Mountain area extreme rainstorm flood forecasting method based on MUL-U2-NET network
By constructing a deep learning model based on the MUL-U2-NET network, the shortcomings of mountain flood forecasting models in terms of accuracy and timeliness were addressed, achieving more accurate flood forecasts, enhancing the model's robustness and generalization ability, and improving the accuracy and timeliness of forecasts.
Patent Information
- Application Number
- CN202511240115.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2025-12-16
AI Technical Summary
Existing flood forecasting models for mountainous areas are inadequate in terms of accuracy and timeliness. In particular, under complex terrain conditions, the accuracy of precipitation forecast data is insufficient and the spatiotemporal resolution is limited. Rainfall stations are unevenly distributed, traditional hydrological models do not adequately describe runoff generation and confluence mechanisms, underlying surface data is not updated in a timely manner, and the coupled models are incomplete, all of which affect the accuracy and timeliness of forecasts.
We adopt a method based on the MUL-U2-NET network to construct a deep learning model that integrates multi-source information, including an input layer, an encoding module, a residual segmentation module, and a decoding module. We use convolutional neural networks (CNN), dense connections (DC), and enhanced attention modules (CBAM) to extract multi-scale features and introduce a non-equal weight loss function for training to optimize model performance.
It improves the accuracy and timeliness of rainstorm and flood forecasts in mountainous areas, can more accurately reflect the runoff generation and confluence mechanisms under complex terrain, provides more comprehensive scientific basis, and provides a guarantee for flood control, disaster reduction and water resource management.
Smart Images
Figure CN121145121A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for forecasting regional extreme rainstorms and floods, and in particular to a method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network. Background Technology
[0002] For flood forecasting in mountainous areas, incorporating high-precision information on the underlying surface, including topography, soil conditions, and human activities, can significantly improve the accuracy and practicality of forecasts. Mountainous terrain is complex and undulating, significantly impacting airflow and precipitation distribution. By integrating high-resolution topographic data, the distribution of mountains and valleys and their guiding role in precipitation flow can be clearly captured, thus accurately predicting the convergence path and velocity of rainwater in mountainous areas. Simultaneously, topography also affects solar radiation and temperature distribution, indirectly influencing water replenishment processes such as snowmelt, which is crucial for flood forecasting. Soil types vary greatly in mountainous areas, with significant differences in infiltration and water-holding capacity. Considering soil information allows for a more accurate assessment of the proportion of precipitation converted into surface runoff and the runoff generation after soil saturation, providing key parameters for flood forecasting. Human activities such as land reclamation and engineering construction also profoundly alter the underlying surface in mountainous areas. These activities can damage vegetation, increase soil erosion, and exacerbate flood risks. Therefore, integrating human activity information helps to more comprehensively assess the potential threat of floods in mountainous areas.
[0003] However, existing precipitation forecast data suffers from insufficient accuracy and limited spatiotemporal resolution, especially under complex terrain conditions, where uneven distribution of rainfall stations leads to significant forecast bias. Updates to underlying surface data are often untimely and inaccurate, with delays in the release of topographic, soil, and land use datasets, hindering timely input into corrective models. Furthermore, traditional hydrological models fail to adequately describe runoff generation and distribution mechanisms in mountainous areas, coupled with incomplete models and limitations in the application of remote sensing technology (such as inversion errors caused by cloud cover). These shortcomings severely impact the accuracy and timeliness of flood forecasts. A mountain flood forecasting model incorporating real-time imagery of the underlying surface, including topography, soil, and human activities, can more precisely reflect the complex runoff generation and distribution mechanisms in mountainous regions, improving the accuracy and timeliness of flood forecasts. This provides a scientific basis for flood control and disaster reduction in mountainous areas and water resource management, effectively safeguarding people's lives and property. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides a method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network.
[0005] To achieve the above objectives, the present invention employs the following method: a method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network, comprising the following steps: Step S1: Acquire and preprocess multi-source data, including hydrological element observation data and image datasets of mountain flood forecasting factors. The image datasets shall at least include extreme precipitation distribution, soil moisture content distribution, geographic elevation distribution, land cover type distribution, water conservancy project type distribution, and nighttime light distribution data. Step S2: Construct a MUL-U2-NET network model that integrates multi-source information. The model includes an input layer, an encoding module, a residual segmentation module, and a decoding module, wherein: The encoding module uses a convolutional neural network (CNN) to extract multi-scale features and introduces dense connections (DC) and a CBAM (Concentration-Enhanced Attention) module. The residual segmentation module extracts deep features through a residual structure with densely connected DCs; The decoding module restores spatial resolution through upsampling operations and extracts image features using a convolutional neural network (CNN) and densely connected networks (DC). Step S3: Train the model using a non-equal-weight loss function; Step S4: Input real-time data into the trained model for flood forecasting.
[0006] As a preferred embodiment of the present invention, the data preprocessing in step S1 specifically includes: Convert hydrological data into image data. Resize and standardize images from different data sources to a uniform size:
[0007] in, The raw pixels of the image data, and They are respectively The mean and standard deviation, The data is standardized. The normalized multi-source image dataset is divided into a training set and a validation set.
[0008] As a preferred embodiment of the present invention, the encoding module uses a convolutional neural network (CNN) to extract multi-scale features, and the specific method for introducing dense connections (DC) and enhanced attention module (CBAM) is as follows: Convolutional Neural Networks (CNNs) are used to extract feature information from spatial fields at different scales of the model input features, thereby extracting multi-scale data source features. Activation functions are applied to introduce nonlinearity, and pooling layers are used to reduce spatial resolution and extract global information through convolutional features. Extract the following: ; Densely connected DCs are applied to convolutional layers, connecting the output of each layer to the input of all subsequent layers, specifically in the following form:
[0009] in, This is the output of the (n+1)th layer. The output of all previous layers is the input of layer n+1; In the input part, the Enhanced Attention Module (CBAM) is applied to calculate channel attention features and spatial attention mechanisms, thereby strengthening the model's feature learning of the image. Specifically, the following is the form: CAM
[0010] SAM
[0011] in, is the input of the nth layer, CAM is the channel attention mechanism, and SAM is the spatial attention mechanism.
[0012] As a preferred embodiment of the present invention, the specific method for the residual segmentation module to extract deep features through the residual structure with DC is as follows: By utilizing a residual segmentation module with DC to extract multi-scale deep features from the encoded portion, the convergence speed is accelerated and the model's versatility is enhanced. Specifically:
[0013] in For the first One residual convolutional layer, This represents the convolution operation. This represents the modified linear unit activation function; where, in the first residual convolutional layer, Take as input features for the model In other residual convolutional layers, This is the output of the previous pooling layer; Pooling layer:
[0014] in, For the first The output of each residual convolutional layer; To perform global average pooling on the input feature map; For the first The output of each pooling layer; Fully connected layer: calculated as follows:
[0015] in, Indicates a fully connected operation; This is the output of the last pooling layer.
[0016] As a preferred embodiment of the present invention, the upsampling operation restores the spatial resolution and extracts image features using CNN and DC, specifically in the following form:
[0017] in This represents the convolution operation. This is the output of the (n+1)th layer. This is the input layer.
[0018] As a preferred embodiment of the present invention, the non-equal weight loss function in the non-equal weight loss function training model in step S3 is:
[0019]
[0020]
[0021] in, It is a non-equal-weighted multidimensional loss function; , These are the weighting coefficients for the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR), respectively. , m The number of rain gauges participating in RMSE calculation within the model training area; The output of the model during the training process; These are the water level observation values from the hydrological station; , This represents the number of grid points within the model training region that participate in PSNR calculation; These are the pixel values corresponding to the model output during the training process. These are the pixel values corresponding to the water level observations at the hydrological station. The maximum pixel value in the model output image and the flood forecast process line observation image during the training process.
[0022] As a preferred embodiment of the present invention, step S3, which involves training the model using a non-equal-weight loss function, further includes: in the validation set, using gradient descent, adjusting the model parameters along the opposite direction of the gradient to iterate until the loss function is minimized, thereby achieving optimal model performance. The parameter update method is as follows:
[0023] in, Indicates specific model parameters, This is the learning rate.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Based on an advanced deep learning scheme, it makes full use of multi-source data that is highly correlated with rainstorm and flood forecasting, fully considers the multi-scale characteristics of rainstorms and complex underlying surface information, establishes a multi-source image dataset for rainstorm and flood in mountainous areas, effectively extracts the refined distribution characteristics of runoff in mountainous areas, incorporates the impact of human activities, and thus obtains a more accurate rainstorm and flood forecast for mountainous areas. Compared with the forecast results of traditional hydrological models, it has more comprehensive data coverage, more thorough data mining, and the flood forecast obtained is more targeted to complex terrain and more objective. 2. After extracting the main spatial modal features based on CNN, considering the discontinuity, suddenness and complexity of the underlying surface information of the rainfall field, the residual network and U2-Net network are fused to extract, process and fuse features at different scales. This improves the global perception of the data and extracts more comprehensive and complex underlying surface spatial features, fully explores the data, reduces the risk of overfitting caused by extreme values, and helps to improve model performance and generalization ability. 3. During the training of the U2-Net model, the enhanced convolutional block attention module (CBAM) and densely connected image segmentation algorithm (DC) are introduced. These can adaptively optimize features by weighting, make full use of global information, enhance the model's expressive power, improve robustness and generalization ability, and accelerate training convergence to achieve efficient and accurate image segmentation. 4. Introduce a new loss function. Based on the evolution and peak characteristics of flood forecasts, use comprehensive evaluation indicators such as root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) to construct a non-equal weighted loss function, making the model more accurate and thus improving the accuracy of mountain torrential rain and flood forecasts. Attached Figure Description
[0025] Figure 1 This is a flowchart of the method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network, as described in this invention. Detailed Implementation
[0026] The following is in conjunction with the appendix Figure 1 The technical solution of the present invention will be further described below.
[0027] The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network provided by this invention has the following process: Figure 1 As shown, the specific steps include: Step S1: Obtain hydrological element data, including observation sequences of hydrological element data such as water level and flow rate at the outlet section of the target watershed; and image datasets of mountain flood forecasting factors, including precipitation, soil moisture content, geospatial data, and human activities, such as spatial distribution image data of extreme rainfall, soil moisture content, geographic elevation, land cover, water conservancy projects, and nighttime light. First, convert the hydrological element data into image data, and adjust images from different data sources to a uniform size and perform standardization processing to ensure that the data range input to the model is consistent and the features are distinct, which is beneficial for model training and convergence. The standardization preprocessing formula is as follows:
[0028] in, The raw pixels of the image data. and They are respectively The mean and standard deviation, The data is standardized. The normalized multi-source image dataset is divided into a training set and a validation set.
[0029] Step S2: For forecasting extreme rainstorms and floods in mountainous areas, a U2-Net network model for forecasting extreme rainstorms and floods in mountainous areas is constructed by fusing real-time hydrological data with multi-source information such as the distribution of extreme precipitation, soil moisture content, geographical elevation, land cover type, water conservancy project type, and nighttime light in the target area. This model mainly includes an input layer, an encoding module, a residual segmentation module, and a decoding module.
[0030] The input layer data includes the multi-source image dataset from step S1. In the encoding module, a convolutional neural network (CNN) is first used to process the multi-source image dataset. Feature information is extracted and fused, and pooling layers are used to reduce spatial resolution and extract global information. Dense connections (DCs) are introduced between two convolutional blocks to ensure the transmission of effective information without redundancy. A CBAM (Concentration-Enhanced Attention) module is applied to the input part of the DC to enhance the weighting of important channels and spatial features, reduce the influence of irrelevant features, and thus better capture multi-scale features in the image. A residual segmentation module with DC is then used to extract deep features from the encoded part. In the decoding module, spatial resolution is restored through upsampling, and convolution and DC are used to refine the feature field of the decoding module. CBAM is applied to the output of each layer. Finally, a final convolution is applied to obtain the final output, which is the evolution line of the extreme rainstorm and flood forecast process.
[0031] Step 2.1 Input Layer: The model receives the multi-source image dataset obtained in step S1. and hydrological station observation data images The input features are integrated at the input layer to form the model input features. .
[0032] Step 2.2 Encoding Module: Use a Convolutional Neural Network (CNN) to encode the input features of the model. Feature information from spatial fields at different scales is extracted, thereby extracting multi-scale data source features. Activation functions are applied to introduce nonlinearity, and pooling layers are used to reduce spatial resolution and extract global information. (Convolutional features) Extract the following:
[0033] Dense connections (DC) are applied to convolutional layers, connecting the output of each layer to the input of all subsequent layers. This enhances feature reuse and gradient flow, thereby improving the model's generalization ability. Specifically, it takes the following form:
[0034] in, This is the output of the (n+1)th layer. This is the output of all previous layers, i.e., the input of layer n+1.
[0035] In the input part, a CBAM (Continuous Enhanced Attention Module) is applied to compute channel attention features and spatial attention mechanisms, thereby enhancing the model's feature learning of the image. Specifically, this takes the following form: CAM
[0036] SAM
[0037] in, is the input of the nth layer, CAM is the channel attention mechanism, and SAM is the spatial attention mechanism.
[0038] Step 2.3 Residual Segmentation Module: A residual segmentation module with DC is used to extract multi-scale deep features from the encoded portion, accelerating convergence and enhancing the model's versatility. Its specific form is as follows:
[0039] in For the first One residual convolutional layer, This represents the convolution operation. This represents the modified linear unit activation function; where, in the first residual convolutional layer, Take as input features for the model In other residual convolutional layers, This is the output of the previous pooling layer; Pooling layer:
[0040] in, For the first The output of each residual convolutional layer; To perform global average pooling on the input feature map; For the first The output of each pooling layer; Fully connected layer: calculated as follows:
[0041] in, Indicates a fully connected operation; This is the output of the last pooling layer; Step 2.4 Decoding Module: This module restores spatial resolution through upsampling and extracts image features using CNN and DC. Specifically, it takes the following form:
[0042] in This represents the convolution operation. This is the output of the (n+1)th layer. This is the input layer.
[0043] Step S3: Input the image dataset from Step S1 into the model from Step S2, and train and optimize the model using the training and validation sets based on the non-equal weight loss function. Specifically, this includes: Step 3.1 Based on the evolution characteristics of the flood forecast hydrograph and fully considering the instantaneous characteristics of the peak value, an innovative loss function, namely the non-equal-weight loss function, is introduced. This loss function integrates two evaluation metrics: Root Mean Square Error (RMSE) and Peak Signal-to-Noise Ratio (PSNR), comprehensively measuring the model's performance by combining the evolution of the flood hydrograph and the peak characteristics. The non-equal-weight loss function is constructed as follows:
[0044]
[0045]
[0046] in, It is a non-equal-weighted multidimensional loss function; , These are the weighting coefficients for the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) indicators, respectively. ,m The number of rain gauges participating in RMSE calculation within the model training area; The output of the model during the training process; These are the water level observation values from the hydrological station; , This represents the number of grid points within the model training region that participate in PSNR calculation; These are the pixel values corresponding to the model output during the training process. These are the pixel values corresponding to the water level observations at the hydrological station. The maximum pixel value in the model output image and the flood forecast process line observation image during the training process.
[0047] By adjusting the weighting coefficients, the influence of various indicators during the training process can be balanced, allowing the model to better adapt to the characteristics of the precipitation field and obtain a more accurate flood forecast process line.
[0048] Step 3.2 uses the non-equal weight loss function obtained in Step 3.1 to train the MUL-U2-NET model that fuses multi-source image data. To avoid overfitting, in the validation set, gradient descent is used to adjust the model parameters along the opposite direction of the gradient, iterating continuously to minimize the loss function and achieve optimal model performance. The parameter update method is as follows:
[0049] in, Indicates specific model parameters, This is the learning rate.
[0050] Step S4: Extreme Rainfall and Flood Forecasting: Substitute the latest hydrological data and image data (same as step S1) into the optimized MUL-U2-NET model from step S3 to perform flood forecasting and output the flood forecast process line.
Claims
1. A method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network, characterized in that, Includes the following steps: Step S1: Acquire and preprocess multi-source data, including hydrological element observation data and image datasets of mountain flood forecasting factors. The image datasets shall at least include extreme precipitation distribution, soil moisture content distribution, geographic elevation distribution, land cover type distribution, water conservancy project type distribution, and nighttime light distribution data. Step S2: Construct a MUL-U2-NET network model that integrates multi-source information. The model includes an input layer, an encoding module, a residual segmentation module, and a decoding module, wherein: The encoding module uses a convolutional neural network (CNN) to extract multi-scale features and introduces dense connections (DC) and a CBAM (Concentration-Enhanced Attention) module. The residual segmentation module extracts deep features through a residual structure with densely connected DCs; The decoding module restores spatial resolution through upsampling operations and extracts image features using a convolutional neural network (CNN) and densely connected networks (DC). Step S3: Train the model using a non-equal-weight loss function; Step S4: Input real-time data into the trained model for flood forecasting.
2. The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network according to claim 1, characterized in that, The data preprocessing in step S1 specifically includes: Convert hydrological data into image data. Resize and standardize images from different data sources to a uniform size: middle, The raw pixels of the image data, and They are respectively The mean and standard deviation, The standardized, normalized multi-source image dataset is divided into a training set and a validation set.
3. The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network according to claim 1, characterized in that, The encoding module uses a convolutional neural network (CNN) to extract multi-scale features, and the specific method of introducing dense connections (DC) and enhanced attention module (CBAM) is as follows: Convolutional Neural Networks (CNNs) are used to extract feature information from spatial fields at different scales of the model input features, thereby extracting multi-scale data source features. Activation functions are applied to introduce nonlinearity, and pooling layers are used to reduce spatial resolution and extract global information through convolutional features. Extract as follows: ； Densely connected DCs are applied to convolutional layers, connecting the output of each layer to the input of all subsequent layers, specifically in the following form: ,in, This is the output of the (n+1)th layer. The output of all previous layers is the input of layer n+1; In the input part, the Enhanced Attention Module (CBAM) is applied to calculate channel attention features and spatial attention mechanisms, thereby strengthening the model's feature learning of the image. Specifically, the following is the form: CAM ; SAM in, is the input of the nth layer, CAM is the channel attention mechanism, and SAM is the spatial attention mechanism.
4. The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network according to claim 1, characterized in that, The specific method by which the residual segmentation module extracts deep features through the residual structure with DC is as follows: By utilizing a residual segmentation module with DC to extract multi-scale deep features from the encoded portion, the convergence speed is accelerated and the model's versatility is enhanced. Specifically: ,in For the first One residual convolutional layer, This represents the convolution operation. This represents the modified linear unit activation function; where, in the first residual convolutional layer, Take as input features for the model In other residual convolutional layers, This is the output of the previous pooling layer; Pooling layer: ,in, For the first The output of each residual convolutional layer; To perform global average pooling on the input feature map; For the first The output of each pooling layer; Fully connected layer: calculated as follows: ,in, Indicates a fully connected operation; This is the output of the last pooling layer.
5. The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network according to claim 1, characterized in that, The upsampling operation restores spatial resolution and extracts image features using CNN and DC, specifically in the following form: ,in This represents the convolution operation. This is the output of the (n+1)th layer. This is the input layer.
6. The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network according to claim 1, characterized in that, In step S3, the non-equal-weight loss function used in training the model is: ； ； ;in, It is a non-equal-weighted multidimensional loss function; , These are the weighting coefficients for the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR), respectively. , m The number of rain gauges participating in RMSE calculation within the model training area; The output of the model during the training process; These are the water level observation values from the hydrological station; , This represents the number of grid points within the model training region that participate in PSNR calculation; These are the pixel values corresponding to the model output during the training process. These are the pixel values corresponding to the water level observations at the hydrological station. The maximum pixel value in the model output image and the flood forecast process line observation image during the training process.
7. The method for forecasting extreme rainstorms and floods in mountainous areas based on the MUL-U2-NET network according to claim 1, characterized in that, Step S3, which involves training the model using a non-equal-weight loss function, further includes: on the validation set, using gradient descent, adjusting the model parameters along the opposite direction of the gradient to iterate until the loss function is minimized, thus achieving optimal model performance. The parameter update method is as follows: ,in, Indicates specific model parameters, This is the learning rate.