Infrared rainfall prediction and model training method and device, and electronic equipment

Abstract

The invention relates to the technical field of deep learning, and particularly relates to an infrared rainfall prediction and model training method and device, and electronic equipment, and the method comprises the steps: obtaining a training set, each sample in the training set comprising a multi-channel brightness temperature feature of a sample region and a corresponding rainfall feature true value; a preset infrared rainfall prediction model is obtained, the infrared rainfall prediction model comprises an input layer, an encoder, a middle layer, a decoder and an output layer, the encoder is additionally provided with an attention mechanism module before down-sampling each time, the middle layer comprises a cavity space pyramid pooling module, the decoder is additionally provided with an attention mechanism module after up-sampling each time, and the output layer comprises a cavity space pyramid pooling module; a weighted mean square error loss function of the model dynamically adjusts an error weight according to precipitation spatial heterogeneity; and training a preset infrared rainfall prediction model by using the training set to obtain a trained infrared rainfall prediction model. According to the technical scheme, infrared rainfall estimation can be carried out more accurately.

Description

Technical Field

[0001] This disclosure relates to the field of deep learning technology, specifically to an infrared precipitation prediction and model training method, apparatus, and electronic equipment. Background Technology

[0002] Over the past two decades, floods and droughts have occurred frequently on both global and regional scales, making in-depth research on precipitation crucial for addressing the challenges of climate change. Geostationary meteorological satellites such as Fengyun-4A / B, the Himawari-8 / 9 series, and the GOES series are equipped with visible-infrared sensors, capable of rapidly capturing short-term, highly variable weather phenomena. Through remote sensing technology, they can provide continuous and high-spatiotemporal resolution information on various cloud parameters, thereby estimating surface precipitation. To obtain large-scale, high-resolution precipitation data, precipitation estimation based on infrared multi-band data has become a research hotspot in this field.

[0003] With the gradual maturation of deep learning technology, it has demonstrated enormous potential in the field of precipitation research. Currently, deep learning can simulate nonlinear precipitation phenomena and generate realistic and spatiotemporally consistent prediction results based on deep neural network models trained with large amounts of observational data. Although existing deep learning models perform well in tasks such as precipitation estimation, they still have some limitations. For example, their ability to capture complex meteorological features and multi-scale information is limited, and the features involved in precipitation estimation, such as cloud layers and precipitation intensity, are widely distributed with large scale differences, and the output is prone to bias towards the majority category (i.e., non-precipitation areas). Therefore, there is an urgent need for an infrared precipitation prediction method to address these existing problems. Summary of the Invention

[0004] To address the problems in related technologies, this disclosure provides an infrared precipitation prediction and model training method, apparatus, and electronic device.

[0005] In a first aspect, this disclosure provides an infrared precipitation prediction and model training method, including:

[0006] Obtain a training set, which includes multiple samples, each of which includes the multi-channel brightness temperature features of a sample region and the corresponding precipitation feature ground values;

[0007] A preset infrared precipitation prediction model is obtained. The infrared precipitation prediction model includes an input layer, an encoder, an intermediate layer, a decoder, and an output layer. The encoder adds an attention mechanism module before each downsampling. The intermediate layer includes a hollow spatial pyramid pooling module. The decoder adds an attention mechanism module after each upsampling. The loss function of the infrared precipitation prediction model includes a weighted mean square error loss function. The weighted mean square error loss function dynamically adjusts the error weights according to the spatial heterogeneity of precipitation.

[0008] The training set is used to train the preset infrared precipitation prediction model to obtain the trained infrared precipitation prediction model.

[0009] In one possible implementation, the weighted mean square error loss function weights the error according to the magnitude of the true value of the precipitation characteristics.

[0010] In one possible implementation, the attention mechanism module includes a convolutional block attention module.

[0011] In one possible implementation, obtaining the training set includes:

[0012] Acquire raw brightness temperature data from the multi-channel scanning imaging radiometer AGRI and raw half-hour precipitation data from the IMERG Final, the final product of global precipitation measurement and multi-satellite precipitation fusion;

[0013] Data within the acceptable range are selected from the multi-channel brightness temperature raw data and the IMERG Final half-hour precipitation raw data to obtain multi-channel brightness temperature data and IMERG Final half-hour precipitation data;

[0014] The multi-channel brightness temperature data and the IMERG Final half-hour precipitation data were spatiotemporally matched.

[0015] Extract multi-channel brightness temperature feature data from the matched multi-channel brightness temperature data and precipitation feature data from the IMERG Final half-hour precipitation data;

[0016] The multi-channel brightness temperature feature data and the precipitation feature data are normalized to obtain the normalized multi-channel brightness temperature feature and the true value of the precipitation feature.

[0017] In one possible implementation, the spatiotemporal matching of the multi-channel brightness temperature data and the IMERG Final half-hour precipitation data includes...

[0018] Based on the start observation time of the multi-channel brightness temperature data, time matching is performed with the corresponding IMERG Final half-hour precipitation data.

[0019] Based on the nominal image points of the multi-channel brightness temperature data, spatial matching is performed with the nominal image points of the IMERG Final half-hour precipitation data using the nearest neighbor method.

[0020] In one possible implementation, the extraction of multi-channel brightness temperature features from the matched multi-channel brightness temperature data includes:

[0021] Extracting BT from the matched multi-channel brightness temperature data 3.75 BT 10.80 BT 12.00 ;

[0022] Calculate the brightness temperature difference ΔBT using the following formula. 10.80-6.25 :

[0023] ΔBT 10.80-6.25 =BT 10.80 -BT 6.25 ;

[0024] Calculate the brightness temperature linear recombination value BT using the following formula. multi-channel :

[0025] BT multi-channel =BT 12.00 +BT 8.55 -2*BT 10.80 ;

[0026] Obtain the BT 3.75 BT 10.80 BT 12.00 ΔBT 10.80-6.25 BT multi-channel Multi-channel brightness temperature characteristics;

[0027] Among them, BT 3.75 BT 8.55 BT 10.80 BT 12.00 These are the brightness temperatures of channels with center wavelengths of 3.75 μm, 8.55 μm, 10.80 μm, and 12.00 μm, respectively.

[0028] Secondly, embodiments of this disclosure provide an infrared precipitation prediction method, including:

[0029] Obtain the current multi-channel brightness temperature raw data of the region to be predicted;

[0030] Extract the current multi-channel brightness temperature feature data from the current multi-channel brightness temperature raw data;

[0031] The current multi-channel brightness temperature feature data is input into a pre-trained infrared precipitation prediction model to obtain the precipitation prediction value output by the infrared precipitation prediction model.

[0032] The pre-trained infrared precipitation prediction model is obtained by training the method described in the first aspect.

[0033] Thirdly, embodiments of this disclosure provide an infrared precipitation prediction model training device, comprising:

[0034] The sample acquisition module is configured to acquire a training set, wherein the training samples in the training set include ground truth values ​​of multi-channel brightness temperature features and precipitation features;

[0035] The model acquisition module is configured to acquire a preset infrared precipitation prediction model. The infrared precipitation prediction model includes an input layer, an encoder, an intermediate layer, a decoder, and an output layer. The encoder adds an attention mechanism module before each downsampling. The intermediate layer includes a hollow spatial pyramid pooling module. The decoder adds an attention mechanism module after each upsampling. The loss function of the infrared precipitation prediction model includes a weighted mean square error loss function. The weighted mean square error loss function dynamically adjusts the error weights according to the spatial heterogeneity of precipitation.

[0036] The model training module is configured to use the training set to train a preset infrared precipitation prediction model to obtain a trained infrared precipitation prediction model.

[0037] Fourthly, embodiments of this disclosure provide an infrared precipitation prediction device, comprising:

[0038] The data acquisition module is configured to acquire the current multi-channel brightness temperature raw data of the area to be predicted;

[0039] The feature extraction module is configured to extract current multi-channel brightness temperature feature data from the current multi-channel brightness temperature raw data;

[0040] The prediction module is configured to input the current multi-channel brightness temperature feature data into a pre-trained infrared precipitation prediction model to obtain the precipitation prediction value output by the infrared precipitation prediction model.

[0041] The pre-trained infrared precipitation prediction model is obtained by training the method described in the first aspect.

[0042] Fifthly, embodiments of this disclosure provide an electronic device including a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method as described in any one of the first aspects.

[0043] According to the technical solution provided in this disclosure, by introducing an attention mechanism module in the encoder and decoder, the sensitivity to key precipitation features can be enhanced, and the influence of weak precipitation intensity or background areas can be reduced. Furthermore, embedding a dilated spatial pyramid pooling module in the intermediate layer can capture both small-scale strong precipitation features and large-scale weak precipitation areas. Dilated convolution can expand the receptive field without losing resolution, enabling the model to simultaneously focus on greater contextual information. The adaptive weighted mean squared error loss mechanism dynamically adjusts the error weights based on the spatial heterogeneity of precipitation and applies physical prior constraints to the loss function to prevent extreme high-value areas from being underestimated. Thus, the model's feature extraction and modeling capabilities for precipitation targets of different scales and intensities can be significantly improved, thereby enhancing the accuracy and precision of the model's prediction results.

[0044] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0045] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings:

[0046] Figure 1 The flowchart illustrates a method for training an infrared precipitation prediction model according to an embodiment of this disclosure.

[0047] Figure 2 This diagram illustrates the structure of an infrared precipitation prediction model provided in an embodiment of the present disclosure.

[0048] Figure 3 A flowchart illustrating an infrared precipitation prediction method provided in an embodiment of this disclosure is shown.

[0049] Figure 4 The following diagrams are shown: the spatial distribution map of infrared precipitation prediction provided in this embodiment, the spatial distribution map of precipitation products from the reference source, the distribution map of prediction results from the convolutional neural network model, and the distribution map of prediction results from the long short-term memory network model.

[0050] Figure 5 This diagram illustrates the structural block diagram of the infrared precipitation prediction model training device provided in an embodiment of the present disclosure.

[0051] Figure 6 A structural block diagram of the infrared precipitation prediction device provided in an embodiment of this disclosure is shown.

[0052] Figure 7 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0053] Figure 8A schematic diagram of the structure of a computer system suitable for implementing the methods of the embodiments of this disclosure is shown. Detailed Implementation

[0054] In the following, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement them. Furthermore, for clarity, portions unrelated to the description of exemplary embodiments have been omitted from the drawings.

[0055] In this disclosure, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, figures, steps, behaviors, components, parts or combinations thereof disclosed in this specification, and are not intended to exclude the possibility of the presence or addition of one or more other features, figures, steps, behaviors, components, parts or combinations thereof.

[0056] It should also be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] Figure 1 A flowchart illustrating a method for training an infrared precipitation prediction model according to an embodiment of this disclosure is shown. Figure 1 As shown, the infrared precipitation prediction model training method includes the following steps:

[0058] In step S101, a training set is obtained, which includes multiple samples, each of which includes the true values ​​of multi-channel brightness temperature features and precipitation features of a sample region.

[0059] In step S102, a preset infrared precipitation prediction model is obtained;

[0060] In step S103, the training set is used to train the preset infrared precipitation prediction model to obtain the trained infrared precipitation prediction model.

[0061] In one possible implementation, the infrared precipitation prediction model training method can be applied to electronic devices such as computers, servers, and computer clusters that are capable of training infrared precipitation prediction models.

[0062] In one possible implementation, the training set includes multiple samples, each sample comprising a multi-channel brightness temperature feature of a sample region and its corresponding true value of precipitation features. The multi-channel brightness temperature feature includes actually measured multi-channel brightness temperature (BT) data that is related to precipitation features; the precipitation feature values ​​include features that characterize precipitation conditions, such as precipitation rate, precipitation rate quality, and other precipitation feature values.

[0063] In one possible implementation, the preset infrared precipitation prediction model is based on the UNet model and includes an input layer, encoder, intermediate layer, decoder, and output layer. The encoder adds an attention mechanism module before each downsampling step, guiding the network to focus on key regions and features before compressing information. The intermediate layer includes a dilated spatial pyramid pooling module, which utilizes parallel convolutions with different dilation rates and global average pooling to efficiently capture multi-scale contextual information and global background without sacrificing resolution. The decoder adds an attention mechanism module after each upsampling step, allowing for fine-tuning of feature weights after restoring feature resolution, optimizing feature fusion and detail reconstruction.

[0064] Here, the specific embedding position of the attention mechanism module in the encoder and decoder is crucial. Embedding the attention mechanism module before downsampling in the encoder allows the network to pre-evaluate and enhance channels and spatial regions crucial for precipitation prediction while preserving original resolution details, and suppressing irrelevant information, thus guiding subsequent downsampling to more effectively retain valuable information. Conversely, placing the attention mechanism module after upsampling in the decoder is because upsampling generates coarser feature maps and may introduce artifacts; applying the attention mechanism module here allows for immediate refinement of high-resolution features, adjusting feature weights to highlight key precipitation structures and boundary details, optimizing the fusion process of skip connections between the decoder and encoder, and ultimately improving the spatial accuracy of precipitation intensity distribution reconstruction.

[0065] In one possible implementation, the Atrous Spatial Pyramid Pooling (ASPP) module comprises three parts: a 1x1 standard convolutional branch (dilution rate = 1) for capturing local details; three 3x3 dilated convolutional branches with progressively increasing dilution rates (e.g., dilution rates of 6, 12, and 18) to expand the receptive field and capture cloud structures for medium, large, and very large-scale precipitation; and a Global Average Pooling (GAP) branch (containing a 1x1 convolution that upsamples the output to the original spatial size) to provide global contextual background at the entire image level. This configuration of dilution rates (6, 12, and 18) aims to cover key precipitation scales ranging from tens to hundreds of kilometers while avoiding the risks of grid artifacts and information discontinuities caused by excessively high dilution rates. The addition of the GAP branch addresses the high dependence of precipitation estimation on regional background. The configuration of the aforementioned Hollow Spatial Pyramid Pooling (ASPP) module enables the model to simultaneously and accurately perceive the characteristics of precipitation systems at different spatial scales in infrared images, and fully integrate key global environmental information, thereby significantly improving the accuracy of identifying precipitation structures and extents at different scales, as well as estimating overall precipitation intensity.

[0066] Example, Figure 2 This diagram illustrates the structure of an infrared precipitation prediction model provided in an embodiment of the present disclosure, as shown below. Figure 2 As shown, the input layer can be a dual 3×3 convolutional module (with ReLU activation function and same padding). The input data received by the infrared precipitation prediction model has a size of n×n×c, where n×n is the pixel size of the input image and c is the number of multi-channel brightness and temperature features. After passing through the dual 3×3 convolutional module, a feature map of size n×n×32 can be generated. Then, the n×n×32 feature map is input to the encoder, which may include... Figure 2 The four downsampling modules shown each include a dual 3×3 convolutional layer, an attention mechanism module, and a max-pooling layer. Downsampling is performed through the max-pooling layer, with image sizes as shown. Figure 2The model gradually decreases the size of the feature map, undergoing four downsampling operations to finally output a feature map of size (n / 16)×(n / 16)×512. To make the model focus more on key regions, an attention mechanism module is added to the encoder before each downsampling. The (n / 16)×(n / 16)×512 feature map output by the encoder can be further processed by an intermediate layer, which includes an ASPP module to extract features at multiple precipitation scales (such as large-scale precipitation, meso-scale precipitation, and small-scale precipitation), enhancing the model's ability to process information corresponding to different precipitation scales. The feature map output by this intermediate layer maintains a size of (n / 16)×(n / 16)×512. Then, in the decoder stage, the model gradually restores the spatial resolution of the feature map through deconvolution operations (upsampling). This decoder may include... Figure 2 The four upsampling modules shown each include a deconvolution module, an attention mechanism module, and two 3×3 convolutional layers. After each upsampling, a 3×3 convolution operation is performed, and the feature maps from the corresponding layers of the encoder are fused using the attention mechanism. The added attention mechanism helps the model focus on more important regions in the input image, extracting more detailed features. After each upsampling, the number of channels is 256, 128, 64, and 32 respectively, ultimately restoring the n×n feature map of the input data with 32 channels. Finally, after the output layer, which is a 1×1 convolution operation, the model reduces the number of channels in the feature map to 1, and the output data has a size of n×n×1, serving as the final output of the model.

[0067] In one possible implementation, during model training, an appropriate batch size can be selected based on the available GPU (Graphics Processing Unit) memory on the electronic device being trained. The batch size should not be set too small or too large to ensure training efficiency. The number of training epochs is set to 500 to 1000, which can be adjusted according to specific circumstances. The optimizer can be selected from Adam Adaptive Moment Estimation or SGD (Stochastic Gradient Descent), etc., and the learning rate should be kept within a reasonable range, typically between 1e-4 and 1e-5. After configuring these parameters, model training can be performed on the GPU, and the model weights with optimal performance can be saved.

[0068] In one possible implementation, the training set can be used to train a preset infrared precipitation prediction model to obtain a trained infrared precipitation prediction model. Specifically, the process may involve inputting multi-channel brightness temperature features of the sample region into the trained infrared precipitation prediction model, executing the model to obtain precipitation prediction values, calculating the error between the predicted precipitation values ​​and the corresponding true precipitation features, calculating the loss function, and updating the model parameters using an optimizer (such as SGD or Adam) to minimize the loss, until the preset number of training iterations is reached or the loss function converges. In this embodiment, the loss function of the infrared precipitation prediction model can be a weighted mean square error loss function. This weighted mean square error loss function dynamically adjusts the error weights based on the spatial heterogeneity of precipitation, thereby increasing the model's sensitivity to specific regions, such as high-precipitation areas, during training and improving prediction accuracy in these areas.

[0069] It's important to note that during model training, a validation set can be used to evaluate model performance. This validation set also includes multiple samples, and the samples in the validation set differ from those in the training set. Typically, the ratio of the number of samples in the validation set to those in the training set is 2:8 or 3:7. The validation set can be used to evaluate model performance after each training round or after a predetermined number of training rounds. Specifically, the multi-channel brightness temperature features of the sample regions in the validation set can be used as input data to the trained infrared precipitation prediction model for precipitation estimation inversion. The trained infrared precipitation prediction model is then executed to obtain predicted precipitation values. The corresponding ground truth values ​​of the precipitation features are used as the validation source to verify the predicted precipitation values, thus testing the performance of the infrared precipitation prediction model. Hyperparameters such as the learning rate and regularization parameters are adjusted based on the performance on the validation set. If the model's performance on the validation set does not improve, training can be stopped early.

[0070] This implementation enhances the sensitivity to key precipitation features by introducing attention mechanisms into the encoder and decoder, reducing the influence of weak precipitation intensity or background areas. A dilated spatial pyramid pooling module is embedded in the intermediate layer to capture both small-scale strong precipitation features and large-scale weak precipitation areas. Furthermore, dilated convolution expands the receptive field without loss of resolution, allowing the model to consider broader contextual information. The adaptive weighted mean squared error loss mechanism dynamically adjusts error weights based on the spatial heterogeneity of precipitation and imposes physical prior constraints on the loss function to prevent underestimation of extreme high-value areas. Thus, the model's feature extraction and modeling capabilities for precipitation targets of different scales and intensities are significantly improved, thereby enhancing the accuracy and precision of the model's predictions.

[0071] In one possible implementation, the weighted mean square error loss function weights the error according to the magnitude of the true value of the precipitation characteristics.

[0072] In this embodiment, the expression for the weighted mean square error loss function is:

[0073]

[0074] Where i = 1, 2, ..., n, and n is the number of samples in the training set; It is the predicted value of the i-th sample; x i w is the true value of the precipitation feature for the i-th sample; i It is the weight corresponding to the i-th sample, and w i ≥0.

[0075] In this embodiment, the precipitation characteristic true value x of the i-th sample can be used as a basis. i Determine the weight w corresponding to the i-th sample. i For example, regions with true precipitation features greater than a predetermined threshold (e.g., 30) are assigned higher weights (e.g., 2.0), while regions with true precipitation features less than or equal to the predetermined threshold are assigned lower weights (e.g., 1.0). Through this weighting strategy, the loss function can increase the model's sensitivity to high-value regions (i.e., regions with abundant precipitation) during training, thereby improving prediction accuracy in these areas. This is particularly suitable for applications requiring optimization of large-scale, high-value predictions, such as precipitation estimation and environmental monitoring.

[0076] In one possible implementation, the attention mechanism module includes a convolutional block attention module.

[0077] In this implementation, the CBAM (Convolutional Block Attention Module) includes two modules: Channel Attention and Spatial Attention. It can enhance the representational power of the feature map. By embedding CBAM in the encoder and decoder, adaptive feature weight allocation can be achieved, thereby enhancing the sensitivity to key precipitation regions.

[0078] The attention mechanism module in this embodiment is CBAM, which can combine channel attention and spatial attention. This dual mechanism allows the model to perform feature selection on two key dimensions, thereby capturing important information of the image more precisely.

[0079] In one possible implementation, obtaining the training set includes:

[0080] Acquire raw AGRI multi-channel brightness temperature data and IMERG Final half-hour precipitation data for the same area during the same time period;

[0081] Data within the acceptable range are selected from the multi-channel brightness temperature raw data and the IMERG Final half-hour precipitation raw data to obtain multi-channel brightness temperature data and IMERG Final half-hour precipitation data;

[0082] The multi-channel brightness temperature data and the IMERG Final half-hour precipitation data were spatiotemporally matched.

[0083] Extract multi-channel brightness temperature feature data from the matched multi-channel brightness temperature data and precipitation feature data from the IMERG Final half-hour precipitation data;

[0084] The multi-channel brightness temperature feature data and the precipitation feature data are normalized to obtain the normalized multi-channel brightness temperature feature and the true value of the precipitation feature.

[0085] In this implementation, the raw brightness temperature data from the AGRI (Advanced Geostationary Radiation Imager) multichannel system can be obtained from FY-4B (Fengyun-4B satellite), and the raw half-hour precipitation data from the IMERG (Integrated Multi-satellite Retrievals for GPM) Final product can be obtained from the meteorological data sharing platform.

[0086] In this implementation, the raw AGRI multi-channel brightness temperature data and the raw IMERGFinal half-hour precipitation data for the same time period can be acquired first. Then, the raw FY-4B AGRI multi-channel brightness temperature data and corresponding latitude and longitude data are read, as well as the raw IMERG Final half-hour precipitation data and corresponding latitude and longitude data are read. Based on the latitude and longitude range of the preset target area (including the longitude range [minimum longitude value, maximum longitude value], and the latitude range [minimum latitude value - maximum latitude value]), the corresponding raw AGRI multi-channel brightness temperature data and raw IMERG Final half-hour precipitation data for the same target area are extracted. In this way, the raw AGRI multi-channel brightness temperature data and raw IMERG Final half-hour precipitation data for the same area at the same time period are acquired.

[0087] In this embodiment, after obtaining the raw data, data quality control is required. Valid and high-quality data are selected from the multi-channel brightness temperature raw data. For example, data within the normal range of multi-channel brightness temperature raw data are selected, and abnormal data are filtered out. High-quality data are selected from the IMERG Final half-hour precipitation raw data. For example, data with a precipitation quality of 0.6 or higher from the IMERG Final half-hour precipitation raw data are selected.

[0088] In this embodiment, after obtaining high-quality multi-channel brightness temperature data and IMERG Final half-hour precipitation data, it is necessary to perform precise temporal and spatial matching of the multi-channel brightness temperature data and IMERG Final half-hour precipitation data to obtain multi-channel brightness temperature data and IMERG Final half-hour precipitation data that are matched with it in time and space.

[0089] In this embodiment, brightness temperature is the temperature (unit: K) when the energy radiated by an object in a specific wavelength band is equivalent to blackbody radiation. It can reflect the radiation characteristics of the atmosphere, clouds, or the Earth's surface. The scattering and emission of cloud particles during precipitation will significantly change the brightness temperature value. Therefore, brightness temperature data that is beneficial to learning precipitation targets can be extracted from the matched multi-channel brightness temperature data, such as channel brightness temperature data that can better reflect precipitation conditions. This could be the brightness temperature data of a single channel or the linear recombination value of the brightness temperatures of two or more channels, etc. Precipitation characteristic data such as precipitation rate can be extracted from the matched IMERG Final half-hour precipitation data.

[0090] In this embodiment, the multi-channel brightness temperature feature data and the precipitation feature data can be normalized to obtain normalized true values ​​for the multi-channel brightness temperature features and the precipitation features. For example, the minimum-maximum normalization method can be used to ensure that the extracted multi-channel brightness temperature feature data and precipitation feature data are within the range of 0 to 1 when input to the model, thus guaranteeing that the range of each input feature value is within the acceptable range of the model. The specific normalization formula is explained below:

[0091] X norm =(XX) min ) / (X max -X min );

[0092] Where X is the feature data before normalization, X min It is the minimum value in the feature data, X max It is the maximum value in the feature data, X norm It is the normalized feature array.

[0093] This implementation method can perform quality optimization, spatiotemporal consistency, feature extraction, and feature normalization on the original data to obtain high-quality training data that meets the model requirements, so as to train an accurate infrared precipitation prediction model based on this data.

[0094] In this embodiment, the spatiotemporal matching of the multi-channel brightness temperature data and the IMERG Final half-hour precipitation data includes...

[0095] Based on the start observation time of the multi-channel brightness temperature data, time matching is performed with the corresponding IMERG Final half-hour precipitation data.

[0096] Based on the nominal image points of the multi-channel brightness temperature data, spatial matching was performed with the IMERG Final half-hour precipitation data using the nearest neighbor method.

[0097] In this embodiment, for time matching, the start observation time of the multi-channel brightness temperature data can be used as a benchmark to match the corresponding IMERG Final half-hour precipitation data. For example, if the start observation time of the multi-channel brightness temperature data is 8:00 AM, then the IMERG Final half-hour precipitation data containing the time of 8:00 AM will be matched with it.

[0098] In this embodiment, nominal image points are used to correspond each pixel in the remote sensing image to latitude and longitude on Earth. This means projecting each pixel in the remote sensing image onto a map projection coordinate system to obtain the latitude and longitude coordinates of each pixel in the remote sensing image on the map projection coordinate system, which are also the nominal image point coordinates. For spatial matching, the nominal image points of the multi-channel brightness temperature data are the nominal image points projected onto the map projection coordinate system from the pixels in the multi-channel brightness temperature data. Similarly, the nominal image points of the IMERG Final half-hour precipitation data are the nominal image points projected onto the map projection coordinate system from the pixels in the IMERG Final half-hour precipitation data. Using the nominal image points of the multi-channel brightness temperature data as a reference, the nearest neighbor method can be used to find the nominal image point in the IMERG Final half-hour precipitation data that is closest to the nominal image point in the multi-channel brightness temperature data. Thus, the pixels corresponding to the two closest nominal image points in the two datasets are matched as a pair of pixels. By matching the position of these pixels, the IMERG Final half-hour precipitation data and the multi-channel brightness temperature data are accurately spatially matched.

[0099] This implementation method uses the start observation time of the multi-channel brightness temperature data as a reference to perform time matching with the corresponding IMERG Final half-hour precipitation data; and uses the nominal image point of the multi-channel brightness temperature data as a reference to perform spatial matching with the IMERG Final half-hour precipitation data using the nearest neighbor method. This results in more accurate spatiotemporal consistency matching, thereby obtaining more matching multi-channel brightness temperature features and their corresponding precipitation feature true values, and training an accurate infrared precipitation prediction model based on this.

[0100] In one possible implementation, the extraction of multi-channel brightness temperature features from the matched multi-channel brightness temperature data includes:

[0101] Extracting BT from the matched multi-channel brightness temperature data 3.75 BT 10.80 BT 12.00 ;

[0102] Calculate the brightness temperature difference ΔBT using the following formula. 10.80-6.25 :

[0103] ΔBT 10.80-6.25 =BT 10.80 -BT 6.25 ;

[0104] Calculate the brightness temperature linear recombination value BT using the following formula. multi-channel :

[0105] BT multi-channel =BT 12.00 +BT 8.55 -2*BT 10.80 ;

[0106] Obtain the BT 3.75 BT 10.80 BT 12.00 ΔBT 10.80-6.25 BT multi-channel Multi-channel brightness temperature characteristics;

[0107] Among them, BT 3.75 BT 8.55 BT 10.80 BT 12.00 These are the brightness temperatures of channels with center wavelengths of 3.75 μm, 8.55 μm, 10.80 μm, and 12.00 μm, respectively.

[0108] In this embodiment, the multi-channel brightness temperature features extracted from the matched multi-channel brightness temperature data have the following five values: BT 3.75 BT 10.80 BT 12.00 ΔBT 10.80-6.25 BTmulti-channel Among them, zero temperature difference ΔBT 10.80-6.25 And brightness temperature linear recombination value BT multi-channel Linear recombination values ​​of brightness temperature belonging to two or more channels.

[0109] Here, BT 3.75 It can identify low-albedo targets and surface types, and is used to observe cloud characteristics, facilitating the identification of cloud types and distribution features. BT 10.80 and BT 12.00 Both are highly sensitive to cloud top temperature, ΔBT 10.80-6.25 It can be used to estimate the height of tropospheric cloud tops, BT multi-channel Cloud top phase is highly sensitive. Factors such as cloud type, cloud top temperature, cloud phase, and tropospheric cloud top height are closely related to the likelihood, intensity, and distribution of precipitation. Different cloud types represent different cloud microphysical processes and thermodynamic states, directly affecting the occurrence and development of precipitation. Cloud top temperature indicates the presence of sufficient ice crystals or supercooled water droplets, crucial for precipitation formation. Different phases correspond to different precipitation mechanisms. Cloud top height reflects the vertical development of clouds, and tropospheric cloud top height more accurately identifies the intensity and extent of heavy convective precipitation. Therefore, BT... 3.75 BT 10.80 BT 12.00 ΔBT 10.80-6.25 BT multi-channe These five brightness temperature data are all channels that can reflect precipitation conditions well. By extracting these five brightness temperature data as multi-channel brightness temperature features, we can learn feature data that is beneficial to precipitation conditions, thereby enabling the learned model to make accurate precipitation predictions.

[0110] This disclosure also provides an infrared precipitation prediction method. Figure 3 A flowchart illustrating an infrared precipitation prediction method provided by an embodiment of this disclosure is shown. Figure 3 As shown, the infrared precipitation prediction method includes the following steps:

[0111] In step S301, the current multi-channel brightness temperature raw data of the region to be predicted is obtained;

[0112] In step S302, based on the current multi-channel brightness temperature raw data, the current multi-channel brightness temperature characteristic data is obtained;

[0113] In step S303, the current multi-channel brightness temperature feature data is input into the pre-trained infrared precipitation prediction model to obtain the precipitation prediction value output by the infrared precipitation prediction model.

[0114] In one possible implementation, the infrared precipitation prediction method can be applied to electronic devices capable of performing infrared precipitation prediction.

[0115] In one possible implementation, the current multi-channel brightness temperature raw data of the region to be predicted can be obtained from FY-4B AGRI. The raw multi-channel brightness temperature data can be subjected to outlier filtering (selecting data within the normal range), feature data extraction (obtaining the relevant channel brightness temperature values), and normalization processing (such as maximum and minimum normalization processing) to obtain the current multi-channel brightness temperature feature data.

[0116] In one possible implementation, the current multi-channel brightness temperature characteristic data can be input into a pre-trained infrared precipitation prediction model. By executing the trained infrared precipitation prediction model, the precipitation prediction value output by the infrared precipitation prediction model can be obtained.

[0117] In one possible implementation, IMERG Final half-hour precipitation data is used as the verification source. The heavy precipitation event of Typhoon Doksuri on July 27, 2023, is selected to evaluate the quality of the prediction results of the infrared precipitation prediction model disclosed herein. The results are then compared with those of commonly used convolutional neural network (CNN) models and long short-term memory network (ConvLSTM) models. Figure 4 The following diagrams are shown: the spatial distribution map of infrared precipitation prediction provided in this embodiment, the spatial distribution map of precipitation products from the reference source, the distribution map of prediction results from the convolutional neural network model, and the distribution map of prediction results from the long short-term memory network model. Figure 4 The four distribution maps shown use different colors of the color card to represent different precipitation rates. It can be clearly seen that the infrared precipitation prediction spatial distribution map of the infrared precipitation prediction model provided in this embodiment is most similar to the spatial distribution map of the precipitation product of the reference source, namely IMERG Final, and is more accurate than the prediction results of commonly used CNN models and ConvLSTM models.

[0118] Table 1 below presents the accuracy evaluation results of the infrared precipitation prediction model, CNN model, and ConvLSTM model disclosed in this publication compared with the IMERG Final precipitation product:

[0119]

[0120] Table 1

[0121] As can be seen from Table 1, the infrared precipitation prediction model provided in this embodiment has the highest accuracy and hit rate. Although the false alarm rate is higher than that of the ConvLSTM model, the difference is small. Moreover, based on the comparison of the relative error of short-term heavy rainfall level samples, the average deviation of short-term heavy rainfall level samples, and the correlation coefficient of short-term heavy rainfall level samples, it can be seen that for the prediction of heavy precipitation areas, the prediction results of commonly used CNN and ConvLSTM models are significantly different from the IMERG Final precipitation product, while the accuracy of the infrared precipitation prediction model provided in this embodiment is relatively better.

[0122] This disclosure also provides an infrared precipitation prediction model training device. Figure 5 This diagram illustrates the structure of an infrared precipitation prediction model training device provided in an embodiment of this disclosure. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 5 As shown, the infrared precipitation prediction model training device includes:

[0123] The sample acquisition module 501 is configured to acquire a training set, wherein the training samples in the training set include ground truth values ​​of multi-channel brightness temperature features and precipitation features.

[0124] The model acquisition module 502 is configured to acquire a preset infrared precipitation prediction model. The infrared precipitation prediction model includes an input layer, an encoder, an intermediate layer, a decoder, and an output layer. The encoder adds an attention mechanism module before each downsampling. The intermediate layer includes a hollow spatial pyramid pooling module. The decoder adds an attention mechanism module after each upsampling. The loss function of the infrared precipitation prediction model includes a weighted mean square error loss function. The weighted mean square error loss function dynamically adjusts the error weights according to the spatial heterogeneity of precipitation.

[0125] The model training module 503 is configured to use the training set to train a preset infrared precipitation prediction model to obtain a trained infrared precipitation prediction model.

[0126] In one possible implementation, the weighted mean square error loss function weights the error according to the magnitude of the true value of the precipitation characteristics.

[0127] In one possible implementation, the attention mechanism module includes a convolutional block attention module.

[0128] In one possible implementation, the sample acquisition module 501 is configured to:

[0129] Acquire raw brightness temperature data from the multi-channel scanning imaging radiometer AGRI and raw half-hour precipitation data from the IMERG Final, the final product of global precipitation measurement and multi-satellite precipitation estimation, for the same area and time period.

[0130] Data within the acceptable range are selected from the multi-channel brightness temperature raw data and the IMERG Final half-hour precipitation raw data to obtain multi-channel brightness temperature data and IMERG Final half-hour precipitation data;

[0131] The multi-channel brightness temperature data and the IMERG Final half-hour precipitation data were spatiotemporally matched.

[0132] Extract multi-channel brightness temperature feature data from the matched multi-channel brightness temperature data and precipitation feature data from the IMERG Final half-hour precipitation data;

[0133] The multi-channel brightness temperature feature data and the precipitation feature data are normalized to obtain the normalized multi-channel brightness temperature feature and the true value of the precipitation feature.

[0134] In one possible implementation, the part of the sample acquisition module 501 that performs spatiotemporal matching of the multi-channel brightness temperature data and the IMERG Final half-hour precipitation data is configured as follows:

[0135] Based on the start observation time of the multi-channel brightness temperature data, time matching is performed with the corresponding IMERG Final half-hour precipitation data.

[0136] Based on the nominal image points of the multi-channel brightness temperature data, spatial matching is performed with the nominal image points of the IMERG Final half-hour precipitation data using the nearest neighbor method.

[0137] In one possible implementation, the portion of the multi-channel brightness temperature features extracted from the matched multi-channel brightness temperature data in the sample acquisition module 501 is configured as follows:

[0138] Extracting BT from the matched multi-channel brightness temperature data 3.75 BT 10.80 BT 12.00 ;

[0139] Calculate the brightness temperature difference ΔBT using the following formula. 10.80-6.25 :

[0140] ΔBT 10.80-6.25 =BT 10.80 -BT 6.25 ;

[0141] Calculate the brightness temperature linear recombination value BT using the following formula. multi-channel :

[0142] BT multi-channel =BT 12.00 +BT8.55 -2*BT 10.80 ;

[0143] Obtain the BT 3.75 BT 10.80 BT 12.00 ΔBT 10.80-6.25 BT multi-channel Multi-channel brightness temperature characteristics;

[0144] Among them, BT 3.75 BT 8.55 BT 10.80 BT 12.00 These are the brightness temperatures of channels with center wavelengths of 3.75 μm, 8.55 μm, 10.80 μm, and 12.00 μm, respectively.

[0145] This disclosure also provides an infrared precipitation prediction device. Figure 6 This diagram illustrates the structure of an infrared precipitation prediction device provided in an embodiment of the present disclosure. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 6 As shown, the infrared precipitation prediction device includes:

[0146] The data acquisition module 601 is configured to acquire the current multi-channel brightness temperature raw data of the area to be predicted;

[0147] Feature extraction module 602 is configured to extract current multi-channel brightness temperature feature data from the current multi-channel brightness temperature raw data;

[0148] The prediction module 603 is configured to input the current multi-channel brightness temperature feature data into a pre-trained infrared precipitation prediction model to obtain the precipitation prediction value output by the infrared precipitation prediction model.

[0149] The pre-trained infrared precipitation prediction model is obtained by training the infrared precipitation prediction model using the above-mentioned infrared precipitation prediction model training method.

[0150] The technical terms and features mentioned in this device embodiment are the same as or similar to those mentioned in the above method embodiment. For explanations and descriptions of the technical terms and features involved in this device, please refer to the explanations of the above method embodiment. They will not be repeated here.

[0151] This disclosure also discloses an electronic device. Figure 7 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0152] like Figure 7As shown, the electronic device 700 includes a memory 701 and a processor 702, wherein the memory 701 is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor 702 to implement the method according to an embodiment of the present disclosure.

[0153] Figure 8 A schematic diagram of the structure of a computer system suitable for implementing the methods of the embodiments of this disclosure is shown.

[0154] like Figure 8 As shown, the computer system 800 includes a processing unit 801, which can execute various processes described in the above embodiments according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the computer system 800. The processing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0155] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed. The processing unit 801 can be implemented as a CPU, GPU, TPU, FPGA, NPU, etc.

[0156] In particular, according to embodiments of this disclosure, the methods described above can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising computer instructions that, when executed by a processor, implement the steps of the methods described above. In such embodiments, the computer program product can be downloaded and installed from a network via communication section 809, and / or installed from removable media 811.

[0157] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0158] The units or modules described in the embodiments of this disclosure can be implemented in software or programmable hardware. The described units or modules can also be located in a processor, and the names of these units or modules do not necessarily constitute a limitation on the unit or module itself.

[0159] In another aspect, this disclosure also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the electronic device or computer system described above; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores one or more programs, which are used by one or more processors to perform the methods described in this disclosure.

[0160] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A method for training an infrared precipitation prediction model, characterized in that, include: Obtain a training set, which includes multiple samples, each of which includes the multi-channel brightness temperature features of a sample region and the corresponding precipitation feature ground values; A preset infrared precipitation prediction model is obtained. The infrared precipitation prediction model includes an input layer, an encoder, an intermediate layer, a decoder, and an output layer. The encoder adds an attention mechanism module before each downsampling. The intermediate layer includes a hollow spatial pyramid pooling module. The decoder adds an attention mechanism module after each upsampling. The loss function of the infrared precipitation prediction model includes a weighted mean square error loss function. The weighted mean square error loss function dynamically adjusts the error weights according to the spatial heterogeneity of precipitation. The training set is used to train the preset infrared precipitation prediction model to obtain the trained infrared precipitation prediction model.

2. The method according to claim 1, characterized in that, The weighted mean square error loss function weights the error according to the magnitude of the true value of precipitation characteristics.

3. The method according to claim 1, characterized in that, The attention mechanism module includes a convolutional block attention module.

4. The method according to claim 1, characterized in that, The acquisition of the training set includes: Acquire raw brightness temperature data from the multi-channel scanning imaging radiometer AGRI and raw half-hour precipitation data from the IMERG Final, the final product of global precipitation measurement and multi-satellite precipitation fusion; Data within the acceptable range are selected from the multi-channel brightness temperature raw data and the IMERG Final half-hour precipitation raw data to obtain multi-channel brightness temperature data and IMERG Final half-hour precipitation data; The multi-channel brightness temperature data and the IMERG Final half-hour precipitation data were spatiotemporally matched. Extract multi-channel brightness temperature feature data from the matched multi-channel brightness temperature data and precipitation feature data from the IMERG Final half-hour precipitation data; The multi-channel brightness temperature feature data and the precipitation feature data are normalized to obtain the normalized multi-channel brightness temperature feature and the true value of the precipitation feature.

5. The method according to claim 4, characterized in that, The process of performing spatiotemporal matching between the multi-channel brightness temperature data and the IMERGFinal half-hour precipitation data includes: Based on the start observation time of the multi-channel brightness temperature data, time matching is performed with the corresponding IMERG Final half-hour precipitation data. Based on the nominal image points of the multi-channel brightness temperature data, spatial matching is performed with the nominal image points of the IMERG Final half-hour precipitation data using the nearest neighbor method.

6. The method according to claim 4, characterized in that, The extracted multi-channel brightness temperature features from the matched multi-channel brightness temperature data include: Extracting BT from the matched multi-channel brightness temperature data 3.75 BT 10.80 BT 12.00 ; Calculate the brightness temperature difference ΔBT using the following formula. 10.80-6.25 : ΔBT 10.80-6.25 =BT 10.80 -BT 6.25 ; Calculate the brightness temperature linear recombination value BT using the following formula. multi-channel : BT multi-channel =BT 12.00 +BT 8.55 -2*BT 10.80 ; Obtain the BT 3.75 BT 10.80 BT 12.00 ΔBT 10.80-6.25 BT multi-channel It features multi-channel brightness temperature characteristics; Among them, BT 3.75 BT 8.55 BT 10.80 BT 12.00 These are the brightness temperatures of channels with center wavelengths of 3.75 μm, 8.55 μm, 10.80 μm, and 12.00 μm, respectively.

7. An infrared precipitation prediction method, characterized in that, include: Obtain the current multi-channel brightness temperature raw data of the region to be predicted; Extract the current multi-channel brightness temperature feature data from the current multi-channel brightness temperature raw data; The current multi-channel brightness temperature feature data is input into a pre-trained infrared precipitation prediction model to obtain the precipitation prediction value output by the infrared precipitation prediction model. The pre-trained infrared precipitation prediction model is obtained by training the method described in any one of claims 1-6.

8. An infrared precipitation prediction model training device, characterized in that, include: The sample acquisition module is configured to acquire a training set, wherein the training samples in the training set include ground truth values ​​of multi-channel brightness temperature features and precipitation features; The model acquisition module is configured to acquire a preset infrared precipitation prediction model. The infrared precipitation prediction model includes an input layer, an encoder, an intermediate layer, a decoder, and an output layer. The encoder adds an attention mechanism module before each downsampling. The intermediate layer includes a hollow spatial pyramid pooling module. The decoder adds an attention mechanism module after each upsampling. The loss function of the infrared precipitation prediction model includes a weighted mean square error loss function. The weighted mean square error loss function dynamically adjusts the error weights according to the spatial heterogeneity of precipitation. The model training module is configured to use the training set to train a preset infrared precipitation prediction model to obtain a trained infrared precipitation prediction model.

9. An infrared precipitation prediction device, characterized in that, include: The data acquisition module is configured to acquire the current multi-channel brightness temperature raw data of the area to be predicted; The feature extraction module is configured to extract current multi-channel brightness temperature feature data from the current multi-channel brightness temperature raw data; The prediction module is configured to input the current multi-channel brightness temperature feature data into a pre-trained infrared precipitation prediction model to obtain the precipitation prediction value output by the infrared precipitation prediction model. The pre-trained infrared precipitation prediction model is obtained by training the method described in any one of claims 1-6.

10. An electronic device, characterized in that, The method includes a memory and a processor, the memory being used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method of any one of claims 1 to 7.

Images