A short-term precipitation prediction method based on gradient activation mapping weight

Abstract

This invention discloses a short-term precipitation forecasting method based on gradient activation mapping weights, belonging to the field of radar echo extrapolation technology. The method acquires continuous time-series radar echo images, preprocesses them, and inputs them into a convolutional neural network to extract multi-scale features. During the training phase, gradient weights for each feature channel are calculated based on backpropagation of prediction errors to generate a spatial attention weight map, dynamically weighting the core region of strong echoes. The weighted features are then input into a gated recurrent unit for temporal evolution modeling, and a decoder outputs the precipitation echo prediction results for multiple future time points. During the inference phase, a heatmap aligned with the echo images is output synchronously, visualizing the prediction basis. This invention can improve the prediction accuracy of key areas of heavy precipitation, suppress background noise interference, and enhance model interpretability.

Description

Technical Field

[0001] This invention relates to the field of radar echo extrapolation technology, specifically to a short-term precipitation forecasting method based on gradient activation mapping weights. Background Technology

[0002] Radar echo extrapolation is a core technology widely used in operational short-term precipitation forecasting. It uses echo sequences obtained from continuous radar observations to predict the spatial distribution and intensity changes of precipitation echoes over the next tens of minutes to several hours, providing crucial support for near-term warnings of small-scale hazardous weather events such as localized thunderstorms and short-duration heavy rainfall. Radar echo sequences not only contain translational motion information but also nonlinear evolution characteristics such as echo generation and dissipation, merging and splitting, and abrupt intensity changes, which are particularly pronounced during the development phase of strong convection. Therefore, this places higher demands on the spatiotemporal joint modeling capabilities of the model.

[0003] Existing traditional extrapolation methods are mostly based on optical flow / correlation tracking and translation assumptions, which have low computational costs but are difficult to characterize the nonlinear evolution of complex weather systems. In recent years, end-to-end deep learning models (such as ConvLSTM, TrajGRU, PredRNN++, etc.) have improved prediction accuracy by learning the spatiotemporal dependencies of historical echo sequences, but they still have common problems in actual operations: First, precipitation echoes have obvious spatial heterogeneity, with strong echo areas accounting for a small proportion and changing rapidly. Conventional networks are easily dominated by large areas of weak echoes / no echoes when learning globally equally, resulting in insufficient focus on key heavy precipitation areas. Second, pixel-level loss (such as MSE / MAE) can easily produce "oversmooth" predictions, resulting in the loss of core area boundaries and detailed textures. Third, deep models lack interpretability, making it difficult to intuitively explain the key areas that the model focuses on and the sources of their contribution to the prediction, which is not conducive to business-side verification and reliable application. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a short-term precipitation forecasting method based on gradient activation mapping weights, which dynamically enhances the focus on key strong echo areas during the forecasting process and simultaneously outputs a short-term precipitation forecast scheme with interpretable and visualized results, so as to balance forecast accuracy, stability and operational availability under complex weather evolution and noise interference scenarios, and solve the problems existing in the background technology.

[0005] Technical solution: The short-term precipitation forecasting method based on gradient activation mapping weights described in this invention includes the following steps:

[0006] (1) Data preparation: acquire radar echo images of continuous time series, and construct input sequence samples after preprocessing. Each sample contains echo images of multiple past times.

[0007] (2) Dual-path feature extraction: The image at each time step in the input sequence is input into a convolutional neural network with shared weights, and multi-scale structural features are extracted along the spatial dimension to generate the corresponding original feature map sequence, while retaining the output of the intermediate layer in the forward propagation process;

[0008] (3) Generate gradient activation mapping weights: During the model training phase, the prediction error is used as the objective function and backpropagated to the original feature map to calculate the gradient contribution of each feature channel to the prediction error; spatial global pooling is performed on the gradient map of each channel to obtain the channel-level importance weights; the channel weights are weighted and linearly fused with the corresponding feature maps, and an initial spatial attention map is generated through a non-linear activation function.

[0009] (4) Attention map refinement: Spatial normalization and threshold pruning are performed on the initial spatial attention map to remove low response regions and generate the final spatial attention weight map. The weight map highlights the strong echo core regions that contribute greatly to the prediction results in the spatial dimension.

[0010] (5) Dynamic weighting of feature maps: The spatial attention weight map is fused with the original feature map element by element. The residual weighting strategy is adopted to enhance the response intensity of key regions while preserving the original feature information, and suppress background noise and feature expression of non-key regions.

[0011] (6) Temporal evolution modeling: The weighted feature map sequence is input into the gated recurrent unit network in chronological order. The gating mechanism is used to selectively memorize and update the spatiotemporal evolution law of precipitation echoes to generate hidden state sequences.

[0012] (7) Decoding the prediction results: The hidden state sequence is input into the decoder network, and the spatial resolution is gradually restored through upsampling and convolution operations to output the precipitation echo prediction images for multiple future times;

[0013] (8) Interpretability visualization: During the model inference stage, the spatial attention weight map is retained and mapped to the resolution of the original input image through interpolation upsampling to generate a heat map that is spatially aligned with the input echo or predicted echo, which is used to intuitively show the key areas that the model focuses on during precipitation forecasting and their dynamic changes.

[0014] Further, step (1) is as follows: noise filtering is performed on the original radar echo image to remove isolated noise points, interpolation method is used to fill in missing pixels, the echo intensity value is linearly normalized to a uniform numerical range, and the image is spatially aligned and cropped according to the geographic coordinates to ensure the spatial consistency of the input sequence.

[0015] Furthermore, in step (2), the convolutional neural network adopts a multi-layer three-dimensional convolutional structure, and simultaneously extracts the spatial structural features of a single frame image and the shallow motion features between multiple frames, providing rich feature representations for the subsequent generation of gradient activation mapping weights.

[0016] Furthermore, in step (3), the channel-level importance weights are obtained as follows: the average value of the gradient map of each feature channel is calculated in the spatial dimension. The average value represents the sensitivity of the channel to the prediction error. The larger the value, the more key precipitation information the channel contains.

[0017] Furthermore, in step (4), spatial normalization adopts the min-max normalization method to map the response value of the initial attention map to the [0,1] interval; threshold pruning sets the response value below the preset threshold to zero in order to highlight the core area of ​​strong echo and eliminate noise interference.

[0018] Furthermore, in step (5), the residual weighting strategy is as follows: the weighted feature map is equal to the original feature map plus the element-wise product of the preset coefficient and the spatial attention weight map; the residual weighting strategy enhances the key regions while maintaining the integrity of the original features.

[0019] Furthermore, in step (7), the decoder network adopts a skip connection structure to directly pass the feature map of the corresponding layer in the encoder to the decoder in order to preserve high-resolution detail information and improve the boundary clarity and texture realism of the predicted echo.

[0020] Furthermore, during model training, a composite loss function is used for parameter optimization. The composite loss function includes a prediction error loss term and an attention consistency constraint term. The prediction error loss term is used to measure the pixel-level difference between the predicted echo image and the real echo image. The attention consistency constraint term is used to guide the attention map generated by the model to maintain a smooth transition between adjacent time steps, avoiding drastic jumps in the attention region.

[0021] Furthermore, the attention consistency constraint is based on the difference between attention maps at adjacent time points. By minimizing this difference, the model’s attention region is made to have temporal continuity and stability, which conforms to the physical laws of precipitation echo evolution.

[0022] Furthermore, during the model training phase, gradient activation mapping weight generation and attention map refinement both participate in backpropagation and parameter updates; during the model inference phase, only forward propagation is performed, and the spatial attention weight map is retained and output for interpretability analysis, without calculating gradient information.

[0023] Furthermore, the heatmap uses color gradients to represent the level of attention. Warm-colored areas represent the core areas of heavy precipitation that the model is highly concerned about, while cool-colored areas represent non-critical areas with lower attention. The heatmap is overlaid with the echo image to form a visualization product, which is used by business personnel to quickly identify the basis for the model's predictions.

[0024] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: Focus on Key Regions: The GAMW module adaptively calculates channel / spatial weights using gradient information, explicitly weighting the core area of ​​strong echoes, reducing the interference of large-area weak echo backgrounds on learning, and improving the characterization ability in heavy precipitation scenarios. Enhanced Interpretability: A heatmap is output simultaneously with the prediction, intuitively reflecting the model's areas of interest in the input echo sequence, which can be used for rapid verification and reliable application of prediction data by the business side. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0027] like Figure 1 As shown, this invention provides a short-term precipitation forecasting method based on gradient activation mapping weights. The core process of this invention is as follows: "Spatial feature extraction—gradient mapping weighted focusing—temporal modeling and prediction—interpretable heatmap output".

[0028] (1) Obtain continuous time series radar echo image data from the radar observation system and construct an input sequence in chronological order. Preferably, the input consists of 5 consecutive frames of radar echo images (corresponding to the past 30 minutes, with a time resolution of 6 min), each frame with a spatial resolution of 480×480 and 1 channel, forming an input tensor containing both time and spatial dimensions, used to predict the next 10 frames of radar echo image sequence (corresponding to the next 60 minutes).

[0029] (2) Preprocess the input radar echo sequence, including noise and clutter suppression, missing measurement filling, pixel value normalization (e.g., linear mapping to the [0,1] interval or segmented normalization by threshold) and spatial clipping / alignment, to reduce the impact of data quality fluctuations on prediction.

[0030] (3) Input the preprocessed sequence into the CNN encoder for spatial feature extraction. The CNN encoder consists of multiple convolutional and downsampling structures, which are used to extract multi-scale spatial structure features from the echo image at each time step to obtain a feature map sequence.

[0031] (4) Introduce the GAMW module between CNN and GRU. The GAMW module calculates the importance weight of each channel based on the gradient information of the feature map of the prediction error, and generates a spatial attention map to dynamically weight the feature map: giving higher attention to the core area of ​​strong echoes, while suppressing the interference of non-critical area features, thereby improving the ability to focus on key areas and enhancing interpretability.

[0032] (5) Input the GAMW-weighted feature map sequence into the GRU layer for temporal modeling. The GRU uses a gating mechanism to selectively memorize and update historical information, capture the temporal evolution of precipitation echoes, and output the hidden state sequence for multiple future moments. Then, the hidden state is mapped to the echo prediction results for the next 10 frames through a decoder (upsampling / deconvolution, etc.).

[0033] (6) During the reasoning or evaluation phase, the spatial attention map output by the GAMW module is upsampled to the original image resolution and superimposed with the input / predicted echo to form a heat map product, which is used to show the distribution of the model’s attention to key precipitation areas and to explain the basis of the prediction.

[0034] (7) During the training phase, a composite loss function and a parameter optimization strategy based on GAMW weights are adopted. The composite loss preferably includes a prediction error term and an attention constraint term: the prediction error term is used to measure the pixel-level difference between the predicted echo and the true echo (such as MSE / MAE); the attention constraint term is used to constrain the consistency between the model's attention and GAMW attention to stabilize the focusing behavior. Parameter updates can use an Adam-type optimizer, and GAMW weights can be introduced into the gradient update to weight the gradients of key channels / regions to accelerate convergence and improve the learning efficiency of key regions.

[0035] The specific process is as follows:

[0036] Data Acquisition and Sample Construction: Raw radar echo images are acquired from the radar echo observation system in chronological order, denoted as... ,in For time indexing, and These represent the height and width of the image, respectively. In this embodiment, continuous values ​​are used. The historical echo image of the frame is used as the input sequence, and the subsequent consecutive frames are taken. The frame-by-frame real echo image is used as the supervision signal and is represented as follows:

[0037]

[0038] in, Represents the model input sequence. Indicates the target sequence to be predicted; ; This indicates the number of time steps to input. This indicates the number of output time steps. In this embodiment, the preferred input time resolution is... Therefore, 5 frames correspond to the past 10 frames correspond to the future .

[0039] Preprocessing: In order to reduce the impact of clutter, isolated noise points and missing pixels on the prediction results, noise suppression, missing pixel filling, normalization and spatial alignment are performed on each frame of the original image in sequence.

[0040] Preferably, noise suppression employs Median filtering is calculated as follows:

[0041]

[0042] in, Indicates the number after noise suppression. Frame image at position Pixel value at; , This represents the relative offset within the filtering window; This indicates the median value operation. The above processing can effectively remove isolated spike noise while preserving the main boundary of the echo.

[0043] For regions with missing data, a weighted interpolation method based on the effective neighborhood is used for filling in the missing data. The calculation formula is as follows:

[0044]

[0045] in, Indicates position The set of surrounding valid neighboring pixels; Represents neighboring pixels The interpolation weight can preferably be determined by the reciprocal of the distance between the neighboring pixel and the missing measurement location; This represents the pixel value after missing measurement filling is completed. (When position...) When the original observation value was valid, it could be directly set to...

[0046] .

[0047] Pixel value normalization can be achieved by linear mapping to... The interval method is as follows:

[0048]

[0049] in, and These represent the lower and upper limits of the echo intensity, respectively; The smallest positive number introduced to prevent the denominator from being zero; This represents a truncation function used to restrict the result to a specific range. Within a certain range. For example, when the raw data is represented in dBZ, the range can be preset based on the statistical range of the dataset or business experience. and .

[0050] Alternatively, a threshold-based segmented normalization method can be used to highlight strong echo regions. The calculation formula is as follows:

[0051]

[0052] in, and Let these represent the low threshold and the high threshold, respectively, and satisfy the following conditions: When the echo intensity is lower than The time-normalized result is set to 0 when the echo intensity is higher than 0. The time-normalized result is set to 1, and the intermediate interval is mapped linearly. This method can improve the model's sensitivity to medium-to-high intensity echo regions while ensuring numerical stability.

[0053] After normalization, the images from all time points are uniformly mapped onto the same target grid to ensure spatial consistency. The expression for this is:

[0054]

[0055] in, Indicates the target grid coordinates; This represents the coordinate mapping relationship from the target grid to the original image grid; This represents the input image after spatial alignment and cropping. In practice, bilinear interpolation can be used to perform pixel resampling after coordinate mapping.

[0056] Spatial feature encoding: This involves encoding the preprocessed images at each time step. The input is a CNN encoder that extracts multi-scale spatial features, resulting in a sequence of feature maps.

[0057]

[0058] in, Indicates the CNN encoder; Indicates encoder parameters; and These represent the height and width of the feature map after downsampling, respectively; Indicates the number of channels; Indicates the first At the [time]th moment Each channel is located in The encoder preferably employs a multi-layer convolution, normalization, and downsampling structure, and can be combined with skip connections to preserve high-resolution detail information.

[0059] Gradient activation mapping weight calculation: During the training phase, to predict the error term. As the objective function, the feature map Perform backpropagation to obtain the gradients for each channel:

[0060]

[0061] in, The loss function is applied to the feature map. The partial derivatives of the value reflect the sensitivity of the location feature to the prediction error; Indicates the channel index.

[0062] Perform global average pooling on the gradient map in the spatial dimension to obtain the importance weights of each channel:

[0063]

[0064] in, Indicates the first At the [time]th moment Channel-level weights for each channel; This represents the total number of pixels in the feature map. The larger the value, the more significant the contribution of the corresponding channel to the current prediction error.

[0065] The feature maps are weighted and linearly fused using channel weights, and an initial attention map is generated using a linear rectified function.

[0066]

[0067] in, This represents the initial gradient activation mapping value; This represents a linear rectified activation function used to suppress negative responses and preserve key positive activation regions.

[0068] Attention map refinement and feature weighting: The initial attention map is normalized to map it to... The interval is calculated using the following formula:

[0069]

[0070] in, This represents the normalized attention value; and They represent the first The minimum and maximum values ​​in the initial attention graph at each time step; It is a very small positive number.

[0071] To further suppress low-response noise, a threshold segmentation method can be used to obtain the final spatial attention weight map:

[0072]

[0073] in, This represents the spatial attention weights ultimately used for feature weighting; and Let represent the low response threshold and the high response threshold, respectively, and satisfy . This process can make the weak response background approach 0, while the strong response core region approaches 1.

[0074] Subsequently, the spatial attention weight map is broadcast by channel and then fused with the original feature map using residual weighting:

[0075]

[0076] in, This represents the weighted characteristic response; This represents the weighting coefficient, used to control the magnitude of attention enhancement, and is preferably taken as a positive value; when The larger the value, the more pronounced the enhancement in the response of the critical region at the corresponding location.

[0077] Temporal modeling and predictive decoding: weighted feature maps After pooling or flattening operations, the input feature vector is fed into the GRU for time series modeling. Let the input feature vector be:

[0078]

[0079] in, Indicates the first The feature vectors input to the GRU at each time step; This represents a global average pooling, flattening, or equivalent feature compression mapping operation.

[0080] The update process of a GRU cell can be represented as follows:

[0081]

[0082]

[0083]

[0084] in, Indicates an update to the door; Indicates that the door is being reset; and These represent the hidden states of the previous and current time points, respectively. Indicates the candidate hidden state; and The weight matrix is ​​trainable. For bias terms; This represents the Sigmoid activation function; Represents the hyperbolic tangent activation function; This indicates element-wise multiplication.

[0085] After obtaining the hidden state, the spatial resolution is restored through the decoder and the future is output. Frame prediction results:

[0086]

[0087] in, Indicates the first Predicted echo images for a future time; Indicates decoder; This represents the decoder parameters. The decoder can use a combination of upsampling and convolution to achieve resolution restoration, and can work with skip connections in the encoder to improve the reconstruction quality of edge and texture details.

[0088] Loss Function and Optimization: During the training phase, a composite loss function consisting of a prediction error term and an attention constraint term is used. The prediction error term can be the mean squared error, defined as:

[0089]

[0090] in, This represents the loss due to prediction error; and These represent the predicted echo and the actual echo at the [missing information] th [missing information]. Location at a future moment The pixel value at that location.

[0091] To ensure the smooth continuity of the attention map over time, attention consistency constraints can be constructed:

[0092]

[0093] in, This indicates a loss of attentional consistency; This term represents the summation of absolute values ​​element by element; it is used to constrain the changes in the attention map at adjacent times from drastic jumps, thus better conforming to the physical law of continuous evolution of precipitation echoes.

[0094] Therefore, the total loss function can be written as:

[0095]

[0096] in, Indicates the total loss; This represents the balance coefficient, used to adjust the weighting relationship between prediction accuracy and attention smoothness.

[0097] For parameter optimization, the Adam optimizer is preferred, and its parameter update can be written as:

[0098]

[0099] in, and They represent the first Second and third Model parameters for the next iteration; Indicates the learning rate; and These represent the bias correction results for the first-order moment estimate and the second-order moment estimate, respectively; A stabilizing term is set up to prevent the denominator from being zero.

[0100] Interpretable output: During the model inference phase, retain and output the final spatial attention weight map. It is then upsampled to the original image resolution, calculated as follows:

[0101]

[0102] in, This represents a heatmap spatially aligned with the original input image. This indicates that the attention map is interpolated and upsampled to a size of [size missing]. The operation can then be performed. The heatmap can then be overlaid with the input echo image or predicted echo image to visually reflect the key strong echo areas and their dynamic changes that the model focuses on during short-term precipitation forecasting.

[0103] experiment:

[0104] Dataset: The HKO-7 real radar echo dataset was used as the experimental dataset. The dataset information is shown in Table 1.

[0105] Table 1 HKO-7 Dataset ;

[0106] Data preprocessing: To improve model convergence and robustness, the original radar echo sequence was preprocessed by normalization, abnormal noise suppression and missing measurement processing, and the samples were organized according to time continuity.

[0107] Training settings: The input is a sequence of 5 consecutive radar echo images, and the output is a prediction sequence of the next 10 frames; the preferred data partition is 70% training set, 10% validation set, and 20% test set; the preferred learning rate is 1e-4.

[0108] Experimental results:

[0109] Model evaluation metrics include image quality metrics (MAE, MSE, SSIM, PSNR), etc. The GAMW-CGNet (a convolutional gated network based on gradient activation mapping weights) of this invention was compared with ConvLSTM (convolutional long short-term memory network), TrajGRU (trajectory-gated recurrent unit network), and PredRNN++ (enhanced predictive recurrent neural network) models on the HKO-7 dataset through unified training and testing. The comparison results are shown in Table 2.

[0110] Table 2 Comparison of Model Accuracy and Performance ;

[0111] Further ablation experiments were conducted to verify the contribution of the GAMW module to the performance improvement. The baseline model without the GAMW module (w / o GAMW) and the model with the GAMW module (w GAMW) were compared under the same settings, and the results are shown in Table 3.

[0112] Table 3 Comparison Results of Ablation Experiments .

Claims

1. A short-term precipitation forecasting method based on gradient activation mapping weights, characterized in that, Includes the following steps: (1) Data preparation: acquire radar echo images of continuous time series, and construct input sequence samples after preprocessing. Each sample contains echo images of multiple past times. (2) Dual-path feature extraction: The image at each time step in the input sequence is input into a convolutional neural network with shared weights, and multi-scale structural features are extracted along the spatial dimension to generate the corresponding original feature map sequence, while retaining the output of the intermediate layer in the forward propagation process; (3) Generate gradient activation mapping weights: During the model training phase, the prediction error is used as the objective function and backpropagated to the original feature map to calculate the gradient contribution of each feature channel to the prediction error; spatial global pooling is performed on the gradient map of each channel to obtain the channel-level importance weights. The channel weights are linearly fused with the corresponding feature maps, and an initial spatial attention map is generated by a non-linear activation function. (4) Attention map refinement: Spatial normalization and threshold pruning are performed on the initial spatial attention map to remove low response regions and generate the final spatial attention weight map. The weight map highlights the strong echo core regions that contribute greatly to the prediction results in the spatial dimension. (5) Dynamic weighting of feature maps: The spatial attention weight map is fused with the original feature map element by element. The residual weighting strategy is adopted to enhance the response intensity of key regions while preserving the original feature information, and suppress background noise and feature expression of non-key regions. (6) Temporal evolution modeling: The weighted feature map sequence is input into the gated recurrent unit network in chronological order. The gating mechanism is used to selectively memorize and update the spatiotemporal evolution law of precipitation echoes to generate hidden state sequences. (7) Decoding the prediction results: The hidden state sequence is input into the decoder network, and the spatial resolution is gradually restored through upsampling and convolution operations to output the precipitation echo prediction images for multiple future times; (8) Interpretability visualization: During the model inference stage, the spatial attention weight map is retained and mapped to the resolution of the original input image through interpolation upsampling to generate a heat map that is spatially aligned with the input echo or predicted echo, which is used to intuitively show the key areas that the model focuses on during precipitation forecasting and their dynamic changes.

2. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, Step (1) is as follows: noise filtering is performed on the original radar echo image to remove isolated noise points, interpolation method is used to fill in missing pixels, the echo intensity value is linearly normalized to a uniform value range, and the image is spatially aligned and cropped according to the geographic coordinates to ensure the spatial consistency of the input sequence.

3. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, In step (2), the convolutional neural network adopts a multi-layer three-dimensional convolutional structure, and simultaneously extracts the spatial structural features of a single frame image and the shallow motion features between multiple frames, providing rich feature representations for the subsequent generation of gradient activation mapping weights.

4. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, In step (3), the channel-level importance weights are obtained as follows: the average value of the gradient map of each feature channel is calculated in the spatial dimension. The average value represents the sensitivity of the channel to the prediction error. The larger the value, the more key precipitation information the channel contains.

5. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, In step (4), spatial normalization adopts the min-max normalization method to map the response value of the initial attention map to the [0,1] interval; threshold clipping sets the response value below the preset threshold to zero in order to highlight the core area of ​​strong echo and eliminate noise interference.

6. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, In step (5), the residual weighting strategy is as follows: the weighted feature map is equal to the original feature map plus the element-wise product of the preset coefficient and the spatial attention weight map; the residual weighting strategy enhances the key regions while maintaining the integrity of the original features.

7. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, In step (7), the decoder network adopts a skip connection structure, which directly transmits the feature map of the corresponding layer in the encoder to the decoder in order to preserve high-resolution detail information and improve the boundary clarity and texture realism of the predicted echo.

8. The short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, During model training, a composite loss function is used for parameter optimization. The composite loss function includes a prediction error loss term and an attention consistency constraint term. The prediction error loss term is used to measure the pixel-level difference between the predicted echo image and the real echo image. The attention consistency constraint is used to guide the attention map generated by the model to maintain a smooth transition between adjacent time steps, avoiding drastic jumps in attention regions.

9. A short-term precipitation forecasting method based on gradient activation mapping weights according to claim 8, characterized in that, The attention consistency constraint is based on the difference between attention maps at adjacent time points. By minimizing the difference, it promotes the continuity and stability of the model's attention area in time, which conforms to the physical laws of precipitation echo evolution.

10. A short-term precipitation forecasting method based on gradient activation mapping weights according to claim 1, characterized in that, During the model training phase, gradient activation mapping weight generation and attention map refinement both participate in backpropagation and parameter updates; during the model inference phase, only forward propagation is performed, and the spatial attention weight map is retained and output for interpretability analysis, without calculating gradient information.

Images