An extreme rainfall prediction method and device based on deep learning and electronic equipment

Abstract

This invention provides a method, device, and electronic device for extreme rainfall prediction based on deep learning. The method includes: collecting multi-source data and preprocessing the multi-source data to construct a multi-source heterogeneous fusion dataset; inputting the multi-source heterogeneous fusion dataset into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process; generating synthetic extreme rainfall samples through a generative adversarial network, mixing the synthetic extreme rainfall samples with real samples for training, and strengthening the learning of extreme rainfall features through a weighted loss function; introducing channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process. This invention achieves accurate prediction and early warning of extreme rainfall events, solving the problem of poor prediction accuracy in existing related technologies.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method, device, and electronic device for predicting extreme rainfall based on deep learning. Background Technology

[0002] Current extreme rainfall prediction mainly relies on numerical weather prediction models and traditional statistical models, but there are many technical bottlenecks: while numerical weather prediction models can simulate atmospheric physical processes, they have high computational complexity, forecast time lag, and insufficient ability to capture small- and medium-scale extreme rainfall events; traditional statistical models are unable to characterize the nonlinear spatiotemporal evolution of rainfall processes, resulting in limited prediction accuracy. In recent years, deep learning technology has been applied in the field of rainfall prediction, but existing methods still have significant shortcomings: First, multi-source data fusion is insufficient, failing to effectively integrate complementary information from satellite, radar, and ground observation data, and the multi-channel characteristics of satellite data and the fine structural characteristics of radar data have not been deeply integrated; second, models do not learn enough about extreme rainfall characteristics, and due to the scarcity of extreme rainfall samples, existing models tend to learn from ordinary rainfall samples, leading to a high false negative rate in extreme rainfall predictions; third, there is a lack of targeted spatiotemporal feature optimization, failing to fully capture the rapid evolution process and spatial clustering characteristics of extreme rainfall; fourth, the prediction threshold is fixed, failing to consider the impact of climate and topographic differences in different regions on the criteria for determining extreme rainfall, resulting in inconsistent warning accuracy.

[0003] Furthermore, existing deep learning-based rainfall prediction methods fail to fully utilize the advantages of wide-area observation of satellite data and the high-resolution characteristics of radar data. The feature extraction process does not highlight key features of extreme rainfall, and the model update mechanism is rigid, making it difficult to adapt to changes in rainfall patterns caused by climate change. Therefore, there is an urgent need to develop an extreme rainfall prediction system that can integrate multi-source data, enhance extreme feature learning, and dynamically adapt to regional differences, thereby improving the accuracy and timeliness of extreme rainfall prediction.

[0004] There is currently no effective solution to the problem of poor accuracy in extreme rainfall prediction using existing related technologies. Summary of the Invention

[0005] This invention provides a deep learning-based method, device, and electronic device for predicting extreme rainfall, which addresses the shortcomings of existing related technologies in terms of poor accuracy in predicting extreme rainfall, and enables accurate prediction and early warning of extreme rainfall events.

[0006] In a first aspect, the present invention provides a deep learning-based method for predicting extreme rainfall, comprising:

[0007] Collect multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0008] The multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0009] Synthetic extreme rainfall samples are generated by generative adversarial networks, and these synthetic extreme rainfall samples are mixed with real samples for training. The learning of extreme rainfall features is enhanced by a weighted loss function.

[0010] By introducing channel attention and spatial attention mechanisms, feature weights are dynamically adjusted to capture the temporal dependencies of rainfall processes;

[0011] Historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the threshold for judging extreme rainfall is optimized to generate extreme rainfall warning results.

[0012] According to a deep learning-based extreme rainfall prediction method provided by the present invention, the multi-source data is preprocessed to construct a multi-source heterogeneous fusion dataset, including:

[0013] Radiometric calibration was performed on the multi-channel satellite observation data, and spatial unification was achieved through equal latitude and longitude projection conversion;

[0014] The radar echo data is subjected to clutter suppression, attenuation correction, and resolution adjustment.

[0015] Outlier removal and interpolation were performed on the ground station observation data;

[0016] Based on the timestamps of satellite multi-channel observation data, linear interpolation is used to synchronize the radar echo data with the ground station observation data, and the spatial resolution of all the multi-source data is unified through linear interpolation.

[0017] The satellite multi-channel observation data, the radar echo data, the ground station observation data, and auxiliary features are stacked according to the channel dimension to obtain a multi-source heterogeneous fusion dataset.

[0018] According to a deep learning-based extreme rainfall prediction method provided by the present invention, the multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process, including:

[0019] At the encoding end, the feature differences between small-scale heavy rainfall and large-scale uniform rainfall are captured through multiple downsampling stages to obtain feature maps at different stages;

[0020] At the decoding end, the feature map obtained at the encoding end is fused with the feature map of the corresponding stage at the decoding end through upsampling to obtain the spatiotemporal evolution features of the process.

[0021] According to the deep learning-based extreme rainfall prediction method provided by the present invention, each downsampling stage of the encoding end consists of a three-dimensional convolutional layer, a batch normalization layer, a GELU activation function, and a residual multi-scale module;

[0022] The three-dimensional convolutional layer is used to achieve feature compression in the spatiotemporal dimension;

[0023] The residual multi-scale module is used to stitch features at different scales, while embedding dilated convolutions to expand the receptive field.

[0024] According to a deep learning-based extreme rainfall prediction method provided by the present invention, at the decoding end, the feature map obtained at the encoding end is fused with the feature map of the corresponding stage at the decoding end through upsampling to obtain the spatiotemporal evolution features of the process, including:

[0025] Feature map size recovery is achieved through transposed 3D convolution.

[0026] By employing skip connections, feature maps from different stages of the encoding end are fused with feature maps from the corresponding stages of the decoding end to obtain spatiotemporal evolution features.

[0027] According to the present invention, a deep learning-based method for predicting extreme rainfall is provided, wherein the generative adversarial network includes a generator and a discriminator;

[0028] The generator uses an improved U-Net structure to generate synthetic extreme rainfall samples that conform to the spatiotemporal distribution of real rainfall.

[0029] The discriminator optimizes the generation quality by distinguishing the synthetic extreme rainfall samples from real samples.

[0030] According to the present invention, a deep learning-based extreme rainfall prediction method is provided, which introduces channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependency of rainfall processes, including:

[0031] Based on the squeeze-excitation mechanism, global average pooling and global max pooling are performed on the channel dimension of the feature map of the spatiotemporal evolution features. Channel weights are learned through a two-layer fully connected network to dynamically strengthen the channel features that are highly correlated with satellite water vapor channels, radar strong echo channels and extreme rainfall.

[0032] Based on the self-attention mechanism, a spatiotemporal attention matrix is ​​constructed to calculate the correlation weight between each spatiotemporal location and other locations, thereby enhancing the spatial clustering characteristics and temporal evolution patterns of the core area of ​​extreme rainfall.

[0033] According to the deep learning-based extreme rainfall prediction method provided by the present invention, historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the extreme rainfall judgment threshold is optimized to generate extreme rainfall warning results, including:

[0034] The distribution of rainfall intensity in real-time observation data is statistically analyzed using a sliding window.

[0035] Input the geographical and climatic feature parameters of the current region into the gradient boosting tree model to obtain the initial threshold;

[0036] The initial threshold is dynamically adjusted based on the real-time rainfall intensity distribution to obtain the extreme rainfall determination threshold for the current area;

[0037] Based on the extreme rainfall determination threshold, it is determined whether extreme rainfall will occur in the current area, and an extreme rainfall warning result is generated.

[0038] Secondly, the present invention also provides a deep learning-based extreme rainfall prediction device, comprising:

[0039] The data acquisition module is used to acquire multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0040] The feature extraction module is used to input the multi-source heterogeneous fusion dataset into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0041] The reinforcement learning module is used to generate synthetic extreme rainfall samples through a generative adversarial network, mix the synthetic extreme rainfall samples with real samples for training, and strengthen the learning of extreme rainfall features through a weighted loss function.

[0042] The weight optimization module is used to introduce channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process;

[0043] The threshold calibration module is used to acquire historical rainfall data and real-time observation data, build an adaptive threshold adjustment model, optimize the extreme rainfall judgment threshold, and generate extreme rainfall warning results.

[0044] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the deep learning-based extreme rainfall prediction method as described in the first aspect above.

[0045] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep learning-based extreme rainfall prediction method as described in the first aspect above.

[0046] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based extreme rainfall prediction method as described in the first aspect above.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The deep learning-based extreme rainfall prediction method provided by this invention addresses the core pain points of existing technologies, such as insufficient multi-source data fusion, inadequate learning of extreme features, lack of spatiotemporal feature capture, and poor regional adaptability. Through multi-module collaborative design, it achieves accurate prediction and early warning of extreme rainfall events, solving the problem of poor prediction accuracy of extreme rainfall in existing related technologies. It is applicable to scenarios such as meteorological disaster early warning, flood control scheduling, and transportation safety assurance. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the deep learning-based extreme rainfall prediction method provided by the present invention;

[0051] Figure 2 This is a flowchart illustrating data acquisition and preprocessing in an embodiment of the present invention;

[0052] Figure 3 This is a flowchart of feature extraction in an embodiment of the present invention;

[0053] Figure 4 This is a flowchart of reinforcement learning in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of spatiotemporal attention in an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of channel attention in an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram illustrating the feature weight optimization process in an embodiment of the present invention;

[0057] Figure 8 This is a schematic diagram of the threshold calibration process in an embodiment of the present invention;

[0058] Figure 9 This is a structural block diagram of the deep learning-based extreme rainfall prediction device provided by the present invention;

[0059] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0061] This invention provides a deep learning-based method for predicting extreme rainfall. Figure 1 This is a flowchart of the deep learning-based extreme rainfall prediction method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0062] Step S101: Collect multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0063] Step S102: Input the multi-source heterogeneous fusion dataset into the multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0064] Step S103: Generate synthetic extreme rainfall samples through a generative adversarial network, mix the synthetic extreme rainfall samples with real samples for training, and enhance the learning of extreme rainfall features through a weighted loss function;

[0065] Step S104: Introduce channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process;

[0066] Step S105: Obtain historical rainfall data and real-time observation data, construct an adaptive threshold adjustment model, optimize the extreme rainfall judgment threshold, and generate extreme rainfall warning results.

[0067] For example, firstly, data from the 14 channels of the FY-4A satellite AGRI sensor, S-band radar echo data, and ground rain gauge observations are collected and preprocessed to construct a multi-source heterogeneous fusion dataset. Then, the preprocessed multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process, obtaining feature representations from low to high dimensions. Next, a generative adversarial network (GAN) is used to generate synthetic extreme rainfall samples. These synthetic extreme rainfall samples are then mixed with real samples for training, and a weighted loss function is used to enhance the learning of extreme rainfall features. Then, channel attention and spatial attention mechanisms are introduced to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process. Finally, an adaptive threshold adjustment model is constructed, and the extreme rainfall judgment threshold is optimized to generate extreme rainfall warning results. Through the above process, addressing the core pain points of existing technologies such as insufficient multi-source data fusion, inadequate learning of extreme features, lack of spatiotemporal feature capture, and poor regional adaptability, this technology achieves accurate prediction and early warning of extreme rainfall events through multi-module collaborative design. This solves the problem of poor prediction accuracy of extreme rainfall in existing related technologies and is applicable to scenarios such as meteorological disaster early warning, flood control scheduling, and transportation safety assurance.

[0068] In some embodiments, step S101 involves preprocessing the multi-source data to construct a multi-source heterogeneous fusion dataset, including: radiometric calibration of satellite multi-channel observation data and spatial unification through isotropic projection conversion; clutter suppression, attenuation correction, and resolution adjustment of radar echo data; outlier removal and interpolation processing of ground station observation data; time synchronization of radar echo data and ground station observation data using linear interpolation based on the timestamps of satellite multi-channel observation data, and unification of the spatial resolution of all multi-source data through linear interpolation; and stacking satellite multi-channel observation data, radar echo data, ground station observation data, and auxiliary features according to channel dimensions to obtain a multi-source heterogeneous fusion dataset.

[0069] Figure 2 This is a flowchart of data acquisition and preprocessing in an embodiment of the present invention, such as... Figure 2 As shown, in this embodiment, preprocessing of multi-source data provides a high-quality dataset for subsequent model training and prediction. For example, during the acquisition phase, 14 channels of data from the FY-4A satellite AGRI sensor (visible light, infrared, and water vapor channels), S-band radar echo data, and rainfall observation data from ground automatic weather stations are acquired simultaneously. The acquisition frequency can be set to 15 minutes / time or 30 minutes / time depending on forecast requirements.

[0070] In the processing phase, the data processing workflow includes three steps: First, satellite multi-channel observation data is radiometrically calibrated to convert DN values ​​into reflectivity or brightness temperature. Geometric correction is performed using the row and column number-latitude / longitude conversion formula, and spatial uniformity is achieved using equal latitude / longitude projection conversion. Radar echo data undergoes clutter suppression, attenuation correction, and resolution adjustment. Ground station observation data undergoes outlier removal (based on the 3σ criterion) and interpolation processing. Second, spatiotemporal alignment is performed. Using the timestamps of the satellite multi-channel observation data as a reference, linear interpolation is used to synchronize the time of radar echo data and ground station observation data. Bilinear interpolation unifies all data to a spatial resolution of 0.05°×0.05°. Finally, feature fusion is performed. Satellite multi-channel observation data, radar echo data, ground station observation data, and auxiliary features (topographic elevation, underlying surface type) are stacked according to the channel dimension to construct an input tensor with dimensions of [time series length × number of channels × height × width]. The time series length can be dynamically adjusted according to the forecast lead time (4 frames for a 1-hour forecast, 12 frames for a 6-hour forecast).

[0071] Based on this, in step S102, the multi-source heterogeneous fusion dataset is input into the multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process, including: capturing the feature differences between small-scale heavy rainfall and large-scale uniform rainfall through multiple downsampling stages at the encoding end to obtain feature maps of different stages; and fusing the feature maps obtained at the encoding end with the feature maps of the corresponding stages at the decoding end through upsampling at the decoding end to obtain the spatiotemporal evolution features of the rainfall process.

[0072] Specifically, Figure 3 This is a flowchart of feature extraction in an embodiment of the present invention, such as... Figure 3 As shown, each downsampling stage at the encoding end consists of a 3D convolutional layer, a batch normalization layer, a GELU activation function, and a residual multi-scale module; the 3D convolutional layer is used to achieve feature compression in the spatiotemporal dimension; the residual multi-scale module is used to concatenate features at different scales, while embedding dilated convolutions to expand the receptive field.

[0073] At the decoding end, the feature map obtained at the encoding end is fused with the feature map of the corresponding stage at the decoding end through upsampling to obtain the spatiotemporal evolution features of the process, including: restoring the feature map size through transposed 3D convolution; and using skip connections to fuse the feature maps of different stages at the encoding end with the feature maps of the corresponding stage at the decoding end to obtain the spatiotemporal evolution features.

[0074] In this embodiment, a multi-scale feature extraction network serves as the core of feature extraction. Based on an encoder-decoder architecture, it integrates 3D convolution and residual multi-scale modules to achieve deep extraction of spatiotemporal features of the rainfall process. The input to the multi-scale feature extraction network is the multi-source fusion tensor corresponding to the multi-source fusion dataset obtained through data fusion preprocessing. Through the collaborative operation of the encoder and decoder, it extracts multi-scale rainfall features from coarse to fine.

[0075] For example, the encoding stage includes four downsampling stages, each consisting of a 3D convolutional layer, a batch normalization layer, a GELU activation function, and a residual multi-scale module. The 3D convolutional layer uses a 3×3×3 kernel with a stride of 2 to achieve spatiotemporal feature compression. The residual multi-scale module includes 3D convolutional branches at three scales: 1×1, 3×3, and 5×5. It fuses features at different scales through feature concatenation and embeds dilated convolutions (with dilation rates of 1, 2, and 4 respectively) to expand the receptive field, effectively capturing the feature differences between small-scale heavy rainfall and large-scale uniform rainfall. The decoding stage includes four upsampling stages. It restores the feature map size through transposed 3D convolutions and uses skip connections to fuse feature maps from different stages of the encoding stage with corresponding feature maps from the decoding stage, reducing information loss during downsampling. The intermediate layers of the network output multi-scale feature maps, providing a foundation for subsequent attention optimization.

[0076] In some embodiments, the generative adversarial network includes a generator and a discriminator; the generator uses an improved U-Net structure to generate synthetic extreme rainfall samples that conform to the spatiotemporal distribution of real rainfall; the discriminator optimizes the generation quality by distinguishing between synthetic extreme rainfall samples and real samples.

[0077] Figure 4 This is a flowchart of reinforcement learning in an embodiment of the present invention, such as... Figure 4 As shown, in this embodiment, to address the model prediction bias caused by the scarcity of extreme rainfall samples, sample augmentation and loss function optimization are used to enhance the model's learning ability for extreme rainfall characteristics. The specific process is implemented through two sub-units: sample generation and sample optimization.

[0078] The sample generation subunit employs a generative adversarial network (GAN) architecture. The generator, based on an improved U-Net structure, takes random noise and features from ordinary rainfall samples as input and outputs synthetic extreme rainfall samples. An attention mechanism is embedded in the generator to ensure the spatiotemporal plausibility of the synthetic samples. The discriminator uses a 3D convolutional neural network to optimize generation quality by distinguishing between real and synthetic extreme rainfall samples. During training, the Wasserstein distance loss function is used to improve the stability of sample generation. Synthetic extreme rainfall samples are mixed with real samples in a 1:1 ratio to construct an expanded dataset.

[0079] The sample optimization subunit employs a hard negative sample mining strategy, screening critical state samples from historical data within 3 hours prior to extreme rainfall events and similar rainfall samples prone to misclassification. The training weights of these samples are then increased using a weighted loss function. The loss function is designed as a combination of Mean Squared Error (MSE) loss and extreme rainfall weighted loss, with the loss weight for extreme rainfall areas set at 3-5 times that for ordinary areas. The formula is as follows:

[0080]

[0081] Where Loss represents the loss function. and For balance coefficient, , `predict_extreme` and `true_extreme` represent the predicted and actual values ​​of the extreme rainfall area, respectively. `WeightedMSE` represents the weighted mean square error for the extreme rainfall area, which means that when calculating the difference between the predicted and actual values, the extreme rainfall sample is given a higher computational weight.

[0082] In some embodiments, step S104 introduces channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process. This includes: based on the squeeze-excitation mechanism, performing global average pooling and global max pooling on the channel dimension of the feature map of the spatiotemporal evolution features; learning channel weights through a two-layer fully connected network to dynamically strengthen the channel features that are highly correlated with satellite water vapor channels, radar strong echo channels, and extreme rainfall; and based on the self-attention mechanism, constructing a spatiotemporal attention matrix, calculating the correlation weight between each spatiotemporal location and other locations to strengthen the spatial clustering characteristics and temporal evolution patterns of the core area of ​​extreme rainfall.

[0083] For example, such as Figure 5-7 As shown, Figure 5 This is a schematic diagram of spatiotemporal attention in an embodiment of the present invention. Figure 6 This is a schematic diagram of channel attention in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the feature weight optimization process in this embodiment of the invention. Feature weights are dynamically allocated through an attention mechanism to enhance the response of features related to extreme rainfall, suppress interference from irrelevant information, and improve the model's prediction accuracy. Channel attention and spatial-temporal attention are introduced, and a serial structure is used to process the feature maps output by the multi-scale feature extraction network sequentially.

[0084] Channel attention is based on the squeeze-and-excitation (SE) mechanism, performing global average pooling and global max pooling on the channel dimensions of the feature map. A two-layer fully connected network learns channel weights, dynamically strengthening channel features highly correlated with extreme rainfall, such as satellite water vapor channels and radar strong echo channels, while suppressing the influence of noisy channels. Spatial-temporal attention employs a self-attention mechanism, constructing a spatiotemporal attention matrix and calculating the association weights between each spatiotemporal location and other locations. It focuses on strengthening the spatial clustering characteristics and temporal evolution patterns of the core extreme rainfall region, while limiting computational complexity through a local attention window whose size can be adaptively adjusted according to the feature map size.

[0085] After attention optimization, the feature map shows a significant enhancement of features related to extreme rainfall, providing more distinctive feature inputs for subsequent predictions.

[0086] Figure 8 This is a schematic diagram of the threshold calibration process in an embodiment of the present invention, as shown below. Figure 8 As shown, based on the above embodiment, step S105 involves acquiring historical rainfall data and real-time observation data, constructing an adaptive threshold adjustment model, optimizing the extreme rainfall determination threshold, and generating an extreme rainfall warning result. This includes: statistically analyzing the rainfall intensity distribution of real-time observation data through a sliding window; inputting the geographical and climatic characteristic parameters of the current region into the gradient boosting tree model to obtain an initial threshold; dynamically adjusting the initial threshold according to the real-time rainfall intensity distribution to obtain the extreme rainfall determination threshold for the current region; and determining whether extreme rainfall will occur in the current region based on the extreme rainfall determination threshold to generate an extreme rainfall warning result.

[0087] In this embodiment, an adaptive threshold adjustment model is constructed to address the differences in extreme rainfall determination criteria across different regions, thereby achieving regionally adaptable extreme rainfall early warning. This embodiment is based on the Gradient Boosting Tree (GBT) algorithm, using historical extreme rainfall event data (1981-2020) as training samples. Input factors include regional longitude, latitude, altitude, historical extreme rainfall frequency, seasonal characteristics, and underlying surface type. The output is the extreme rainfall determination threshold for that region.

[0088] The threshold calibration process includes: first, statistically analyzing the rainfall intensity distribution of real-time observation data using a sliding window (30 days); second, inputting the geographical and climatic characteristic parameters of the current region into the gradient boosting tree model to obtain an initial threshold; and finally, dynamically adjusting the initial threshold based on the real-time rainfall intensity distribution, with the adjustment range not exceeding ±20% of the initial threshold, ensuring that the threshold conforms to both regional characteristics and recent rainfall patterns. This embodiment ultimately outputs a dynamically adjusted extreme rainfall threshold of 0.5-5 mm / h, and supports real-time threshold updates every 24 hours, with the ability to trigger emergency updates based on extreme weather events.

[0089] Furthermore, this method provides continuous optimization capabilities, enabling dynamic updates of model parameters through incremental learning algorithms without retraining the entire network, thus reducing computational costs. New observation data and prediction results are periodically collected, and the prediction error is calculated. When the error exceeds a preset threshold (MAE > 0.5 mm / h), the incremental update process is triggered.

[0090] The update process employs a strategy of freezing the underlying network parameters and fine-tuning the upper-layer parameters. Only the parameters of the top convolutional and fully connected layers of the multi-scale feature extraction network, as well as the model parameters during dynamic threshold calibration, are updated. The training data for incremental learning uses a combination of "new data + historical key samples," including extreme rainfall samples and samples prone to misclassification, ensuring that the model does not lose existing knowledge during the update process. The update cycle can be set to 7-30 days according to actual needs, and manual triggering of emergency updates is supported.

[0091] Using the above method, accurate predictions of extreme rainfall in the next 1-6 hours are achieved, outputting the probability of extreme rainfall occurrence, peak intensity, and affected area. The spatial resolution of the prediction is no less than 0.05°×0.05°, the temporal resolution is 15 minutes / time, and the prediction accuracy is no less than 85%. This method can dynamically adapt to the climate and topographic differences of different regions, and continuously optimize the model performance through incremental learning, providing efficient technical support for meteorological disaster early warning, flood control scheduling, and transportation safety assurance.

[0092] The present invention also provides a deep learning-based extreme rainfall prediction device. The deep learning-based extreme rainfall prediction device provided by the present invention will be described below. The deep learning-based extreme rainfall prediction device described below can be referred to in correspondence with the deep learning-based extreme rainfall prediction method described above. Figure 9 This is a structural block diagram of the deep learning-based extreme rainfall prediction device provided by the present invention, as shown below. Figure 9 As shown, the device includes:

[0093] The data acquisition module 901 is used to acquire multi-source data, preprocess the multi-source data, and construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0094] The feature extraction module 902 is used to input the multi-source heterogeneous fusion dataset into the multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0095] The reinforcement learning module 903 is used to generate synthetic extreme rainfall samples through a generative adversarial network, mix the synthetic extreme rainfall samples with real samples for training, and strengthen the learning of extreme rainfall features through a weighted loss function.

[0096] The weight optimization module 904 is used to introduce channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process.

[0097] The threshold calibration module 905 is used to acquire historical rainfall data and real-time observation data, build an adaptive threshold adjustment model, optimize the extreme rainfall judgment threshold, and generate extreme rainfall warning results.

[0098] In operation, this device first acquires data from 14 channels of the FY-4A satellite AGRI sensor, S-band radar echo data, and ground rain gauge observation data. The acquired data is preprocessed to construct a multi-source heterogeneous fusion dataset. Then, the feature extraction module 902 inputs the preprocessed dataset into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process, obtaining feature representations from low to high dimensions. The reinforcement learning module 903 then generates synthetic extreme rainfall samples using a generative adversarial network. These synthetic extreme rainfall samples are mixed with real samples for training, and a weighted loss function is used to enhance the learning of extreme rainfall features. Next, the weight optimization module 904 introduces channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process. Finally, the threshold calibration module 905 constructs an adaptive threshold adjustment model and optimizes the extreme rainfall judgment threshold, generating extreme rainfall warning results. Through the above process, addressing the core pain points of existing technologies such as insufficient multi-source data fusion, inadequate learning of extreme features, lack of spatiotemporal feature capture, and poor regional adaptability, this technology achieves accurate prediction and early warning of extreme rainfall events through multi-module collaborative design. This solves the problem of poor prediction accuracy of extreme rainfall in existing related technologies and is applicable to scenarios such as meteorological disaster early warning, flood control scheduling, and transportation safety assurance.

[0099] In addition, this device also includes a real-time update module, which continuously absorbs new observation data to update the model parameters through an incremental learning algorithm, without the need to retrain the entire network, ensuring the predictive adaptability of the model in different climate stages. The incremental update cycle can be set to 7-30 days.

[0100] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 10 As shown, the electronic device may include: a processor 1001, a communication interface 1002, a memory 1003, and a communication bus 1004, wherein the processor 1001, the communication interface 1002, and the memory 1003 communicate with each other via the communication bus 1004. The processor 1001 can call logical instructions in the memory 1003 to execute a deep learning-based extreme rainfall prediction method, which includes:

[0101] Collect multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0102] The multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0103] Synthetic extreme rainfall samples are generated by generative adversarial networks, and these samples are mixed with real samples for training. The learning of extreme rainfall features is enhanced by a weighted loss function.

[0104] By introducing channel attention and spatial attention mechanisms, feature weights are dynamically adjusted to capture the temporal dependencies of rainfall processes;

[0105] Historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the threshold for judging extreme rainfall is optimized to generate extreme rainfall warning results.

[0106] Furthermore, the logical instructions in the aforementioned memory 1003 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0107] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the deep learning-based extreme rainfall prediction method provided by the above methods, the method comprising:

[0108] Collect multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0109] The multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0110] Synthetic extreme rainfall samples are generated by generative adversarial networks, and these samples are mixed with real samples for training. The learning of extreme rainfall features is enhanced by a weighted loss function.

[0111] By introducing channel attention and spatial attention mechanisms, feature weights are dynamically adjusted to capture the temporal dependencies of rainfall processes;

[0112] Historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the threshold for judging extreme rainfall is optimized to generate extreme rainfall warning results.

[0113] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the deep learning-based extreme rainfall prediction method provided by the methods described above, the method comprising:

[0114] Collect multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data;

[0115] The multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process;

[0116] Synthetic extreme rainfall samples are generated by generative adversarial networks, and these samples are mixed with real samples for training. The learning of extreme rainfall features is enhanced by a weighted loss function.

[0117] By introducing channel attention and spatial attention mechanisms, feature weights are dynamically adjusted to capture the temporal dependencies of rainfall processes;

[0118] Historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the threshold for judging extreme rainfall is optimized to generate extreme rainfall warning results.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for predicting extreme rainfall, characterized in that, include: Collect multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data; The multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process; Synthetic extreme rainfall samples are generated by generative adversarial networks, and these synthetic extreme rainfall samples are mixed with real samples for training. The learning of extreme rainfall features is enhanced by a weighted loss function. By introducing channel attention and spatial attention mechanisms, feature weights are dynamically adjusted to capture the temporal dependencies of rainfall processes; Historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the threshold for judging extreme rainfall is optimized to generate extreme rainfall warning results; The multi-source heterogeneous fusion dataset is input into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process, including: At the encoding end, the feature differences between small-scale heavy rainfall and large-scale uniform rainfall are captured through multiple downsampling stages to obtain feature maps at different stages; At the decoding end, the feature map obtained at the encoding end is fused with the feature map of the corresponding stage at the decoding end through upsampling to obtain the spatiotemporal evolution features of the process; Each downsampling stage of the encoding end consists of a three-dimensional convolutional layer, a batch normalization layer, a GELU activation function, and a residual multi-scale module. The three-dimensional convolutional layer is used to achieve feature compression in the spatiotemporal dimension; The residual multi-scale module is used to stitch features at different scales and simultaneously embed dilated convolution to expand the receptive field. At the decoding end, the feature map obtained at the encoding end is fused with the feature map of the corresponding stage at the decoding end through upsampling to obtain the spatiotemporal evolution features of the process, including: Feature map size recovery is achieved through transposed 3D convolution. By employing skip connections, feature maps from different stages of the encoding end are fused with feature maps from the corresponding stages of the decoding end to obtain spatiotemporal evolution features; By introducing channel attention and spatial attention mechanisms, feature weights are dynamically adjusted to capture the temporal dependencies of rainfall processes, including: Based on the squeeze-excitation mechanism, global average pooling and global max pooling are performed on the channel dimension of the feature map of the spatiotemporal evolution features. Channel weights are learned through a two-layer fully connected network to dynamically strengthen the channel features that are highly correlated with satellite water vapor channels, radar strong echo channels and extreme rainfall. Based on the self-attention mechanism, a spatiotemporal attention matrix is ​​constructed to calculate the correlation weight between each spatiotemporal location and other locations, thereby enhancing the spatial clustering characteristics and temporal evolution patterns of the core area of ​​extreme rainfall. Historical rainfall data and real-time observation data are acquired, an adaptive threshold adjustment model is constructed, and the threshold for determining extreme rainfall is optimized to generate extreme rainfall warning results, including: The distribution of rainfall intensity in real-time observation data is statistically analyzed using a sliding window. Input the geographical and climatic feature parameters of the current region into the gradient boosting tree model to obtain the initial threshold; The initial threshold is dynamically adjusted based on the real-time rainfall intensity distribution to obtain the extreme rainfall determination threshold for the current area; Based on the extreme rainfall determination threshold, it is determined whether extreme rainfall will occur in the current area, and an extreme rainfall warning result is generated.

2. The deep learning-based extreme rainfall prediction method according to claim 1, characterized in that, The multi-source data is preprocessed to construct a multi-source heterogeneous fusion dataset, including: Radiometric calibration was performed on the multi-channel satellite observation data, and spatial unification was achieved through equal latitude and longitude projection conversion; The radar echo data is subjected to clutter suppression, attenuation correction, and resolution adjustment. Outlier removal and interpolation were performed on the ground station observation data; Based on the timestamps of satellite multi-channel observation data, linear interpolation is used to synchronize the radar echo data with the ground station observation data, and the spatial resolution of all the multi-source data is unified through linear interpolation. The satellite multi-channel observation data, the radar echo data, the ground station observation data, and auxiliary features are stacked according to the channel dimension to obtain a multi-source heterogeneous fusion dataset.

3. The deep learning-based extreme rainfall prediction method according to claim 1, characterized in that, The generative adversarial network includes a generator and a discriminator; The generator uses an improved U-Net structure to generate synthetic extreme rainfall samples that conform to the spatiotemporal distribution of real rainfall. The discriminator optimizes the generation quality by distinguishing the synthetic extreme rainfall samples from real samples.

4. A deep learning-based extreme rainfall prediction device, used to implement the deep learning-based extreme rainfall prediction method according to any one of claims 1-3, characterized in that, include: The data acquisition module is used to acquire multi-source data and preprocess the multi-source data to construct a multi-source heterogeneous fusion dataset; the multi-source data includes satellite multi-channel observation data, radar echo data, and ground station observation data; The feature extraction module is used to input the multi-source heterogeneous fusion dataset into a multi-scale feature extraction network to extract the spatiotemporal evolution features of the rainfall process; The reinforcement learning module is used to generate synthetic extreme rainfall samples through a generative adversarial network, mix the synthetic extreme rainfall samples with real samples for training, and strengthen the learning of extreme rainfall features through a weighted loss function. The weight optimization module is used to introduce channel attention and spatial attention mechanisms to dynamically adjust feature weights and capture the temporal dependencies of the rainfall process; The threshold calibration module is used to acquire historical rainfall data and real-time observation data, build an adaptive threshold adjustment model, optimize the extreme rainfall judgment threshold, and generate extreme rainfall warning results.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the deep learning-based extreme rainfall prediction method as described in any one of claims 1 to 3.

Images