Lightweight precipitation nowcasting method based on frequency domain perception cbam and hollow spatial pyramid

Abstract

The application discloses a kind of light precipitation nowcasting methods based on frequency domain perception CBAM and hollow spatial pyramid, first acquire data and pre-process, then construct light U-Net backbone network: construct a backbone network of U-Net form, the U-Net is further carried out multi-scale context feature extraction using encoder-decoder symmetric architecture;Then decode and jump connection, by decoder from the fusion feature step-by-step is carried out up sampling operation;The feature map output by the last layer of decoder, through a 1×1 output convolution layer, it is directly mapped into multiple frames continuous forecast field, to realize from input radar chart sequence to output forecast chart sequence end-to-end sequence mapping;Finally, the model is trained and optimized, after training is completed, model can generate high-quality short-time precipitation forecast result according to the input radar chart sequence.The method realizes high-precision, low false alarm, high computational efficiency short-time precipitation forecast.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting, specifically a lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid. Background Technology

[0002] Currently, precipitation forecasting methods are mainly divided into two categories: methods based on traditional physical models and methods based on deep learning. Traditionally, meteorologists rely on Numerical Weather Prediction (NWP) models for forecasting. NWP uses supercomputers to solve atmospheric dynamic equations. With its rigorous physical foundation and continuous computational optimization, it can accurately describe weather phenomena, and its accuracy can be gradually improved by increasing computational resources. However, its dependence on high-performance computing resources and long computation time makes it difficult to meet the needs of short-term (short-term) precipitation forecasting.

[0003] Radar echo extrapolation, another classic method, uses radar observation data from previous and current moments to estimate the velocity and direction of the echo, thereby predicting its future location and intensity. This method offers rapid response and has been widely applied to real-time tracking of severe convective systems and nowcasting rainfall. Optical flow is a typical example, possessing strong spatiotemporal continuity modeling capabilities, but it has significant limitations: it is sensitive to image noise, leading to increased optical flow estimation errors; furthermore, it has high computational complexity, requiring substantial resources in complex scenarios.

[0004] In recent years, with the rise of deep learning technology, Convolutional Neural Networks (CNNs) have demonstrated outstanding performance in radar echo extrapolation tasks due to their advantages in image feature extraction and spatial correlation capture. However, pure CNN methods often produce blurry extrapolation results. To address this, researchers have turned to advanced architectures such as Generative Adversarial Networks (GANs) or diffusion models, which significantly improve the realism of predicted details.

[0005] Despite the significant progress made by the above methods, the field of short-term precipitation forecasting still faces two key challenges: First, the prediction accuracy for extreme events such as heavy precipitation is insufficient. Short-term heavy precipitation events are characterized by small sample sizes and uneven distribution. Insufficient training data makes it difficult for models to capture their complex spatiotemporal evolution patterns, resulting in significant biases in areas of heavy precipitation and affecting the accuracy of intensity predictions for extreme events. Existing models perform poorly in terms of echo boundaries, local details, and overall structure fidelity, making it difficult to meet the needs of refined forecasting.

[0006] Second, the model architecture is complex and inference efficiency is low. In pursuit of high accuracy and detailed reproduction, existing model architectures are becoming increasingly complex, with the number of parameters skyrocketing. Radar data itself is high-dimensional and massive, which further amplifies the computational demands, leading to high training and inference costs and making it difficult to achieve real-time deployment in resource-constrained real-world business environments (such as edge devices).

[0007] Therefore, how to design a precipitation nowcasting model that can accurately characterize the multi-scale structure of heavy precipitation systems, especially the boundary and detailed features, while maintaining a lightweight design to achieve efficient inference is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] This invention addresses the following technical problems in existing precipitation nowcasting technologies: traditional physics-based numerical forecasting models have insufficient accuracy and high computational resource consumption in short-term nowcasting (e.g., 0-6 hours); existing deep learning models (e.g., standard U-Net) have limitations in capturing multi-scale, especially high-frequency detailed information, and have weak ability to suppress false alarms (i.e., predicting areas of precipitation that have not actually occurred), and have a large number of model parameters, making them difficult to deploy on resource-constrained edge computing devices.

[0009] To address the aforementioned technical problems, this invention provides a lightweight precipitation nowcasting method based on frequency domain sensing CBAM and a void spatial pyramid. This method achieves high-precision, low-false-alarm, and high-computational-efficiency short-term precipitation forecasts by fusing frequency domain information with a spatial attention mechanism, combining multi-scale contextual feature extraction and lightweight network design.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void spatial pyramid, comprising the following steps: Step 1: Data Acquisition and Preprocessing: Acquire the radar precipitation dataset and preprocess it to obtain the radar precipitation map sequence and the corresponding actual precipitation data values. This preprocessing step aims to convert the raw radar data into a standardized time-series image format suitable for input to deep learning models.

[0011] Step 2: Construct a lightweight U-Net backbone network: Construct a U-Net-style backbone network, which adopts an encoder-decoder symmetric architecture. The encoder includes an input module and four cascaded downsampling modules. To enhance feature extraction capabilities while maintaining a lightweight design, frequency-aware convolutional attention modules are introduced at the input of the downsampling modules in the first two layers (i.e., the two layers closest to the input) to strengthen the capture of high-frequency details and global frequency domain features. Standard convolutional attention modules are used at the input and output of the downsampling modules in the last two layers (i.e., the two deeper layers) to enhance the channel and spatial selectivity of semantic features.

[0012] Step 3: Multi-scale contextual feature extraction: A dilated spatial pyramid pooling module is applied to the bottleneck layer features between the encoder and decoder. This module effectively captures multi-scale contextual features without reducing the feature map resolution by setting multiple parallel dilated convolution branches with different dilation rates, and fuses the outputs of each branch to finally output a fused feature containing rich multi-scale information.

[0013] Step 4, Decoding and Skip Connections: The decoder performs upsampling operations step by step starting from the fused features. Each upsampling level uses transposed convolution to double the feature map space size and halve the number of channels. Simultaneously, the upsampled features at that level are spliced ​​with the features of the corresponding layer (i.e., the layer with the same resolution) of the encoder through skip connections to fuse shallow detail information. The spliced ​​features are further processed using depthwise separable convolutions to significantly reduce the number of model parameters and computational cost while maintaining accuracy.

[0014] Step 5, Output Mapping: The feature map output from the last layer of the decoder is directly mapped to a multi-frame continuous prediction field through a 1×1 output convolutional layer, thereby realizing end-to-end sequence mapping from the input radar map sequence to the output prediction map sequence.

[0015] Step 6, Model Training and Optimization: The constructed model is trained end-to-end, and a hybrid loss function is used to optimize the model parameters. This hybrid loss function integrates multiple loss terms to supervise the model's performance in areas such as pixel accuracy, structural similarity, false alarm suppression, and frequency domain characteristics. After training, the model can generate high-quality short-term precipitation forecasts based on the input radar image sequence.

[0016] Preferably, the frequency domain-aware convolutional attention module mentioned in step 2 above has the following internal structure: a channel attention submodule, a spatial attention submodule, a frequency domain attention submodule, and a feature fusion module.

[0017] Preferably, the frequency domain attention submodule is implemented as follows: First, a two-dimensional real fast Fourier transform is performed on the input feature map to obtain a complex spectrum, from which the amplitude spectrum and phase spectrum are separated. In the frequency domain plane, the spectrum is divided into low-frequency, mid-frequency, and high-frequency regions based on radial frequency coordinates. Amplitude scaling factors for low and high frequencies are learned using learnable scalar parameters. Subsequently, a continuous weight matrix for the entire frequency domain plane is obtained through linear interpolation at the radial frequency, and after processing by an activation function, the frequency domain attention weight matrix is ​​finally generated. This mechanism enables the network to adaptively emphasize or suppress different frequency components, particularly enhancing high-frequency information that characterizes precipitation texture details.

[0018] Preferably, the spatial attention submodule is implemented as follows: the input feature map is sequentially passed through 1×1 convolutional layers to adjust the number of channels, followed by batch normalization and activation functions; then, multiple cascaded 3×3 convolutional layers (each followed by batch normalization and activation functions) are used to learn the feature importance in the spatial dimension; finally, the number of channels is adjusted back to the original dimension through 1×1 convolutional layers, and the final spatial attention weight map is generated through the Sigmoid activation function and dropout operation.

[0019] Preferably, the channel attention submodule is implemented as follows: the input feature map is subjected to global average pooling and global max pooling respectively, compressing each channel into a scalar; then batch normalization and activation functions are performed; the channel dimensions are transformed by a shared multilayer perceptron; finally, the channel attention weight map is generated by the sigmoid activation function and dropout operation.

[0020] Preferably, to effectively suppress false alarms (i.e., incorrectly predicting precipitation in a background area without precipitation), this invention innovatively introduces a background false activation penalty mechanism into the frequency domain perceptual convolutional attention module. This mechanism generates a set of background pixels from the input radar precipitation map sequence based on a preset threshold. Then, the squares of the model's predicted values ​​on this set of background pixels are summed and normalized to obtain a background penalty term. This background penalty term is multiplied element-wise with the aforementioned channel attention weight map and spatial attention weight map, thereby achieving feature suppression of the background region on the output feature map, significantly reducing the risk of false alarms generated by the model in areas without precipitation.

[0021] Preferably, the feature fusion module is implemented as follows: the generated channel attention weight map, spatial attention weight map, and frequency domain attention weight map are concatenated along the channel dimension. The concatenated feature map is then passed sequentially through a 3×3 convolutional layer, batch normalization, activation function, and dropout operation, followed by a Sigmoid activation function to generate the final fusion weights. Finally, the fusion weights are multiplied element-wise with the input feature map. Furthermore, preferably, the module introduces a residual connection, with the final output equal to the product of the fusion weights and the input feature map, plus a learnable residual coefficient multiplied by the frequency domain modulation feature map. The generation process of the frequency domain modulation feature map is as follows: the feature map output by the convolutional attention module is subjected to a two-dimensional real fast Fourier transform again to obtain its amplitude and phase spectra; a learnable amplitude weight matrix is ​​constructed in the frequency domain plane, and this matrix is ​​multiplied point-by-point with the original amplitude spectrum while keeping the phase spectrum unchanged; finally, a two-dimensional real fast Fourier inverse transform is performed on the weighted complex spectrum to obtain the frequency-enhanced spatial feature map. This residual design further enhances the network's utilization of key frequency domain information.

[0022] Preferably, the hollow spatial pyramid pooling module in step 3 above introduces a specific implementation of multi-dilation rate dilated convolution. The method is as follows: the preliminary precipitation prediction result and the corresponding actual precipitation value from the current step are concatenated with the original radar precipitation map sequence input to U-Net along the channel dimension to form the original sequence. The hollow spatial pyramid pooling module extracts multi-scale features from this original sequence through four parallel 3×3 dilated convolution branches with dilation rates of 3, 6, 12, and 18, respectively. The outputs of each branch are fused with the output of a global average pooling branch at the channel level. The fused output sequence is then compressed using a 1×1 convolution, and a frequency-domain aware convolutional attention module is used again for feature purification and fusion, finally outputting the bottleneck features. This design ensures that the bottleneck layer features have extremely strong multi-scale expressive power.

[0023] Preferably, the hybrid loss function described in step 6 above is obtained by weighted summation of the following four parts: weighted mean squared error loss, structural similarity loss, background spurious activation penalty loss, and Fourier loss.

[0024] As a preferred approach, the weighted mean squared error loss works as follows: First, for each forecast frame of each sample in each batch, the squared difference between the predicted field and the actual field is calculated pixel by pixel. Then, this difference is multiplied by a heavy precipitation weighting factor (which is 1 plus the product of the heavy precipitation weighting factor and the heavy precipitation region mask; the heavy precipitation region mask is 1 when the actual field pixel value exceeds a preset threshold, and 0 otherwise). Next, the weighted mean squared error is normalized by dividing the above weighted mean squared error by the mean of the heavy precipitation weighting factor over the entire batch and spatial dimensions. Finally, the normalized result is multiplied by a preset frame weight for each frame and averaged over the batch and frame dimensions. This loss function effectively solves the problem of sample imbalance in precipitation data (heavy precipitation samples are far fewer than weak precipitation or non-precipitation samples).

[0025] As a preferred approach, structural similarity loss is employed: for each forecast frame, a structural similarity index between the predicted and actual fields is calculated, and this index is subtracted from 1 to obtain the single-frame loss. This loss is then multiplied by the corresponding frame weight, and finally averaged over all frames. This loss term guides the model to learn the overall morphology and texture structure of precipitation.

[0026] As a preferred approach, the background spurious activation penalty loss works by pre-defining a set of background pixels based on the real field (e.g., considering pixels with actual precipitation values ​​below a very low threshold as background). The sum of squares of the pixel values ​​belonging to this background set in the prediction field is normalized by dividing by the total number of background pixels, and then averaged across batches and frames. This loss term directly penalizes the model for producing spurious predictions in the background region.

[0027] As a preferred approach, the Fourier loss is as follows: For each forecast frame, a two-dimensional real fast Fourier transform is performed on both the predicted and actual fields, and the amplitude spectra are obtained. One is added to the amplitude spectrum, and the logarithm is taken. The absolute value of the difference between the two logarithmic amplitude spectra is calculated, multiplied by the frame weight corresponding to that frame, and finally averaged over the batch and frame dimensions. This loss term ensures that the model accurately fits the real data in the frequency domain, especially for the recovery of high-frequency details.

[0028] This invention has the following characteristics and beneficial effects: This invention addresses the challenge of existing technologies in capturing the complex details and extreme intensity of heavy precipitation systems by proposing an innovative network structure. First, a frequency-domain perceptual CBAM module is introduced into the shallow layer of the encoder. Through two-dimensional Fast Fourier Transform (FFT), the low-frequency and high-frequency components of the feature map are adaptively modulated, effectively enhancing the model's ability to model high-frequency details such as echo boundaries and local strong centers. Second, a hollow spatial pyramid pooling (ASPP) module is integrated into the bottleneck layer. This module utilizes multi-dilation rate dilated convolution to capture multi-scale contextual information, improving the model's ability to express the overall structure of the precipitation system. As shown in Table 1 of the embodiments, this invention outperforms existing methods in all key metrics on the Shanghai radar dataset: compared to the best baseline model, the Critical Success Index (CSI) is improved by 5.3%, the Heidke Skill Score (HSS) by 5.2%, and the Structural Similarity Index (SSIM) by 4.9%, while the Perceptual Image Patch Similarity (LPIPS) is reduced by 25.3%. This indicates that this invention can not only more accurately pinpoint precipitation areas but also generate forecast fields with clearer boundaries and more complete structures.

[0029] This invention improves accuracy while fully considering the stringent requirements of inference efficiency in real-world business scenarios. By employing depthwise separable convolutions throughout the encoder and decoder, combined with a non-recursive single forward propagation design, the computational complexity and inference latency of the model are significantly reduced. As shown in Table 3 of the embodiments, the model of this invention has 9.86M parameters, a single-sequence floating-point operation (FLOPs) of only 7.69G, a single-frame computation as low as 0.385G, and an inference latency of only 6.13 milliseconds. Compared with SmaAt-UNet, which is also based on U-Net, this invention increases inference latency by only 60% (6.13ms vs 3.83ms) with a limited increase in the number of parameters, but achieves a significant improvement of 21.6% in the CSI metric (0.3788 vs 0.3115). Compared with the computationally intensive ConvGRU (FLOPs up to 247.09G, latency 30.59ms), this invention improves inference speed by nearly 5 times and reduces computation by more than 30 times. This demonstrates that the invention achieves an excellent balance between accuracy and efficiency, and has the ability to run in real time on resource-constrained edge devices.

[0030] This invention addresses the common false alarm problem in precipitation forecasting by introducing a background spurious activation penalty term (L_bg) into the hybrid loss function. This term identifies truly non-precipitating background pixels by using a preset threshold and applies a sum-of-squares normalization penalty to the predicted output in these regions, thereby forcing the model to reduce false alarms in background areas during training. As shown in Tables 1 and 2 of the embodiments, the complete model significantly improves on multiple CSI metrics compared to the variant that removes the frequency domain attention module, indirectly demonstrating that the synergistic effect of the background penalty term and the frequency domain attention mechanism effectively reduces false echoes and improves forecast reliability.

[0031] This invention introduces a heavy precipitation region mask into the frequency domain sensing CBAM and hollow spatial pyramid pooling (ASPP) modules. Regions exceeding a set threshold are assigned a value of 1 to preserve their complete features, while regions below a threshold are assigned a value of 0 to suppress their complete features, thereby enhancing feature extraction from heavy precipitation regions. A hybrid loss function incorporating this heavy precipitation region mask is used to calculate the total loss. Regions exceeding a set threshold are assigned a value of 1 to enhance their impact on model training, while regions below a threshold are assigned a value of 0 to suppress their impact on model training, thus improving the prediction accuracy for heavy precipitation regions. Furthermore, time-weighted calculations are used to enhance the accuracy of short-term precipitation areas and the continuity of long-term trends.

[0032] This invention further ensures the physical rationality of the forecast results by designing a composite loss function. Specifically, the frame weighting mechanism (assigning higher weights to the first and farthest frames) strengthens the model's learning of short-term accuracy and long-term trend continuity; the Fourier loss term (L_Fourier) ensures that the high and low frequency component distributions of the predicted and true fields are consistent in the logarithmic amplitude spectrum space, echoing the shallow frequency-domain sensing CBAM of the encoder, jointly ensuring the fidelity of the forecast field in the frequency domain. (See the embodiments.) Figure 5 As shown in the visualization results, the present invention can better maintain the shape and boundary integrity of heavy precipitation areas in long-term forecasts, avoiding the problems of over-smoothing or structural breakage common in traditional methods.

[0033] In summary, this invention, through its innovative frequency-domain sensing CBAM-ASPP collaborative architecture, lightweight deep separable convolutional design, and composite loss function for precipitation forecasting tasks, solves the two core pain points of "insufficient prediction accuracy" and "low inference efficiency" in existing technologies, providing a practical solution with high accuracy, high efficiency, and low false alarm rate for short-term precipitation nowcasting. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the lightweight U-Net model based on frequency domain sensing CBAM and void space pyramid of the present invention; Figure 2 A schematic diagram illustrating the amplitude spectrum visualization results of a single sample from the Shanghai dataset; Figure 3 This is a structural diagram of a frequency-domain enhanced CBAM module. Figure 4 A schematic diagram illustrating a case of frequency domain attention (2D FFT) processing; Figure 5 A schematic diagram illustrating examples of the prediction results from each model; Figure 6 Example visualization of encoder and bottleneck layer feature maps. Detailed Implementation

[0035] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the present invention. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0037] This embodiment provides a lightweight precipitation nowcasting method based on frequency-domain sensing CBAM and a void space pyramid. This method constructs and trains a lightweight neural network model called FCA-UNet (Frequency-enhanced CBAM ASPP U-Net) to achieve end-to-end prediction from historical radar echo sequences to future multi-frame precipitation fields. The following section combines... Figures 1 to 6 The specific implementation steps, module structure, and parameter settings of the present invention will be described in detail.

[0038] like Figure 1 As shown, the FCA-UNet model constructed in this invention adopts an encoder-decoder U-Net backbone network. The encoder is responsible for multi-scale feature extraction from the input multi-frame radar image sequence, and the decoder is responsible for progressively restoring the bottleneck features to the original spatial resolution and outputting the prediction result. The specific steps are as follows: Step 1: Obtain the radar precipitation dataset and perform preprocessing. In this embodiment, the Shanghai meteorological radar echo dataset is used, which contains 1545 samples and a total of 38625 image frames. Each image frame is a single-channel radar reflectivity factor map, and the spatial resolution is uniformly set to H×W=256×256 pixels after preprocessing. The input sequence is composed of radar images from the past 5 consecutive time points (6 minutes apart), forming the input tensor. Where B is the batch size (set to 16 in this embodiment). The input frame number is [number]. The corresponding output sequence (true value) is the radar image for the next 20 consecutive time points, i.e., the forecast frame number. All image data is normalized before being input into the model, scaling pixel values ​​to the [0, 1] range.

[0039] Step 2: Construct the U-Net backbone network. The encoder consists of one input module and four cascaded downsampling modules (i.e., four coding levels).

[0040] The input tensor X first passes through an initial convolutional block, which contains a 3×3 convolutional layer (stretch 1, padding 1), increasing the number of input channels. Mapped to base channel number This is followed by batch normalization (BN) and the ReLU activation function. The output feature map is denoted as x_0, and its size is... .

[0041] Furthermore, the encoder's four downsampling modules (layers 1 to 4) sequentially halve the spatial resolution and double the number of channels of the feature map. Each downsampling module's input first passes through a 2×2 max-pooling layer with a stride of 2 for downsampling. Subsequently, the downsampled feature map is fed into a feature extraction unit.

[0042] The core structure of this feature extraction unit is: Depthwise Separable Convolution (DSC) + Block Convolutional Attention (CBAM). The Depthwise Separable Convolution is used to extract features while maintaining lightweight architecture. Specifically, it is implemented by first performing a 3×3 channel-wise convolution, followed by a 1×1 pointwise convolution. Each convolutional layer is followed by Batch Normalization (BN) and ReLU activation functions.

[0043] It should be noted that the key improvement of this invention lies in the use of different types of CBAM modules in coding layers of different depths: For the downsampling modules in the first two layers (layer 1 and layer 2), the attention mechanism adopts frequency-aware convolutional attention module (CBAM). This module introduces an additional frequency-aware attention submodule on the basis of standard CBAM (channel attention + spatial attention) to enhance the representation of high-frequency details such as precipitation echo boundaries and local textures.

[0044] For the downsampling modules in the last two layers (layer 3 and layer 4), the attention mechanism adopts the standard convolutional attention module (CBAM), which only includes channel attention and spatial attention to reduce the computational complexity of deep features and focus on the extraction of semantic features.

[0045] Through the above four levels of downsampling, the encoder finally outputs a multi-scale feature map. to ,in This is a bottleneck layer feature, and its size is... (When H=W=256, the spatial dimension is reduced to 16×16).

[0046] Furthermore, the frequency-domain aware CBAM module is the core of this invention for achieving detail enhancement. For example... Figure 3 As shown, this module includes four sub-modules: a channel attention sub-module, a spatial attention sub-module, a frequency domain attention sub-module, and a feature fusion module. The specific implementation of each sub-module will be described in detail below with reference to this embodiment.

[0047] (1) Channel attention submodule: Input feature map (size is) First, global average pooling and global max pooling are used respectively to compress the spatial information of each channel into a scalar, resulting in two... The two descriptors are then fed into a shared multilayer perceptron (MLP) containing a hidden layer with C / r hidden units (r is the reduction rate, set to 16 in this embodiment). The outputs of the MLP are fused by element-wise summation, and then passed through a sigmoid activation function and a dropout operation (dropout rate set to 0.1) to generate the final channel attention weight map. This weight map is used to adaptively scale each channel of the input feature map.

[0048] (2) Spatial attention submodule: Input feature map First, the number of channels is adjusted to C' (in this embodiment, C' = 64) through a 1×1 convolutional layer, and then BN and PReLU activation are performed. Next, the feature importance in the spatial dimension is learned through two cascaded 3×3 convolutional layers (each followed by BN and PReLU). Finally, the number of channels is adjusted back to the original number of channels C through a 1×1 convolutional layer, and a spatial attention weight map is generated through the Sigmoid activation function and Dropout operation (dropout rate 0.1). .

[0049] (3)Frequency-domain attention sub-module: This is one of the core innovations of this invention. The implementation method is as follows: The input is first transformed into the frequency-domain representation through the Fast Fourier Transform (FFT), and then the high and low frequency weights are learned through the Radial Frequency Modulation mechanism to strengthen the mid and high frequency details (such as echo boundaries and local heavy precipitation centers). After the Inverse Fourier Transform (IFFT), a frequency-domain attention weight map is generated through the Sigmoid activation function.

[0050] Specifically, frequency-domain transformation: The input feature map (or the intermediate output of the CBAM module) first performs a two-dimensional real Fast Fourier Transform (2D Real FFT) to obtain a complex spectrum.

[0051] Radial frequency division: On the frequency-domain plane, the radial frequency coordinate r is defined, and its calculation formula is: where, is the frequency-domain coordinate, and r ∈ [0, 1]. The spectrum is divided into three regions according to the r value: the low-frequency region (r ≤ 0.3), the mid-frequency (0.3 < r < 0.7), and the high-frequency (r ≥ 0.7) frequency bands.

[0052] Learnable amplitude modulation: Two learnable scalar parameters are independently maintained for each feature channel: the low-frequency scaling factor and the high-frequency scaling factor . Through the Softplus activation function, it is ensured that , > 0, and the initial value is about 1. For any point in the frequency domain, its corresponding amplitude modulation weight is obtained by linear interpolation of on the radial frequency: Weighting and inverse transformation: Multiply the calculated weight matrix The original amplitude spectrum is multiplied point-by-point to obtain the modulated amplitude spectrum, while the phase spectrum remains unchanged. Finally, a two-dimensional inverse real fast Fourier transform (2D IFFT) is performed on the modulated complex spectrum to map the signal from the frequency domain back to the spatial domain, resulting in a frequency-enhanced feature map. .

[0053] Furthermore, the feature fusion module: This module is used to fuse attention information from the channel, spatial, and frequency domains. First, it fuses the channel attention weight map... Spatial attention weight map and the frequency domain modulation feature map output by the frequency domain attention submodule The feature maps are concatenated along the channel dimension. The concatenated feature maps are then passed sequentially through a 3×3 convolutional layer (compressing the number of channels back to their original value), Batch Normalization (BN), PReLU activation, and Dropout (with a dropout rate of 0.1), before being processed by a Sigmoid activation function to generate the final fused weight map. Finally, the module's output. It consists of two residuals, and its expression is as follows: in, It is the output feature map of the standard CBAM (channel + spatial attention) for input x. It is a learnable residual coefficient scalar, initialized to 0.1, used to balance the contributions of spatial and frequency domain features. This residual design facilitates gradient backpropagation and contributes to training stability. and These represent the two-dimensional real FFT and its inverse transform, respectively, which map the features from the spatial domain to the frequency domain and then restore them. According to radial frequency The constructed learnable amplitude weights are: low and high frequencies are distinguished by radial coordinates in the frequency domain plane, with each frequency corresponding to a learnable scalar. The intermediate frequency is obtained by linear interpolation of the two, and this is guaranteed by the softplus function during implementation. . For spectral amplitude, The phase factor of the original spectrum; product This means that while keeping the phase unchanged, only the amplitude is scaled, and the inverse transformation yields the spatial characteristics after frequency domain modulation. The residual coefficients are learnable scalars, initially set to small positive values ​​so that the frequency domain branch contributes to the output from the early stages of training, and the summation of residuals facilitates gradient backpropagation and is beneficial to training stability.

[0054] It should be noted that in this embodiment, a background false activation penalty is introduced in the frequency domain aware convolutional attention module, and the suppression of false alarm regions is enhanced based on a threshold guidance mechanism. The background false activation penalty is to pass the input radar precipitation map sequence through a pre-set threshold. A background pixel set is generated, and the squares of the predicted values ​​on this set are summed and normalized to generate a background penalty term. This background penalty term is then element-wise multiplied with the channel attention weight map and the spatial attention weight map to output a feature map with the same dimension as the input sequence, achieving feature suppression in the background region. By inputting the features initially extracted by depthwise separable convolutions into the CBAM module with a frequency domain attention branch at the encoder downsampling layer and decoder upsampling layer of U-Net, multi-level feature enhancement is achieved.

[0055] Specifically, background spurious activation penalty: To further suppress false alarms generated by the model in areas without precipitation, this invention introduces a background spurious activation penalty mechanism into the aforementioned attention module. Specifically, based on a pre-set threshold... (In this embodiment, the value is set to 0.1 mm / h, corresponding to extremely weak precipitation or no precipitation.) A set of background pixels is generated from the input radar precipitation map sequence. (i.e., the actual precipitation value is less than) (pixel positions). Then, the sum of squares of the model's predictions on this set is normalized to obtain a background penalty term. (Its definition is explained in the loss function section below). This background penalty term is multiplied element-wise with the channel attention weight map and the spatial attention weight map, thereby explicitly suppressing the background region in the module's output feature map, forcing the model to reduce the activation values ​​in these regions.

[0056] Step 3: Features in the bottleneck layer between the encoder and decoder The ASPP module is applied to capture multi-scale contextual features. For example... Figure 1 As shown in the bottleneck layer, the specific implementation process of the ASPP module in this embodiment is as follows: Input concatenation: To provide richer spatiotemporal information, the preliminary precipitation prediction result of the current iteration step and the corresponding actual precipitation data are concatenated with the original radar precipitation map sequence input to the encoder along the channel dimension to form an enhanced original sequence. In this embodiment, this operation only uses the actual values ​​to provide a stronger supervision signal during the training phase; during the inference phase, the actual values ​​are replaced by the model's own predictions.

[0057] Parallel Branches: The ASPP module contains four parallel feature extraction branches: A 1×1 convolutional layer branch is used to preserve local details.

[0058] Three 3×3 dilated convolution branches are used, with dilation rates of 3, 6, and 12, respectively. Additionally, a 3×3 dilated convolution branch with a dilation rate of 18 is included to capture a very large range of contextual information. Each dilated convolutional layer is followed by Batch Normalization (BN) and ReLU activation.

[0059] A global average pooling branch is used to aggregate image-level semantic information. The output of this branch is upsampled through 1×1 convolution and bilinear interpolation to restore it to the same spatial size as the input feature map.

[0060] Feature Fusion: The output feature maps of all the above branches are concatenated along the channel dimension. Then, a 1×1 convolutional layer is used to compress the channels of the fused features, adjusting the number of channels back to the same as the bottleneck layer feature x_5. Finally, to further refine the features, the compressed feature map is input again into a frequency domain-aware CBAM module (its structure is the same as described in step 2) for feature extraction and fusion. The output of this module is the final bottleneck fused feature. .

[0061] Step 4: Decoder output from ASPP module Starting from this point, upsampling operations are performed step by step. The decoder also contains four levels, symmetrical to the four levels of the encoder.

[0062] Upsampling operation: Each level of upsampling first doubles the feature map space size (×2) and halves the number of channels (÷2) by using a transposed convolutional layer (or by using bilinear interpolation followed by a 3×3 convolution).

[0063] Skip connection: Skip connection is performed between the upsampled feature map and the output feature map of the corresponding layer in the encoder (i.e., the same spatial resolution). Specifically, this is done by concatenating along the channel dimension.

[0064] Feature Refinement: The concatenated feature maps are then processed through a depthwise separable convolutional block, which consists of two depthwise separable convolutional layers (3×3 depthwise convolution + 1×1 pointwise convolution, each followed by BN and ReLU) to fuse shallow details and deep semantic information while maintaining lightweightness.

[0065] After four levels of upsampling, the spatial resolution of the feature map output by the decoder is restored to the original input size H×W.

[0066] Step 5: Extract the feature map output from the last layer of the decoder (number of channels is...). The number of channels is mapped to the number of prediction frames through a 1×1 output convolutional layer. This 1×1 convolutional layer does not change the spatial resolution, and its output is the final multi-frame prediction field. This enables end-to-end mapping from input radar sequences to output forecast sequences.

[0067] Step 6: Train the FCA-UNet model using an end-to-end approach. In this embodiment, the Adam optimizer (β1=0.5, β2=0.999) is used, with an initial learning rate of 1e-3 and a cosine annealing scheduling strategy for decay. The weight decay coefficient is set to 1e-2. The total training epochs are set to 200, and early stopping is used, stopping training when the loss on the validation set shows no improvement for 20 consecutive epochs.

[0068] This invention employs a hybrid loss function To optimize the model parameters, the loss function is obtained by weighted summation of four parts: The specific calculation methods for each loss term and the hyperparameter settings used in this embodiment are as follows: Weighted Mean Squared Error Loss : This loss is designed to enhance the model's regression accuracy in areas of heavy precipitation and to address the sample imbalance problem in precipitation data.

[0069] Where B=16 is the batch size. =20 is the number of forecast frames. and These are the predicted field and the actual field of the b-th sample in frame c, respectively. It is a binary mask for heavy precipitation, when The value is 1 if the rainfall is ≥ the threshold (0.5 mm / h in this embodiment), otherwise it is 0. β is the heavy rainfall weighting factor, which is set to 1.2 in this embodiment. The operator ⊙ indicates element-wise multiplication. Denominator The weighting factor is the mean of the weighting factor across the entire batch and spatial dimensions, used for normalization and stabilizing the loss scale. The frame weight of frame c is used to apply differentiated attention to different frames in the forecast sequence: In this embodiment, the first 12 frames (short-term proximity) are given a higher weight of 2.3, the 8th to 20th frames (longer time-term) are given a weight of 1.55, and the remaining frames are given a weight of 1.0.

[0070] Structural Similarity Loss (SSIM Loss) : This loss is used to measure the similarity between the predicted field and the real field in terms of brightness, contrast, and structure, guiding the model to generate prediction results with higher visual quality.

[0071] Wherein, SSIM(·) is the structural similarity index calculation function, and Δ is the dynamic range of pixel values ​​(in this embodiment, the dynamic range of the normalized image is [0, 1], so Δ=1). Use the same frame weights as in L_MSE.

[0072] Background False Activation Penalty Loss : This loss directly penalizes the model for producing false predictions in the background region (the actual area with no precipitation).

[0073] in, To be based on the real scene and preset threshold (In this embodiment, the set of background pixels is set to 0.1 mm / h, i.e.) < The location. This represents the total number of pixels in the set. This loss forces the model to predict values ​​close to 0 in the background region.

[0074] Fourier Loss : This loss constrains the consistency between the predicted and actual fields in the frequency domain, ensuring that the model can learn the correct global and local frequency features, especially for high-frequency information of precipitation boundaries.

[0075] The method of first adding 1 to the amplitude spectrum, then taking the logarithm, and finally calculating the L1 norm can effectively measure the difference between the two spectra and suppress the influence of abnormally large values. This is also a frame weight.

[0076] In this embodiment, the weight hyperparameters of each loss term in the hybrid loss function are set as follows: , , , .

[0077] Four hyperparameters control the weighted mean square error and structural similarity, respectively. The synergistic effect of each loss term is as follows: the weighted mean square error (MSE) ensures the regression accuracy of precipitation intensity and focuses on constraining the numerical deviation in areas of heavy precipitation; the SSIM loss measures the similarity between the forecast field and the actual field at the structural level, compensating for the neglect of image structure by MSE; the background false activation penalty specifically suppresses false predictions in areas without precipitation, reducing the false alarm rate in operational forecasts; and the Fourier loss constrains the consistency of high and low frequency components in the frequency domain, echoing the shallow frequency domain sensing CBAM of the encoder and strengthening the characterization of echo boundaries and details.

[0078] In this embodiment, the training parameters are as follows: the maximum training epoch is set to 200. To improve training efficiency and prevent overfitting, training is terminated when the model's performance on the validation set shows no improvement for 20 consecutive epochs. To prevent overfitting of the loss curve, this embodiment sets the model training frequency to once every two epochs.

[0079] Further adjustments are made to the learning rate scheduler, each monitoring the corresponding loss on the validation set. The optimization process employs the Adam algorithm, where... , The weight decay coefficient is set to Specifically, the regional weighting coefficient w was experimentally determined to be 1.2, the weight of the first frame (the first 12 frames) was 2.3, and the weight of the far frames (starting from the 8th frame) was 1.55.

[0080] Training was conducted based on the above experimental configuration, and each model was evaluated on a test set using the optimal parameters. This section first compares the model with the selected benchmark models ConvGRU, SmaAt-UNet, SimVP, and PhyDnet. Comparative experiments were performed on the Shanghai dataset, using multi-scale CSI, HSS, LPIPS, and SSIM as benchmarks.

[0081] Table 1 lists the mean squared error (MSE), F1 score, critical success index (CSI), Heidke skill score (HSS), and Matthews correlation coefficient (MCC) for all comparison methods at different thresholds. To compare the performance of each model under different precipitation thresholds, hourly precipitation maps were generated by overlaying the first 12 frames of each model's predictions to calculate each index. In terms of overall MSE, this invention outperforms several comparison methods at thresholds of ≥10 mm / h and ≥20 mm / h.

[0082] Table 1 shows the experimental results of the model on the Shanghai radar dataset. ↑ indicates that the higher the index, the better; ↓ indicates that the lower the index, the better.

[0083] The proposed FCA-UNet model demonstrates superior performance on the Critical Success Index (CSI) and Heidke Skill Score (HSS) metrics. While FCA-UNet's scores on CSI, CSI-pool4, CSI-pool16, and HSS are slightly lower than the model described in Chapter 3, the differences are controlled within 1%, and it significantly outperforms benchmark methods such as ConvGRU, SimVP, SmaAt-UNet, and PhyDnet. Specifically, CSI-pool4 and CSI-pool16 calculate CSI by pooling the forecast field at fine-grained and coarse-grained spatial scales, respectively, thus evaluating the model's hit rate performance across multiple scales. FCA-UNet ranked highly in both pooling scales, demonstrating its robustness and stability in fine- and meso-scale binary classification prediction tasks.

[0084] FCA-UNet achieved state-of-the-art results in both the Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). Its SSIM score surpasses ConvGRU and all other comparable models; simultaneously, it performs exceptionally well on the LPIPS metric, indicating that the generated predicted field more closely resembles the observed real field in terms of structural integrity and visual perception quality. Overall, FCA-UNet achieves comparable classification accuracy to other models while demonstrating significant advantages in structural fidelity and perceptual quality. This is attributed to the enhancement of details and boundary features by the frequency-domain perceptual CBAM, the effective fusion of multi-scale contextual information by ASPP, and the meticulous design of SSIM and Fourier domain constraints in the loss function.

[0085] Table 2 Ablation experiments with region weighting coefficients

[0086] In terms of prediction consistency and image quality assessment, the Heidke Skill Score (HSS), Perceptual Patch Similarity Index (LPIPS), and Structural Similarity Index (SSIM) quantify the reliability and fidelity of the model's generated results from multiple dimensions. The HSS value of the full FCA-UNet model is significantly higher than that of the three variant models: FCA-UNet with the frequency domain attention CBAM and ASPP modules removed, FCA-UNet with only the CBAM module removed, and FCA-UNet with only the ASPP module removed. Compared with the full model, the HSS values ​​of these variants are reduced by more than 12%, indicating that the synergistic effect of the frequency domain attention CBAM, ASPP module, and the designed composite loss function can significantly improve the stability and consistency of prediction results and effectively suppress bias propagation.

[0087] Similarly, the full FCA-UNet outperforms the variant models on both LPIPS and SSIM metrics, demonstrating varying degrees of performance improvement. This result highlights the unique contribution of the frequency-domain attention CBAM module in reducing perceptual discrepancies and enhancing detail preservation, as well as the crucial role of the ASPP module in multi-scale context aggregation.

[0088] Comprehensive analysis shows that all three variant models underperform the complete FCA-UNet across all metrics, with the variant that simultaneously removes both the CBAM and ASPP modules experiencing the most significant performance degradation. This is primarily due to the lack of frequency domain attention, which prevents the model from explicitly enhancing the mid-to-high frequency features of echo boundaries, thus weakening its ability to characterize details. Simultaneously, the absence of ASPP limits the model's receptive field, hindering its ability to effectively capture the multi-scale spatiotemporal evolution of precipitation systems, thereby amplifying mesoscale forecast bias and ultimately leading to a dual degradation in core metrics and visual quality. By integrating the frequency domain attention mechanism, the ASPP module, and Fourier domain loss constraints, our proposed method not only enhances the model's core prediction accuracy and multi-scale feature representation capabilities but also significantly improves result consistency and image generation quality, fully validating the rationality and effectiveness of the proposed architecture design.

[0089] To verify the effectiveness of the proposed FCA module in enhancing attention to areas of heavy precipitation, this embodiment uses Grad-CAM to generate activation heatmaps of the module. Grad-CAM (Gradient-weighted Class Activation Mapping) is a visualization technique used to understand the decision-making process of convolutional neural networks. Its calculation involves two steps: calculating the gradient weights of the target class, as shown in the following formula: in, Z represents the importance weight of the k-th feature map to category c, and Z is the spatial size of the feature map. The model outputs a score for category c. The first step represents the activation value of the k-th feature map at position (i,j). The second step involves weighted summation of the generated gradient weights, followed by ReLU filtering to obtain the final result. This method uses a threshold of 0.5 mm / h to binarize the output, transforming the prediction task into a classification task to generate activation heatmaps. This embodiment generates activation heatmaps for the FCA module and the traditional CBAM module in the encoder based on the input sequence. In the activation heatmaps, a redder color indicates a higher activation level, while a bluer color indicates a lower activation level.

[0090] It can be seen that when the depth is 0, FCA's region of interest is relatively concentrated and localized, responding significantly to areas of heavy precipitation in the input, making it suitable for capturing early, fine patterns. CBAM, on the other hand, has a more dispersed region of interest with weaker contours, exhibiting slightly stronger global perception and broader early feature coverage but lower accuracy. When the depth is 1, the resolution decreases, and compared to CBAM, FCA's highlighted areas expand and its contours become clearer. As the resolution continues to decrease, FCA demonstrates better feature focusing and semantic aggregation capabilities, resulting in more accurate pattern capture of sequence data.

[0091] In summary, under the same input and observed values, the proposed FCA-UNet model effectively maintains the morphology and boundary integrity of heavy precipitation areas in short, medium, and long-term forecasts, exhibiting superior structural coherence and visual plausibility, especially in the long-term stage. This performance is highly consistent with the leading results of SSIM and LPIPS in quantitative metrics, further validating the key role of the frequency-domain sensing CBAM and ASPP modules in enhancing the structural fidelity and sensing quality of the forecast field.

[0092] Furthermore, this chapter evaluated the model's parameter count, single-sequence inference latency, total computational cost (FLOPs), and single-frame computational cost, with results shown in Table 3. The parameter count directly reflects the model size and memory usage; inference latency determines the real-time deployment capability in practical business scenarios; and FLOPs quantify the theoretical computational complexity. The following sections provide an in-depth analysis from three dimensions: the correspondence between parameter count and computational cost, inference latency analysis, and the positioning of FCA-UNet in the efficiency-performance tradeoff.

[0093] Table 3. Parameter count, inference latency of a single sequence, computational cost, and computational cost per frame for each model.

[0094] In terms of parameter size, FCA-UNet has the largest size at 9.86 M, followed by SmaAt-UNet (6.03 M), while ConvGRU (1.22 M), SimVP (0.35 M), and PhyDnet (2.47 M) are relatively smaller. This is mainly due to the fact that ConvGRU and PhyDnet reuse parameters in the time dimension through recursion or physical constraint mechanisms, and SimVP uses a lightweight backbone. FCA-UNet, on the other hand, is based on the U-Net architecture and incorporates layer-by-layer CBAM (including shallow frequency domain branches) and the bottleneck ASPP module, resulting in parameters mainly composed of convolution and attention weights, which is relatively large. However, this increase is an intentional trade-off to enhance structural fidelity and multi-scale frequency domain modeling.

[0095] Computational cost (FLOPs) and inference latency do not simply correspond to the number of parameters: ConvGRU, despite its small number of parameters, achieves 247.09 G FLOPs and a latency of 30.59 ms per sequence, limited by recursive unrolling and sequence dependencies that restrict parallelism; PhyDnet has the highest FLOPs and latency, related to its iterative physical modules; SimVP and SmaAt-UNet achieve lower overhead through single forward passes. Although FCA-UNet has a large number of parameters, it employs depthwise separable convolutions and non-recursive forward passes, achieving FLOPs per sequence / frame comparable to SmaAt-UNet, with a latency of only 6.13 ms, significantly better than ConvGRU / PhyDnet, on par with SimVP / TSPF-GAN, and only slightly higher than SmaAt-UNet, thus effectively controlling efficiency amplification.

[0096] In summary, FCA-UNet maintains efficient computation and latency levels despite its relatively large scale, comparable to or better than most models, demonstrating a good trade-off between efficiency and performance and validating the rationality of its design.

[0097] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid, characterized in that, Includes the following steps: Step 1: Obtain radar precipitation dataset and preprocess it to obtain radar precipitation map sequence and actual precipitation data values; Step 2: Construct the U-Net backbone network. The U-Net adopts an encoder-decoder configuration. The encoder includes one input module and four downsampling modules. The input of the downsampling modules in the first two layers adopts frequency-aware convolutional attention modules, while the input and output of the downsampling modules in the last two layers adopt standard convolutional attention modules. Step 3: Apply the hollow spatial pyramid pooling module to the bottleneck layer features between the encoder and decoder, capture multi-scale contextual features through parallel branches, and output fused features; Step 4: The decoder upsamples from the fused features level by level. At each level, the spatial size is doubled and the channels are halved through transposed convolution. The features are then spliced ​​with the corresponding layer features of the encoder by skip connection. Finally, depthwise separable convolution is used for processing. Step 5: The output features of the last layer of the decoder are mapped to a multi-frame prediction field through a 1×1 output convolutional layer, realizing end-to-end mapping from sequence to sequence; Step 6: Train the model end-to-end, optimize the parameters using a hybrid loss function, and generate short-term precipitation forecast results.

2. The lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 1, characterized in that, The frequency domain-aware convolutional attention module includes a channel attention submodule, a spatial attention submodule, a frequency domain attention submodule, and a feature fusion module.

3. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 2, characterized in that, The frequency domain attention submodule is implemented as follows: the input radar precipitation map sequence is first subjected to a two-dimensional real fast Fourier transform to obtain a complex spectrum and separate the amplitude and phase; low-frequency, mid-frequency and high-frequency components are distinguished on the frequency domain plane according to the radial frequency coordinate, and the amplitude scaling factor of low-frequency and high-frequency components is learned through a learnable scalar; the learned weights are obtained by linear interpolation of low-frequency and high-frequency scaling on the radial frequency, and the frequency domain attention weight matrix is ​​generated through an activation function.

4. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 3, characterized in that, The spatial attention submodule is implemented as follows: the input radar precipitation map sequence is processed by adjusting the number of channels through a 1×1 convolutional layer, followed by batch normalization and activation functions; then, multiple 3×3 convolutional layers are passed, with each convolutional layer followed by batch normalization and activation functions, to learn features in the spatial dimension; finally, the number of channels is adjusted by a 1×1 convolutional layer and expanded back to the original spatial dimension, and a spatial attention weight map is generated by using a sigmoid activation function and dropout operation.

5. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 4, characterized in that, Background false activation penalty is introduced in the frequency domain aware convolutional attention module to enhance the suppression of false alarm regions based on a threshold guidance mechanism. The background false activation penalty is generated by judging the input radar precipitation map sequence through a pre-set threshold to generate a background pixel set, summing the squares of the predicted values ​​on this set and normalizing them to generate a background penalty term. The background penalty term is multiplied element-wise with the channel attention weight map and the spatial attention weight map to output a feature map with the same dimension as the input sequence, and feature suppression is achieved in the background region.

6. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 5, characterized in that, The implementation method of the channel attention submodule is as follows: inputting a radar precipitation map sequence, compressing each channel into a scalar through global average pooling and max pooling, and then performing batch normalization and activation functions; next, adjusting the number of channels through a shared multilayer perceptron, and generating a channel attention weight map through a sigmoid activation function and dropout operation.

7. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 6, characterized in that, The feature fusion module is implemented as follows: the channel attention weight map, spatial attention weight map, and frequency domain attention weight map are concatenated along the channel dimension. The concatenated sequence is then passed sequentially through a 3×3 convolutional layer, batch normalization, activation function, and dropout operation. A fusion weight is then generated using a Sigmoid activation function. Finally, the fusion weight is multiplied element-wise with the input feature map to generate the final output. The final output is equal to the feature map output by the convolutional attention module on the input feature map, plus a learnable residual coefficient multiplied by the frequency domain modulation feature map. The frequency domain modulation feature map is generated as follows: a two-dimensional real fast Fourier transform is performed on the feature map output by the convolutional attention module to obtain a complex spectrum. Its amplitude and phase spectra are extracted. A learnable amplitude weight matrix is ​​constructed on the frequency plane according to radial frequencies. This weight matrix is ​​multiplied point-by-point with the original amplitude spectrum while keeping the phase spectrum unchanged. Finally, a two-dimensional real fast Fourier inverse transform is performed on the weighted complex spectrum to obtain the frequency-modulated spatial feature map.

8. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 1, characterized in that, The hollow spatial pyramid pooling module introduces multi-dilation rate dilated convolution. The implementation method is as follows: the preliminary precipitation prediction result and the actual value are respectively concatenated with the radar precipitation map sequence input to U-Net to form the original sequence. The hollow spatial pyramid pooling module extracts multi-scale features of the original sequence through 3×3 dilated convolution branches with dilation rates of 3, 6, 12 and 18, generates multi-branch output, and performs channel-level fusion with the global average pooling branch. The fused output sequence is channel-compressed using 1×1 convolution. The frequency domain aware convolutional attention module is used for feature extraction and fusion, and finally the bottleneck feature is output.

9. A lightweight precipitation nowcasting method based on frequency domain sensing CBAM and void space pyramid as described in claim 1, characterized in that, The hybrid loss function is obtained by weighted summation of weighted mean squared error loss, structural similarity loss, background spurious activation penalty loss, and Fourier loss; where... The weighted mean square error loss is calculated as follows: First, for each sample and each forecast frame in each batch, the square of the pixel-by-pixel difference between the predicted field and the actual field is calculated. This squared value is then multiplied by a factor that includes a heavy precipitation weighting factor, which is 1 plus the product of the heavy precipitation weighting factor and the heavy precipitation region mask. The heavy precipitation region mask is set to 1 when the pixel value in the actual field reaches a preset threshold and otherwise to 0. Then, the weighted mean square error is divided by the mean of the heavy precipitation weighting factor in both the batch and spatial dimensions for normalization. Finally, the normalized result is multiplied by a preset frame weight in each frame and averaged in both the batch and frame dimensions. The structural similarity loss is calculated as follows: for each prediction frame, the structural similarity index between the predicted field and the real field is calculated, and the single-frame structural similarity loss is obtained by subtracting the index from 1. Then, it is multiplied by the frame weight corresponding to that frame, and finally, the average is calculated over all frames. The calculation method for the background false activation penalty loss is as follows: a background pixel set is pre-defined based on the real field, the sum of squares of the pixel values ​​in the prediction field that belong to the background set is calculated, the sum is divided by the total number of background pixels for normalization, and then the average is calculated in the batch and frame dimensions. The Fourier loss is calculated as follows: perform a two-dimensional real fast Fourier transform on the predicted field and the true field of each forecast frame, take the amplitude spectrum, add 1 to the amplitude spectrum and take the logarithm, calculate the absolute value of the difference between the two logarithmic amplitude spectra, multiply by the frame weight corresponding to the frame, and finally calculate the average over the batch and frame dimensions.

Images