Local area rainfall intensity prediction method, device, system and storage medium

Abstract

The application relates to a local area rainfall intensity prediction method, device, system and computer readable storage medium. The method comprises the following steps: S1, shallow layer features are extracted from a continuous m frame rainfall intensity image sequence after preprocessing; S2, the shallow layer features are dimensionally reduced and transformed based on a Metaformer to extract deep layer features; S3, the deep layer features are dimensionally increased and transformed based on the Metaformer to output high-dimensional features; and S4, the high-dimensional features are mapped to target output, that is, a subsequent n frame rainfall intensity image sequence, wherein 0 < m <= n <= 20. The application can better fit the rainfall evolution process in a larger space-time range of a radar echo image sequence, so that the local area rainfall intensity prediction precision is improved.

Description

Technical Field

[0001] This application relates to the field of meteorological forecasting technology, and more specifically, to a method, apparatus, system, and computer-readable storage medium for predicting local rainfall intensity. Background Technology

[0002] Short-term (0-6 hours) localized rainfall intensity forecasting is an important issue in the field of weather forecasting. It provides weather guidance for people's daily life and economic activities, and plays an important role in promoting economic development and protecting people's lives and property.

[0003] Traditional localized rainfall intensity prediction primarily relies on radar echo extrapolation based on optical flow. This method assumes smooth motion and constant total rainfall intensity, estimating convective cloud movement patterns from continuous radar echo images and then using a semi-Lagrange advection model to predict future radar echoes. However, this method struggles to achieve high-precision rainfall intensity predictions when faced with dynamic, nonlinear motion patterns and rapidly changing rainfall intensities.

[0004] Machine learning methods that leverage deep neural network models, which possess excellent fitting capabilities for nonlinear processes, to model the spatiotemporal processes of radar echo sequence using a data-driven approach are gaining increasing attention. For example, ConvLSTM-based deep neural network models capture the spatiotemporal patterns of past radar echo sequence sequences and predict future radar echoes. This model, based on the classic LSTM (Long Short-Term Memory) network, replaces the fully connected neural network with a convolutional neural network to infer the hidden states corresponding to the image sequence, thus enabling it to simultaneously learn spatial correlations within temporal correlation learning. Consequently, it achieves higher prediction accuracy in short-term, localized radar echo sequence prediction tasks. To address the challenge of training ConvLSTMs due to their relatively large number of parameters, existing technologies have proposed Conv GRU (Gate Recurrent Unit) models to capture the spatiotemporal patterns of radar echo sequence sequences and predict future radar echoes. Since Conv GRU's convolution is based on an invariant local connectivity template, it lacks effective learning capabilities for the variable local spatial structure features of continuous radar echo sequence. The Trajectory GRU model was proposed to address this by first learning the local connectivity structure of the convolution operation before feature extraction, thereby improving the prediction accuracy of radar echo sequences. To learn multimodal rainfall patterns, the GA-GRU model has also been proposed for fitting rainfall processes. Compared to the Conv GRU model, its predicted radar echo sequences are sharper, and the accuracy of rainfall intensity prediction is also improved.

[0005] However, existing deep learning-based rainfall intensity prediction methods still have shortcomings in capturing spatiotemporal information over a large range, and their rainfall intensity prediction performance is greatly limited when dealing with rainfall events with a large spatiotemporal range. Summary of the Invention

[0006] The technical problem to be solved by this application is to provide a method, apparatus, system and computer-readable storage medium for predicting local rainfall intensity that can capture spatiotemporal features over a large range to improve the accuracy of rainfall intensity prediction, in view of the above-mentioned deficiencies of the prior art.

[0007] To address its technical problem, this application proposes a method for predicting rainfall intensity in a local area in its first aspect, comprising the following steps:

[0008] S1. Extract shallow features from the preprocessed sequence of m consecutive rainfall intensity maps;

[0009] S2. Based on Metaformer, the shallow features are subjected to dimensionality reduction transformation to extract deep features;

[0010] S3. Based on Metaformer, perform dimensionality up-transformation on the deep features to output high-dimensional features;

[0011] S4. Map the high-dimensional features to the target output, i.e., the subsequent n-frame rainfall intensity map sequence, where 0 <m≤n≤20。

[0012] In one embodiment of the local rainfall intensity prediction method according to this application, step S1 of extracting shallow features further includes:

[0013] S11. Perform the first pointwise convolution on the continuous m-frame rainfall intensity map, and perform feature transformation in the three dimensions of channel, height and width.

[0014] S12. Perform the first feature normalization on the feature map after the first pointwise convolution;

[0015] S13. Perform the first nonlinear mapping on the feature map after the first normalization process;

[0016] S14. Perform a second pointwise convolution on the feature map obtained in step S13;

[0017] S15. Perform a second feature normalization on the feature map that has undergone the second pointwise convolution.

[0018] S16. Perform a second nonlinear mapping on the feature map after the second normalization process to obtain the shallow features of the input rainfall intensity map.

[0019] According to one embodiment of the local rainfall intensity prediction method described in this application, step S2 further includes performing the following steps S21-S25 sequentially at least once:

[0020] S21. Normalize the input feature map and then perform separable convolution calculation;

[0021] S22. Add the feature map calculated in step S21 to the input feature map to obtain the first new feature map;

[0022] S23. Normalize the first new feature map, and then perform dimensionality increase and then dimensionality decrease transformation in the channel dimension;

[0023] S24. Add the transformed feature map from step S23 to the first new feature map to obtain the second new feature map;

[0024] S25. Perform dimensionality-up transformation on the channel dimension and dimensionality-down transformation on the height and width dimensions on the second new feature map.

[0025] According to one embodiment of the local rainfall intensity prediction method described in this application, step S3 further includes performing the following steps S31-S35 sequentially at least once:

[0026] S31. Normalize the input feature map and then perform separable convolution calculation;

[0027] S32. Add the feature map calculated in step S31 to the input feature map to obtain the third new feature map;

[0028] S33. Normalize the features of the third new feature map, and then perform dimensionality-upgrading and dimensionality-reducing transformation in the channel dimension;

[0029] S34. Add the transformed feature map from step S33 to the third new feature map to obtain the fourth new feature map;

[0030] S35. Perform dimensionality reduction transformation on the channel dimension and dimensionality increase transformation on the height and width dimensions on the fourth new feature map.

[0031] In one embodiment of the local rainfall intensity prediction method according to this application, step S4 further includes:

[0032] S41. Perform the first deconvolution calculation on the feature map obtained in step S3, perform dimensionality reduction transformation in the channel dimension, and dimensionality increase transformation in the height and width dimensions.

[0033] S42. Perform feature normalization on the feature map after the first deconvolution calculation;

[0034] S43. Perform the first nonlinear mapping on the normalized feature map;

[0035] S44. Perform a second deconvolution calculation on the feature map obtained in step S43, perform dimensionality reduction transformation in the channel dimension, and dimensionality increase transformation in the height and width dimensions.

[0036] S45. Perform a second nonlinear mapping on the feature map after the second deconvolution to output n frames of rainfall intensity maps.

[0037] In one embodiment of the local rainfall intensity prediction method according to this application, the method further preprocesses radar echo image data into a rainfall intensity map before step S1, specifically including:

[0038] S01, Denoise the radar echo image;

[0039] S02. Convert the denoised radar echo map z into a rainfall intensity map x according to the following formula:

[0040]

[0041] Where a = 58.53 and b = 1.56;

[0042] S03. Normalize the rainfall intensity map x according to the following formula:

[0043]

[0044] Where, x i，j Let be the pixel value of the original rainfall intensity map x at pixel (i, j), mean be the mean of all pixel values ​​in the rainfall intensity map dataset, and δ be the standard deviation of all pixel values ​​in the rainfall intensity map dataset. This represents the new pixel value at pixel (i, j) after standard normalization of the rainfall intensity map x.

[0045] To address its technical problem, this application proposes a regional rainfall intensity prediction device in a second aspect, comprising an encoder and a decoder, wherein:

[0046] The encoder comprises, in sequence:

[0047] The feature extraction module is used to extract shallow features from the preprocessed sequence of m consecutive rainfall intensity maps;

[0048] At least one first Metaformer module is used to perform dimensionality reduction transformation on the shallow features based on Metaformer to extract deep features;

[0049] The decoder includes, in sequence:

[0050] At least one second Metaformer module is used to perform a dimensionality-up transformation on the deep features based on Metaformer to output high-dimensional features;

[0051] The mapping module is used to map the high-dimensional features to the target output, i.e., the subsequent n-frame sequence of rainfall intensity maps, where 0 <m≤n≤20。

[0052] According to one embodiment of the local rainfall intensity prediction device described in this application, the device further includes a preprocessing module for preprocessing radar echo image data into a rainfall intensity map.

[0053] To address its technical problem, this application proposes a localized rainfall intensity prediction system in a third aspect, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the steps of the aforementioned localized rainfall intensity prediction method.

[0054] In order to solve its technical problem, this application proposes a computer-readable storage medium in a fourth aspect, which stores a computer program that, when executed by a processor, implements the steps of the local area rainfall intensity prediction method as described above.

[0055] The local area rainfall intensity prediction method, apparatus, system, and computer-readable storage medium of this application have the following beneficial effects: The local area rainfall intensity prediction method and apparatus according to the embodiments of this application combine an encoder-decoder structure and a Metaformer module to fit the spatiotemporal evolution process of the rainfall intensity map sequence. By using several Metaformer modules with large convolution kernels to capture spatiotemporal features over a larger range, it is possible to better fit the rainfall evolution process over a larger spatiotemporal range in the radar echo map sequence, thereby improving the accuracy of local area rainfall intensity prediction. Attached Figure Description

[0056] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0057] Figure 1 This is a schematic diagram of the logic structure of a local area rainfall intensity prediction device according to an embodiment of this application;

[0058] Figure 2 This is a schematic diagram of the logical structure of the feature extraction module in one embodiment of this application;

[0059] Figure 3 This is a schematic diagram of the logical structure of the Metaformer module in one embodiment of this application;

[0060] Figure 4 This is a schematic diagram of the logical structure of a separable convolution module in one embodiment of this application;

[0061] Figure 5 This is a schematic diagram of the logical structure of a feedforward network module in one embodiment of this application;

[0062] Figure 6 This is a schematic diagram of the logical structure of a transition module in one embodiment of this application;

[0063] Figure 7 This is a schematic diagram of the logical structure of the mapping module in one embodiment of this application;

[0064] Figure 8 This is a flowchart of a local area rainfall intensity prediction method according to an embodiment of this application;

[0065] Figure 9 This is a schematic diagram of the logical structure of a local area rainfall intensity prediction system according to an embodiment of this application. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, the embodiments and features described herein can be combined with each other unless otherwise specified.

[0067] This application combines encoder and decoder structures with Metaformer architecture to fit the spatiotemporal evolution of rainfall intensity map sequences, constructing a Metaformer-based local regional rainfall intensity prediction model. The Metaformer-based local regional rainfall intensity prediction model is trained using rainfall intensity maps from historical data to solve for various parameters of the model. Then, the trained Metaformer-based local regional rainfall intensity prediction model can be used to predict local regional rainfall intensity maps, calculating subsequent n frames of rainfall intensity maps from the input m-frame rainfall intensity map sequence.

[0068] Accordingly, this application proposes a local area rainfall intensity prediction device. Figure 1 A schematic diagram of the logic structure of a local area rainfall intensity prediction device 100 according to an embodiment of this application is shown. Figure 1As shown in the figure, the local area rainfall intensity prediction device 100 is composed of an encoder 10 and a decoder 20. The input is m frames of preprocessed past rainfall intensity maps, and the output is the subsequent n frames of rainfall intensity maps. Generally, 0 < m ≤ n ≤ 20. The encoder 10 is sequentially composed of a feature extraction module 11 and two first Metaformer modules 12a and 12b. The decoder 20 is sequentially composed of two second Metaformer modules 21a and 21b and a mapping module 22. Among them, the feature extraction module 11 of the encoder 10 extracts shallow features from the continuous m-frame rainfall intensity map sequence after preprocessing. The two first Metaformer modules 12a and 12b of the encoder 10 sequentially perform dimensionality reduction transformation on the shallow features based on Metaformer to extract deep features. The two second Metaformer modules 21a and 21b of the decoder 20 sequentially perform dimensionality increase transformation on the deep features based on Metaformer to output high-dimensional features. The mapping module 22 of the decoder 20 maps the high-dimensional features to the target output, that is, the subsequent n-frame rainfall intensity map sequence. The encoder-decoder architecture of the local area rainfall intensity prediction device 100 according to the embodiment of the present application is designed to first reduce the dimension and then increase the dimension, aiming to extract effective deep representations of rainfall spatio-temporal patterns through an autoencoding mechanism. The transformations of all modules are implemented by neural network layers and the parameters of the neural network layers are solved through data driving.

[0069] For specific reference Figure 2 As shown in the figure, the feature extraction module 11 is further sequentially composed of a point-wise convolution layer (Point-wised Convolution Layer) 111, a batch normalization layer (BNL, Batch Normalization Layer) 112, a ReLU activation function layer 113, a point-wise convolution layer 114, a batch normalization layer 115, and a ReLU activation function layer 116. Among them, the point-wise convolution layer 111 performs the first point-wise convolution on the continuous m-frame rainfall intensity maps, performs feature transformation in three dimensions of channels, height, and width. The batch normalization layer 112 performs the first feature normalization on the feature map after the first point-wise convolution. The ReLU activation function layer 113 performs the first non-linear mapping on the feature map after the first normalization process. The point-wise convolution layer 114 performs the second point-wise convolution on the feature map obtained by the ReLU activation function layer 113. The batch normalization layer 115 performs the second feature normalization on the feature map after the second point-wise convolution. Then, the ReLU activation function layer 116 performs the second non-linear mapping on the feature map after the second normalization process to obtain the shallow features of the input rainfall intensity map. Among them, the convolution kernels of the two point-wise convolution layers 111 and 114 can be 3×3, and the downsampling rate is 1 / 2.

[0070] Thus, the input of the feature extraction module 11 is m frames of rainfall intensity maps x ∈ Rm×H×W After the first pointwise convolutional layer 111 calculates, the output feature map is completed. The output feature map is calculated after the second pointwise convolutional layer 114. Feature extraction module 11 finally outputs feature maps Where R represents a multidimensional array or tensor, m is the number of feature map channels, H is the height of the rainfall intensity map, and W is the width of the rainfall intensity map.

[0071] See details Figure 3 As shown, the two first Metaformer modules 12a and 12b and the two second Metaformer modules 21a and 21b are each composed of a batch normalization layer 121, a separable convolution module 122, a skip connection module 123, a batch normalization layer 124, a feed forward network module 125, a skip connection module 126, and a transition module 127 in sequence. For the two first Metaformer modules 12a and 12b of encoder 10, batch normalization layer 121 normalizes the input feature map, then separable convolution module 122 performs separable convolution calculation, skip connection module 123 adds the feature map calculated by separable convolution module 122 to the input feature map to obtain a first new feature map, batch normalization layer 124 normalizes the first new feature map, then feedforward network module 125 performs dimensionality increase and then dimensionality decrease transformation in the channel dimension, skip connection module 126 adds the feature map transformed by feedforward network module 125 to the first new feature map to obtain a second new feature map, then transition module 127 performs dimensionality increase transformation in the channel dimension and dimensionality decrease transformation in the height and width dimensions on the second new feature map. For the two second Metaformer modules 21a and 21b of the decoder 20, the batch normalization layer 121 normalizes the input feature map, and then the separable convolution module 122 performs separable convolution calculation. The skip connection module 123 adds the feature map calculated by the separable convolution module 122 to the input feature map to obtain a third new feature map. The batch normalization layer 124 normalizes the third new feature map, and then the feedforward network module 125 performs dimensionality increase and then dimensionality decrease transformation in the channel dimension. The skip connection module 126 adds the feature map transformed by the feedforward network module 125 to the third new feature map to obtain a fourth new feature map. Then the transition module 127 performs dimensionality decrease transformation in the channel dimension and dimensionality increase transformation in the height and width dimensions on the fourth new feature map.

[0072] Therefore, the input to the first Metaformer module 12a is the feature map. The first Metaformer module 12a ultimately outputs a feature map. The input to the first Metaformer module 12b is a feature map. The first Metaformer module 12b ultimately outputs a feature map. The input to the second Metaformer module 21a is the feature map. The second Metaformer module 21a ultimately outputs a feature map. The input to the second Metaformer module 21b is the feature map. The second Metaformer module 21b ultimately outputs a feature map. Where H represents the height of the rainfall intensity map and W represents the width of the rainfall intensity map.

[0073] See details Figure 4 As shown, the separable convolutional module 122 in each Metaformer module is further composed of a channel-wise convolutional layer 1221, a batch normalization layer 1222, a ReLU activation function layer 1223, a pointwise convolutional layer 1224, a batch normalization layer 1225, a ReLU activation function layer 1226, a channel-wise convolutional layer 1227, a batch normalization layer 1228, and a ReLU activation function layer 1229 in sequence. Specifically, pointwise convolutional layer 1224 performs feature transformations in the height and width dimensions of the feature map, while channelwise convolutional layers 1221 and 1227 perform feature transformations in the channel dimension of the feature map. Batch normalization layers 1222, 1225, and 1228 normalize the feature map, and ReLU activation function layers 1223, 1226, and 1229 perform nonlinear mappings on the feature map. The purpose is to separate the channel dimension during the feature transformation process, enabling the module to perform feature transformations with a relatively small number of parameters. The kernel size of pointwise convolutional layer 1224 is k×k, while the kernel size of channelwise convolutional layers 1221 and 1227 is 1×1. Specifically, the value of k is preferably 31 to use a large kernel to capture a wider range of spatiotemporal features.

[0074] Therefore, the input to the separable convolutional module 122 is the feature map f∈R. c×h×w Then the input to the first channel-wise convolutional layer 1221 is the feature map f∈R. c×h×w The output is a feature map f∈R 4c×h×w The input to the pointwise convolutional layer 1224 is the feature map f∈R. 4c×h×w The output is a feature map f∈R 4c×h×w The input to the second channel-wise convolutional layer 1227 is the feature map f∈R. 4c×h×w The output is a feature map f∈R c×h×wThe separable convolutional module 122 ultimately outputs a feature map f∈R. c×h×w Where c is the number of feature map channels, h is the height of the feature map, and w is the width of the feature map.

[0075] See details Figure 5 As shown, the feedforward network module 125 in each Metaformer module is further composed of a batch normalization layer 1251, a channel-wise convolutional layer 1252, a batch normalization layer 1253, a GeLU activation function layer 1254, a channel-wise convolutional layer 1255, a batch normalization layer 1256, and a ReLU activation function layer 1257 in sequence. Specifically, the channel-wise convolutional layers 1252 and 1255 perform feature transformation along the channel dimension of the feature map, the activation function layers 1254 and 1257 perform non-linear mapping on the feature map, and the batch normalization layers 1251, 1253, and 1256 normalize the feature map. This enhances the model's feature extraction capability by performing dimensionality-upgrading and then dimensionality-reducing transformations on the feature map along the channel dimension. The kernel size of the channel-wise convolutional layers 1252 and 1255 is 1×1.

[0076] Therefore, the input to the feedforward network module 125 is the feature map f∈R. c×h×w The input to the first channel-wise convolutional layer 1252 is the feature map f∈R. c×h×w The output is a feature map f∈R 4c×h×w The input to the second channel-wise convolutional layer 1255 is the feature map f∈R. 4c×h×w The output is a feature map f∈R c×h×w The feedforward network module 125 ultimately outputs a feature map f∈R. c×h×w Where c is the number of feature map channels, h is the height of the feature map, and w is the width of the feature map.

[0077] See details Figure 6 As shown, the transition module 127 in each Metaformer module is further composed of a channel-wise convolutional layer 1271, a batch normalization layer 1272, a ReLU activation function layer 1273, a pointwise convolutional layer 1274, a batch normalization layer 1275, and a ReLU activation function layer 1276 in sequence. Specifically, the pointwise convolutional layer 1274 performs feature transformations in the height and width dimensions of the feature map, the channel-wise convolutional layer 1271 performs feature transformations in the channel dimension of the feature map, the ReLU activation function layers 1273 and 1276 perform non-linear mappings on the feature map, and the batch normalization layers 1272 and 1275 normalize the feature map, thereby achieving dimensionality increase or decrease of the input feature map. The kernel size of the pointwise convolutional layer 1274 is 3×3, and the kernel size of the channel-wise convolutional layer 1271 is 1×1.

[0078] Therefore, when transition module 127 is a submodule of the first Metaformer modules 12a and 12b, the input of transition module 127 is the feature map f∈R. c×h×w The input to the channel-wise convolutional layer 1271 is the feature map f∈R. c×h×w The output is a feature map f∈R 2c×h×w The input to the pointwise convolutional layer 1274 is the feature map f∈R. 2c×h×w The output is a feature map. Transition module 127 finally outputs feature map When transition module 127 is used as a submodule of the second Metaformer modules 21a and 21b, the input of transition module 127 is the feature map f∈R. c×h×w The input to the channel-wise convolutional layer 1271 is the feature map f∈R. c×h×w The output is a feature map. The input to the pointwise convolutional layer 1274 is the feature map. The output is a feature map. Transition module 127 finally outputs feature map Where c is the number of feature map channels, h is the height of the feature map, and w is the width of the feature map.

[0079] See also Figure 7 As shown, the mapping module 22 is further composed of a deconvolution layer 221, a batch normalization layer 222, a ReLU activation function layer 223, a deconvolution layer 224, and a ReLU activation function layer 225 in sequence. Specifically, the deconvolution layer 221 performs a first deconvolution calculation on the feature map obtained from the second Metaformer module 21b, performing dimensionality reduction in the channel dimension and dimensionality increase in the height and width dimensions. The batch normalization layer 222 normalizes the feature map after the first deconvolution calculation. The ReLU activation function layer 223 performs a first nonlinear mapping on the normalized feature map. Then, the deconvolution layer 224 performs a second deconvolution calculation on the obtained feature map, performing dimensionality reduction in the channel dimension and dimensionality increase in the height and width dimensions. Finally, the ReLU activation function layer 225 performs a second nonlinear mapping on the feature map after the second deconvolution, outputting the target n-frame rainfall intensity map.

[0080] Therefore, the input to the mapping module 22 is the feature map f∈R. c×h×w The input to the first deconvolutional layer 221 is the feature map f∈R. c×h×w The output is a feature map. The input to the second deconvolutional layer 224 is the feature map. The output is a feature map f∈R n×H×W The mapping module 22 ultimately outputs n frames of rainfall intensity maps f∈R.n×H×W This represents the prediction result of the local rainfall intensity prediction device 100. Where c is the number of feature map channels, h is the height of the feature map, w is the width of the feature map, H is the height of the rainfall intensity map, and W is the width of the rainfall intensity map.

[0081] Furthermore, the local area rainfall intensity prediction device 100 in the above embodiments of this application also includes a preprocessing module for preprocessing radar echo image data into a rainfall intensity map before performing local area rainfall intensity prediction. The specific implementation of radar echo map preprocessing will be described in detail later.

[0082] According to the local rainfall intensity prediction device 100 of the above embodiments of this application, a local rainfall intensity prediction model based on Metaformer is constructed. The training of the model is carried out in two stages: in the first stage, consecutive m+n frames of preprocessed rainfall intensity maps are randomly selected from historical data for training; in the second stage, consecutive m+n frames of preprocessed rainfall intensity maps are selected from historical data in a window of size m for training. Wherein, m is the number of input past rainfall intensity map sequence frames, and n is the number of target rainfall intensity map sequence frames. The loss function for model training is the sum of balanced mean square error B-MSE (see formula (1)) and balanced absolute error B-MAE (see formula (4)), and the gradient descent optimizer is used to solve for the parameters of the convolutional layer and batch normalization layer of the model.

[0083]

[0084]

[0085] In formulas (1) and (2), N is the number of frames in the rainfall intensity map sequence, H and W are the height and width of the rainfall intensity map, respectively, and x n，i，j For the observation value at position (i, j) of the rainfall intensity map in frame n, w is the predicted value for the location (i, j) of the rainfall intensity map in the nth frame. n，i，j Let (i, j) be the error weight for the position (i, j) of the rainfall intensity map in the nth frame. The weight for each frame of the rainfall intensity map is calculated as shown in the following formula (3):

[0086]

[0087] Where x is the rainfall intensity value before standard normalization.

[0088] Based on the aforementioned local rainfall intensity prediction device, this application also proposes a local rainfall intensity prediction method. Figure 8A flowchart of a localized rainfall intensity prediction method 200 according to an embodiment of this application is shown. See also Figure 8 As shown, the local rainfall intensity prediction method 200 includes the following steps:

[0089] In step S210, the radar echo image data is preprocessed into a rainfall intensity map. According to a specific embodiment of this application, preprocessing the radar echo image data into a rainfall intensity map specifically includes:

[0090] Step S211: Denoise the radar echo image;

[0091] Step S212, convert the denoised radar echo map z into a rainfall intensity map x according to the following formula (4):

[0092]

[0093] Where a = 58.53 and b = 1.56;

[0094] Step S213: Normalize the rainfall intensity map x according to the following formula (5):

[0095]

[0096] Where, x i，j Let be the pixel value of the original rainfall intensity map x at pixel (i, j), mean be the mean of all pixel values ​​in the rainfall intensity map dataset, and δ be the standard deviation of all pixel values ​​in the rainfall intensity map dataset. This represents the new pixel value at pixel (i, j) after standard normalization of the rainfall intensity map x.

[0097] Then, in step S220, shallow features are extracted from the preprocessed sequence of m consecutive rainfall intensity maps. According to a specific embodiment of this application, the extraction of shallow features in step S220 further includes:

[0098] Step S221: Perform the first pointwise convolution on the continuous m-frame rainfall intensity map, and perform feature transformation in the three dimensions of channel, height and width;

[0099] Step S222: Perform the first feature normalization on the feature map after the first pointwise convolution;

[0100] Step S223: Perform the first nonlinear mapping on the feature map after the first normalization process;

[0101] Step S224: Perform a second pointwise convolution on the feature map obtained in step S13;

[0102] Step S225: Perform a second feature normalization on the feature map that has undergone the second pointwise convolution;

[0103] Step S226: Perform a second nonlinear mapping on the feature map after the second normalization process to obtain the shallow features of the input rainfall intensity map.

[0104] Then, in step S230, the shallow features are subjected to dimensionality reduction transformation based on the Metaformer module to extract deep features. According to a specific embodiment of this application, step S230, which involves performing dimensionality reduction transformation on the shallow features obtained in step S220 based on Metaformer to extract deep features, further includes:

[0105] Step S231: Normalize the input feature map and then perform separable convolution calculation.

[0106] Step S232: Add the feature map calculated in step S231 to the input feature map to obtain the first new feature map;

[0107] Step S233: Normalize the features of the first new feature map, and then perform dimensionality increase and then dimensionality decrease transformation in the channel dimension;

[0108] Step S234: Add the transformed feature map from step S233 to the first new feature map to obtain the second new feature map;

[0109] Step S235: Perform dimensionality-up transformation on the channel dimension and dimensionality-down transformation on the height and width dimensions on the second new feature map.

[0110] In a further embodiment of this application, steps S231-S235 can be executed sequentially at least once in step S230. For example, steps S231-S235 can be executed again on the feature map output by step S235 to finally extract the deep features of the rainfall intensity map.

[0111] Then, in step S240, the deep features are subjected to dimensionality upscaling transformation based on the Metaformer module to output high-dimensional features. According to a specific embodiment of this application, step S240, which involves performing dimensionality upscaling transformation on the deep features obtained in step S230 based on the Metaformer module to output high-dimensional features, further includes:

[0112] Step S241: Normalize the input feature map and then perform separable convolution calculation.

[0113] Step S242: Add the feature map calculated in step S241 to the input feature map to obtain a third new feature map;

[0114] Step S243: Normalize the features of the third new feature map, and then perform dimensionality increase and then dimensionality decrease transformation in the channel dimension;

[0115] Step S244: Add the feature map obtained after the transformation in step S243 to the third new feature map to obtain a fourth new feature map;

[0116] Step S245: Perform a dimensionality reduction transformation on the fourth new feature map in the channel dimension and a dimensionality increase transformation in the height and width dimensions.

[0117] In a further embodiment of the present application, steps S241 - S245 in step S240 can be cyclically executed at least once in sequence. For example, the feature map output in step S245 can be used to execute steps S241 - S245 again, and finally, the high-dimensional features of the rainfall intensity map are extracted.

[0118] Then in step S250, map the high-dimensional features to the target output, that is, the subsequent n-frame rainfall intensity map sequence, where 0 < m ≤ n ≤ 20. In a specific embodiment of the present application, step S250 further includes:

[0119] Step S251: Perform a first deconvolution calculation on the feature map obtained in step S240, perform a dimensionality reduction transformation in the channel dimension, and a dimensionality increase transformation in the height and width dimensions;

[0120] Step S252: Perform feature normalization on the feature map after the first deconvolution calculation;

[0121] Step S253: Perform a first non-linear mapping on the feature map after the normalization process;

[0122] Step S254: Perform a second deconvolution calculation on the feature map obtained in step S253, perform a dimensionality reduction transformation in the channel dimension, and a dimensionality increase transformation in the height and width dimensions;

[0123] Step S255: Perform a second non-linear mapping on the feature map after the second deconvolution, and output the target n-frame rainfall intensity map.

[0124] The local area rainfall intensity prediction method 200 according to the above embodiments of the present application is proposed based on the local area rainfall intensity prediction device 100 in the foregoing embodiments. For further specific implementations of each method step, reference can also be made to the detailed descriptions of each module of the local area rainfall intensity prediction device 100 above.

[0125] Based on the above local area rainfall intensity prediction method of the present application, the present application also proposes a local area rainfall intensity prediction system 300. Refer to Figure 9As shown, the local rainfall intensity prediction system 300 includes a processor 310 and a memory 320, which are communicatively connected. The memory 320 stores a computer program, which, when executed by the processor 310, causes the processor 310 to implement the steps of the local rainfall intensity prediction method of the foregoing embodiments of this application.

[0126] This application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the local area rainfall intensity prediction method 300 of the foregoing embodiments of this application.

[0127] The local rainfall intensity prediction method and apparatus according to the above embodiments of this application combine an encoder-decoder structure and Metaformer modules to fit the spatiotemporal evolution of rainfall intensity map sequences. By employing several Metaformer modules with large convolutional kernels to capture spatiotemporal features over a larger range, it can better fit the rainfall evolution process over a larger spatiotemporal range in radar echo map sequences, thereby improving the prediction accuracy of local rainfall intensity. Experiments were conducted on the HKO-7 dataset using the local rainfall intensity prediction method and apparatus according to the above embodiments of this application. Taking a 5-frame input and 20-frame prediction as an example, the comparison results between the proposed model and Last Frame, ROVER, 2d CNN, 3d CNN, and TrajGRU models are shown in Tables 1 and 2. Specifically, the model's prediction evaluation uses common rainfall prediction evaluation metrics, namely: CSI (Critical Success Index) and HSS (Heidke Skill Score). As shown in Tables 1 and 2, compared with existing models, this application achieves higher prediction accuracy in both CSI and HSS metrics.

[0128] Table 1: Comparison of CSI index results (r represents rainfall intensity value)

[0129]

[0130] Table 2: Comparison of HSS index results (r represents rainfall intensity value)

[0131]

[0132] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting rainfall intensity in a local area, characterized in that, Includes the following steps: S1. Extract shallow features from the preprocessed sequence of m consecutive rainfall intensity maps; S2. Based on Metaformer, the shallow features are subjected to dimensionality reduction transformation to extract deep features; S3. Based on Metaformer, perform dimensionality up-transformation on the deep features to output high-dimensional features; S4. Map the high-dimensional features to the target output, i.e., the subsequent n-frame rainfall intensity map sequence, where 0 <m≤n≤20; Step S2 further includes performing the following steps S21-S25 sequentially at least once: S21. Normalize the input feature map and then perform separable convolution calculation; S22. Add the feature map calculated in step S21 to the input feature map to obtain the first new feature map; S23. Normalize the first new feature map, and then perform dimensionality increase and then dimensionality decrease transformation in the channel dimension; S24. Add the transformed feature map from step S23 to the first new feature map to obtain the second new feature map; S25. Perform dimensionality-increasing transformation on the channel dimension and dimensionality-reducing transformation on the height and width dimensions on the second new feature map respectively; Step S3 further includes performing the following steps S31-S35 sequentially at least once: S31. Normalize the input feature map and then perform separable convolution calculation; S32. Add the feature map calculated in step S31 to the input feature map to obtain the third new feature map; S33. Normalize the features of the third new feature map, and then perform dimensionality-upgrading and dimensionality-reducing transformation in the channel dimension; S34. Add the transformed feature map from step S33 to the third new feature map to obtain the fourth new feature map; S35. Perform dimensionality reduction transformation on the channel dimension and dimensionality increase transformation on the height and width dimensions on the fourth new feature map.

2. The method for predicting local rainfall intensity according to claim 1, characterized in that, The extraction of shallow features in step S1 further includes: S11. Perform the first pointwise convolution on the continuous m-frame rainfall intensity map, and perform feature transformation in the three dimensions of channel, height and width. S12. Perform the first feature normalization on the feature map after the first pointwise convolution; S13. Perform the first nonlinear mapping on the feature map after the first normalization process; S14. Perform a second pointwise convolution on the feature map obtained in step S13; S15. Perform a second feature normalization on the feature map that has undergone the second pointwise convolution. S16. Perform a second nonlinear mapping on the feature map after the second normalization process to obtain the shallow features of the input rainfall intensity map.

3. The method for predicting local rainfall intensity according to claim 1, characterized in that, Step S4 further includes: S41. Perform the first deconvolution calculation on the feature map obtained in step S3, perform dimensionality reduction transformation in the channel dimension, and dimensionality increase transformation in the height and width dimensions. S42. Perform feature normalization on the feature map after the first deconvolution calculation; S43. Perform the first nonlinear mapping on the normalized feature map; S44. Perform a second deconvolution calculation on the feature map obtained in step S43, perform dimensionality reduction transformation in the channel dimension, and dimensionality increase transformation in the height and width dimensions. S45. Perform a second nonlinear mapping on the feature map after the second deconvolution to output n frames of rainfall intensity maps.

4. The method for predicting local rainfall intensity according to claim 1, characterized in that, The method further preprocesses the radar echo image data into a rainfall intensity map before step S1, specifically including: S01, Denoise the radar echo image; S02. Convert the denoised radar echo map z into a rainfall intensity map x according to the following formula: Where a = 58.53 and b = 1.56; S03. Normalize the rainfall intensity map x according to the following formula: Where, x i,j Let be the pixel value of the original rainfall intensity map x at pixel (i,j), mean be the mean of all pixel values ​​in the rainfall intensity map dataset, and δ be the standard deviation of all pixel values ​​in the rainfall intensity map dataset. This represents the new pixel value at pixel (i,j) after standard normalization of the rainfall intensity map x.

5. A device for predicting rainfall intensity in a local area, characterized in that, Includes encoders and decoders, where: The encoder comprises, in sequence: The feature extraction module is used to extract shallow features from the preprocessed sequence of m consecutive rainfall intensity maps; At least one first Metaformer module is used to perform dimensionality reduction transformation on the shallow features based on Metaformer to extract deep features; The decoder includes, in sequence: At least one second Metaformer module is used to perform a dimensionality-up transformation on the deep features based on Metaformer to output high-dimensional features; The mapping module is used to map the high-dimensional features to the target output, i.e., the subsequent n-frame sequence of rainfall intensity maps, where 0 <m≤n≤20; Wherein, each of the at least one first Metaformer module is used in a cyclical manner for: The input feature map is normalized, and then separable convolution is performed. The feature map calculated by the separable convolution is added to the input feature map to obtain the first new feature map; The first new feature map is normalized, and then dimensionality-up and dimensionality-down transformation is performed on the channel dimension. The transformed feature map is added to the first new feature map to obtain the second new feature map; The second new feature map is subjected to dimensionality increase transformation in the channel dimension and dimensionality decrease transformation in the height and width dimensions, respectively. Each of the at least one second Metaformer module is used sequentially in a loop for: The input feature map is normalized, and then separable convolution is performed. The feature map calculated by the separable convolution is added to the input feature map to obtain a third new feature map; The third new feature map is normalized, and then dimensionality-upgrading and dimensionality-reducing transformations are performed on the channel dimension. The transformed feature map is added to the third new feature map to obtain the fourth new feature map; The fourth new feature map is subjected to dimensionality reduction transformation in the channel dimension and dimensionality increase transformation in the height and width dimensions.

6. The local rainfall intensity prediction device according to claim 5, characterized in that, It also includes a preprocessing module for preprocessing radar echo image data into a rainfall intensity map.

7. A localized rainfall intensity prediction system, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the steps of the local area rainfall intensity prediction method as described in any one of claims 1-4.

8. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, implements the steps of the local area rainfall intensity prediction method as described in any one of claims 1-4.

Images