 480×480 precipitation echo image into four 256×256 images starting from each of the four vertices, and by segmenting a 960×960 precipitation echo image into four 512×512 images starting from each of the four vertices, a sufficient number of data samples are ensured, while slight overlap is considered to ensure more complete precipitation information is entered into the network.

[0148] S203. Input X-band radar network echo image data with different time and spatial resolutions in pairs to train and optimize the constructed radar precipitation echo time interpolation spatial super-resolution deep neural network model based on DConvLSTM.

[0149] As an example, during model training, the input consists of two adjacent time frames (Fi and Fi+1) at a 6-minute interval between 960 meters, i.e., data values ​​taken at a 6-minute interval after downsampling (1 minute or 1.5 minutes). Simultaneously, the input is paired with three adjacent time frames (Fi, F(i+1) / 2, and Fi+1) at a 3-minute interval between 480 meters, used for model data verification and parameter tuning. X-band radar network echo image data with over 14,000 data samples each at 960 meters 6-minute and 480 meters 3-minute resolutions are input into the model in pairs, forming one sample per pair, to train and optimize the aforementioned DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution deep neural network model.

[0150] As an optional implementation, the above-mentioned S300, inputting preprocessed S-band radar echo data into the trained and optimized deep neural network model, includes:

[0151] S301. Acquire raw data from the S-band radar;

[0152] Guangdong Province has established an S-band dual-polarization radar network. S-band radar has a wide coverage area and high reliability of detection data, but its spatiotemporal resolution is poor, with a spatial resolution of 1,000 meters and a temporal resolution of 6 minutes, which cannot meet the requirements of refined meteorological analysis and severe weather forecasting.

[0153] The raw S-band radar data in this embodiment of the invention is from April 17, 2019 to May 19, 2024. Based on the observation time and S-band radar data path, the raw S-band radar data for the corresponding time period (once every 6 minutes) is retrieved according to the principle of time proximity.

[0154] S302. Preprocess the raw S-band radar data, including:

[0155] To determine whether there is precipitation in the raw S-band radar data, the method for determining precipitation is the same as the X-band method described above.

[0156] After quality control of the raw S-band radar data with precipitation, the quality control includes removing isolated precipitation echoes and removing non-meteorological echoes (such as ocean waves, birds, and ground objects) from the S-band radar data.

[0157] S303. Generate radar combined reflectivity product CREF with the same latitude and longitude range as the X-band radar network data.

[0158] Based on the preprocessed S-band radar data described above, a 1000-meter, 6-minute resolution S-band radar composite reflectivity product (CREF) is generated. CREF is a radar meteorological product that provides a two-dimensional image of the ability of precipitation particles or other scattering bodies in the atmosphere within the radar coverage area to reflect radar waves. CREF is obtained by combining and selecting the maximum value from data scanned by the radar at different elevation angles; therefore, it can display the strongest reflectivity perpendicular to the radar beam direction within the radar's detection range.

[0159] S304. Using the high spatiotemporal resolution echo data of the X-band phased array radar as an aid, the echo data of the S-band dual polarization radar is registered.

[0160] Using the high spatiotemporal resolution echo data of the X-band phased array radar as an aid, the echo data of the S-band dual-polarization radar is registered. Registration refers to cropping the observation coverage area of ​​the S-band dual-polarization radar according to the mosaic area of ​​the X-band phased array radar.

[0161] S304. Input the S-band radar combined reflectivity product CREF into the trained and optimized deep neural network model.

[0162] Input the S - band radar composite reflectivity product CREF with a resolution of 6 minutes for every 1000 meters into the optimized deep neural network model through training.

[0163] As an optional implementation manner, the above S400 outputs the S - band radar precipitation echo time - interpolated spatial super - resolution data through the deep neural network model, including:

[0164] Generate radar echo data after time interpolation and spatial super - resolution processing through the deep neural network model, and output data with a resolution of 500m and a time interval of 3 minutes.

[0165] As an optional implementation manner, the present invention also includes evaluating the results to ensure the accuracy and reliability of the data.

[0166] Through the observation data of 47 X - band phased - array radars networked in Guangdong Province, perform spatio - temporal super - resolution processing on the echoes of the S - band radar. After processing, the time resolution is improved from 6 minutes to 3 minutes, and the spatial resolution is improved from 1000 meters to 500 meters. Evaluate and compare the method of the present invention and the traditional bicubic interpolation super - resolution method. The results of the evaluation indexes are shown in Table 1 below:

[0167] Table 1: Comparison results of index evaluation between the present invention and the optical flow method

[0168]

[0169] The statistical values of PSNR and SSIM in the above table show that the image data processed by the time - interpolated spatial super - resolution method of the present invention are significantly better than the traditional method of performing bicubic interpolation after extrapolation by the optical flow method in terms of peak signal - to - noise ratio (PSNR) and structural similarity (SSIM).

[0170] At the same time, as shown in the appendix Figure 6-8 shown, the appendix Figure 6 is the meteorological radar observation echo image at the corresponding moment, the appendix Figure 7 is the super - resolution image of bicubic interpolation after the traditional optical flow method, and the appendix Figure 8 is the image after time - interpolated spatial super - resolution based on DConvLSTM of the present invention; among them, blue represents weak echoes (0dBZ < ZH < 20dBZ), yellow represents medium echoes (20dBZ ≤ ZH < 35dBZ), and orange represents strong echoes (ZH ≥ 35dBZ). Through the observation results of the three figures, it can be seen that the image after time - interpolated spatial super - resolution based on DConvLSTM of the present invention is better than the bicubic interpolation method under weak, medium, and strong precipitation echoes, showing excellent performance, as Figure 7The results show that during periods of rapid change in precipitation echoes, traditional optical flow methods followed by bicubic interpolation super-resolution images exhibit abnormal deformations in certain regions, while the images obtained by temporal interpolation spatial super-resolution based on DConvLSTM in this invention are well aligned with the precipitation echoes observed by radar in the corresponding regions.

[0171] Based on the evaluation results from both aspects, this invention demonstrates that the method can significantly improve the spatiotemporal resolution of radar echoes, highlighting the potential of the DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution deep neural network model in the application of spatiotemporal super-resolution of meteorological radar data. This invention, through its neural network architecture, can perform high-quality interpolation and resolution enhancement of radar echo data in both temporal and spatial dimensions. The model effectively utilizes the characteristics of radar data from different bands (such as the high spatiotemporal resolution of the X-band and the stability of S-band radar data), achieving collaborative fusion of multi-source data. The resulting radar echo images have higher temporal and spatial resolution, enabling refined forecasting of severe weather, significantly improving the limitations of raw radar observation data, and enhancing the practicality and prediction accuracy of the data. It provides more refined and high-quality data support for meteorological monitoring and severe weather forecasting. This method features multi-source data fusion and bidirectional temporal modeling, demonstrating significant application prospects and promotional value.

[0172] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0173] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A radar precipitation echo time interpolation spatial super-resolution method based on DConvLSTM, characterized in that, It comprises: The radar precipitation echo time interpolation spatial super-resolution deep neural network model based on DConvLSTM is constructed, specifically comprising: The input layer of the network model is set, and the input radar echo data sequence is composed of two adjacent time frames; The multi-scale data features of the input two adjacent time frames are extracted to obtain the feature map data viewL1 of the parallel processing channel L1, the feature map data viewL2 of the parallel processing channel L2, and the feature map data viewL3 of the parallel processing channel L3; The two adjacent time frame feature map data viewL1, viewL2 and viewL3 of the three scales of parallel processing channels L1, L2 and L3 are pyramidized in the spatial scale and spatially aligned and enhanced in the time sequence features to improve the ability to capture spatial deformation features; The time sequence of the radar echo image is interpolated by the bidirectional deformable convolution long short-term memory network to generate an interpolated frame of a new time between two times; The image spatial resolution is improved by feature integration and up-sampling to generate a high spatial resolution radar echo image; The output layer of the network model is set, and the radar echo image with doubled spatial resolution and doubled time resolution is output.

2. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 1, wherein, The multi-scale data features of the input two adjacent time frames are extracted to obtain the feature map data viewL1 of the parallel processing channel L1, including: The radar echo data sequence features of the input two adjacent time frames are stacked by the batch normalization-free residual module 10 times to obtain the feature map data viewL1 of the parallel processing channel L1.

3. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 2, wherein, The feature map data viewL2 of the parallel processing channel L2 is obtained, including: The obtained feature map data viewL1 is processed by the convolution layer conv and the LeakyReLU activation layer to obtain the feature map data viewL2 of the parallel processing channel L2.

4. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 3, characterized in that, The feature map data viewL3 of the parallel processing channel L3 is obtained, including: The obtained feature map data viewL2 is processed again by the convolution Conv and the LeakyReLU activation layer to obtain the feature map data viewL3 of the parallel processing channel L3.

5. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 4, wherein, The two adjacent time frame feature map data viewL1, viewL2 and viewL3 of the three scales of parallel processing channels L1, L2 and L3 are pyramidized in the spatial scale and spatially aligned and enhanced in the time sequence features to improve the ability to capture spatial deformation features, including: The two adjacent time frame feature map data viewL1, viewL2 and viewL3 of the three scales of parallel processing channels L1, L2 and L3 are jointly input into the pyramid, cascaded and deformable convolution module PCD, and these features are fused into the intermediate time frame rate by the PCD module to perform spatial alignment and enhancement of the time sequence features and improve the ability to capture spatial deformation features.

6. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 5, wherein, The image spatial resolution is improved by feature integration and up-sampling to generate a high spatial resolution radar echo image, including: The feature data processed by the bidirectional deformable convolution LSTM is first subjected to data shape adjustment, and then fused by two convolution layers Conv operation and LeakyReLU activation function to obtain a reconstructed feature map, and then enters the up-sampling Up stage to generate a high-resolution radar echo image.

7. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 6, wherein, The deep neural network model constructed by training and tuning using the preprocessed X-band radar networking data includes: Obtaining X-band radar networking original data; Preprocessing the X-band radar networking original data to obtain X-band radar networking echo image data with different time and spatial resolutions; Inputting the X-band radar networking echo image data with different time and spatial resolutions in pairs to train and tune the deep neural network model of radar precipitation echo time interpolation spatial super-resolution based on DConvLSTM.

8. The DConvLSTM-based radar precipitation echo temporal interpolation spatial super-resolution method according to claim 7, characterized in that, Inputting the preprocessed S-band radar echo data into the trained and tuned deep neural network model, including: Obtaining S-band radar original data, preprocessing the S-band radar original data to generate radar combined reflectivity product CREF with the same longitude and latitude range as the X-band radar networking data, and inputting the S-band radar combined reflectivity product CREF into the trained and tuned deep neural network model.

Images