A method and system for predicting urban waterlogging based on a classification-regression cascade architecture

Abstract

This invention discloses a method and system for predicting urban flooding based on a classification-regression cascade architecture. It acquires multi-source heterogeneous spatiotemporal data of the target prediction location, constructs a feature vector of the target point, inputs it into a random forest classifier, and outputs a judgment result on whether it belongs to a significant flooding area; if not, it outputs a safety baseline value; if yes, it constructs a static spatial tensor within a local spatial window and a dynamic spatiotemporal tensor on a preset time backtracking window within that window, and inputs them respectively into a two-branch deep learning regression model. The static branch uses multi-scale convolution and channel attention to extract spatial topological constraint features, while the dynamic branch uses a convolutional long short-term memory network to extract the temporal lag effect of the rainfall process; the output features of the two branches are fused and mapped to a predicted flooding depth value. This invention improves the prediction accuracy of extreme flooding scenarios through physical diversion, reducing the prediction error of deep water areas by approximately 30%, and can serve as a real-time early warning system for urban flooding.

Description

Technical Field

[0001] This invention relates to the field of smart city disaster prevention and mitigation technology, and more specifically to a method and system for predicting urban waterlogging based on a classification-regression cascade architecture. Background Technology

[0002] Driven by both global climate change (increased extreme rainfall) and rapid urbanization (hardening of underlying surfaces and a surge in impervious surfaces), urban flooding is becoming increasingly sudden and destructive. Traditional physical hydrodynamic models (such as SWMM) are computationally overloaded when dealing with large-scale real-time early warning systems and struggle to obtain high-precision real-time parameters of underground pipe networks. While existing data-driven models are computationally fast, they often treat the water accumulation process as a single linear process, neglecting the significant differences in physical mechanisms between shallow water stages (dominated by random disturbances) and deep water stages (dominated by runoff generation and confluence patterns). This results in severely insufficient model fitting accuracy under extreme flooding scenarios.

[0003] Problems existing in the prior art include:

[0004] (1) Confusion of physical processes leads to prediction bias: Existing technologies usually use a single end-to-end regression model to process the full range of water levels. The random disturbances in shallow water and the flow generation and confluence patterns in deep water have fundamental differences in physical mechanisms. A single model is difficult to balance the learning of both at the same time, which easily produces an averaging effect, leading to a serious underestimation of extreme high water levels. (2) Unscientific extraction of time response features: Existing models often arbitrarily set the backtracking time when constructing rainfall features, and fail to align features according to the actual response lag period of urban hydrology; (3) Weak ability to identify multi-scale spatial features: The impact of urban underlying surface (slope, road network, river network) on water accumulation has a scale effect, and traditional convolutional structures are difficult to take into account both local micro-topographic features and macro-drainage area features.

[0005] Therefore, there is an urgent need in this field for an urban waterlogging prediction method that can decouple physical processes and scientifically extract multi-temporal and spatial features. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide an urban waterlogging prediction method based on a classification-regression cascade architecture that overcomes or at least partially solves the above problems; To achieve the above objectives, the present invention adopts the following technical solution: S1. Obtain multi-source heterogeneous spatiotemporal data of the target prediction location. The multi-source heterogeneous spatiotemporal data includes static feature data for characterizing the spatial properties of the underlying surface, and dynamic time-series feature data for characterizing the changes of the rainfall process over time. S2. Construct a target point feature vector corresponding to the target prediction location based on the multi-source heterogeneous spatiotemporal data, input the target point feature vector into the classification model, and obtain the determination result of whether the target prediction location belongs to a significant water accumulation area; S3. If the determination result is negative, output a preset safety benchmark value as the predicted water depth at the target prediction location. S4. If the determination result is yes, then based on the multi-source heterogeneous spatiotemporal data, with the target prediction location as the center, construct a static spatial tensor within a local spatial window, and a dynamic spatiotemporal tensor within the local spatial window that covers a preset time backtracking window. The static spatial tensor and the dynamic spatiotemporal tensor are respectively input into the static branch and the dynamic branch of the pre-trained two-branch deep learning regression model; The static branch is used to extract spatial topological constraint features, and the dynamic branch is used to extract the temporal lag effect of the rainfall process; the output features of the two branches are fused to generate the predicted water depth at the target prediction location.

[0007] Preferably, the static spatial feature data includes terrain slope, impervious surface ratio, distance from river, population density, GDP density, and road network density; the dynamic temporal feature data includes cumulative precipitation at multiple preset time scales.

[0008] Preferably, the classification model is a random forest classifier, and its training process includes: binarizing the actual observed water depth with a preset physical boundary to generate hard labels; traversing the predicted probabilities output by the classification model through cross-validation, plotting the Precision-Recall curve, and finding the probability threshold that maximizes the F1-Score as the classification decision threshold.

[0009] Preferably, the static spatial tensor and the dynamic spatiotemporal tensor are constructed in the following ways: The local spatial window is a 7×7 grid window centered on the target prediction location, and the static spatial tensor is composed of static spatial feature data within the window; the preset time backtracking window is 12 time steps sampled at 1-hour intervals, with the current time as the endpoint, and the dynamic spatiotemporal tensor is composed of dynamic temporal feature data within the 7×7 grid window and at the 12 time steps.

[0010] Preferably, the static branch performs the following steps: The extracted static spatial tensor is subjected to two-dimensional convolution to extract initial spatial features; Multi-scale spatial encoding is performed on the initial spatial features to extract features from different receptive fields; Channel attention weighting is applied to the multi-scale spatially encoded features to output spatial features; The multi-scale spatial encoding uses parallel 1×1 convolutional kernels, 3×3 convolutional kernels, and 3×3 dilated convolutional kernels with an inflation rate of 2. The output feature maps of each convolutional kernel are concatenated along the channel dimension and used as the input for channel attention weighting. The channel attention weighting compresses the features of each channel through global average pooling, generates channel weight coefficients through two fully connected layers, and multiplies the weight coefficients with the input feature map channels of the channel attention weighting module to achieve feature recalibration. Preferably, the dynamic branch performs the following steps: The extracted dynamic spatiotemporal tensor is input into a two-layer stacked two-dimensional convolutional long short-term memory network for processing; The first layer extracts the spatial features of each time step and retains the time dimension, outputs a three-dimensional tensor, and then connects to a batch normalization layer. The second layer integrates information from all time steps and collapses the time dimension to output spatiotemporal features.

[0011] Preferably, the step of fusing the output features of the two branches to generate the predicted water depth at the target location specifically includes: The spatial features output by the static branch and the spatiotemporal features output by the dynamic branch are concatenated along the channel dimension. The concatenated features are then flattened into a one-dimensional vector by passing them through a convolutional layer, a global average pooling layer, a Dropout layer, and a fully connected layer, and mapped to a dimensionless predicted value. The dimensionless predicted value is then denormalized to obtain the predicted water depth.

[0012] Preferably, the dual-branch deep learning regression model uses HuberLoss as the loss function during training, employs the AdamW optimizer to perform backpropagation, and uses a gradient pruning mechanism to prune the gradient norm, with the upper limit of the gradient norm set to 5.0.

[0013] Based on the same inventive concept, this invention also discloses an urban waterlogging prediction system based on a classification-regression cascade architecture, used to execute any of the above-mentioned urban waterlogging prediction methods based on a classification-regression cascade architecture, including: The data acquisition module is used to acquire multi-source heterogeneous spatiotemporal data of the target prediction location. The multi-source heterogeneous spatiotemporal data includes static feature data for characterizing the spatial properties of the underlying surface, and dynamic time-series feature data for characterizing the changes of the rainfall process over time. The classification and screening module is used to construct the target point feature vector corresponding to the target prediction location based on the multi-source heterogeneous spatiotemporal data, input the target point feature vector into the classification model, obtain the determination result of whether the target prediction location belongs to a significant water accumulation area, if not, output the preset safety benchmark value, if yes, trigger the regression prediction module; The regression prediction module, which embeds a dual-branch deep learning regression model, is used to construct a static spatial tensor within a local spatial window and a dynamic spatiotemporal tensor covering a preset time backtracking window within the local spatial window, based on the multi-source heterogeneous spatiotemporal data and with the target prediction location as the center. The static branch extracts spatial topological constraint features, and the dynamic branch extracts the temporal lag effect of the rainfall process. After fusion, the predicted value of water accumulation depth is generated. The output module is used to output the predicted water depth.

[0014] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: (1) Improve the accuracy of extreme values ​​by decoupling the physical stages: The present invention adopts a classification-regression cascade architecture. The pre-classifier performs diversion and cutoff of the shallow water interference stage, and the dedicated regression model focuses on fitting the high water level pattern. This effectively solves the problem of conservative prediction of extreme values ​​caused by the "long tail effect of samples" in traditional machine learning models, and greatly reduces the prediction error of extreme water accumulation scenarios.

[0015] (2) Scientific capture of spatiotemporal lag effect: This invention sets a historical time window that matches the hydrological response lag period, enabling the model to perceive the impact of the saturation effect caused by the previous rainfall on the subsequent water accumulation, thereby improving the accuracy of time series prediction.

[0016] (3) Taking into account both local micro-topography and spatial feature perception of macro-region: By constructing a static feature tensor within a local spatial window and using a feature extraction branch with multi-scale extraction capability, this invention can not only accurately identify the confluence characteristics of local low-lying points, but also perceive the comprehensive runoff pressure caused by the hardening of the underlying surface in the region. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is the network architecture of the dual-branch deep learning regression model provided in this embodiment of the invention; Figure 2This is a performance comparison between the cascaded architecture provided in this embodiment of the invention and a single direct regression model; Figure 3 A comparison of prediction curves from different methods under a typical urban flooding event provided in the embodiments of the present invention; Figure 4 This is a block diagram of an urban waterlogging prediction system based on a classification-regression cascade architecture provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a method for predicting urban waterlogging based on a cascaded architecture, comprising: S1. Obtain multi-source heterogeneous spatiotemporal data of the target prediction location. The multi-source heterogeneous spatiotemporal data includes static feature data for characterizing the spatial properties of the underlying surface, and dynamic time-series feature data for characterizing the changes of the rainfall process over time. S2. Construct a target point feature vector corresponding to the target prediction location based on multi-source heterogeneous spatiotemporal data, input the target point feature vector into the classification model, and obtain the determination result of whether the target prediction location belongs to a significant water accumulation area. S3. If the determination result is negative, output the preset safety benchmark value as the predicted water depth at the target prediction location. S4. If the determination result is yes, then based on multi-source heterogeneous spatiotemporal data, with the target prediction location as the center, construct a static spatial tensor within a local spatial window, and a dynamic spatiotemporal tensor within the local spatial window that covers the preset time backtracking window. The static spatial tensor and the dynamic spatiotemporal tensor are respectively input into the static branch and the dynamic branch of the pre-trained two-branch deep learning regression model; The static branch is used to extract spatial topological constraint features, while the dynamic branch is used to extract the temporal lag effect of the rainfall process. The output features of the two branches are fused to generate the predicted water depth at the target location.

[0021] The following detailed explanation of steps S1-S4 is provided with reference to specific embodiments: S101. Acquisition and Preprocessing of Multi-Source Heterogeneous Spatiotemporal Data This example uses a city's urban flooding monitoring area as an example to perform the following data preparation operations: (a) Data Acquisition and Alignment First, gridded spatiotemporal slice data containing metadata for the target area is acquired. This slice data includes static spatial feature data and dynamic temporal feature data for each grid point. The static spatial feature data includes six attributes: terrain slope, proportion of impervious surfaces, distance from river, population density, GDP density, and road network density. The dynamic temporal feature data is the cumulative precipitation time series at multiple preset time scales for each grid point.

[0022] Based on 30m high-precision terrain data, the nearest neighbor interpolation method was used to resample all static feature data (terrain slope, impervious surface ratio, distance from river, population density, GDP density, road network density) to ensure that the static geographic layer is strictly aligned on the spatial grid and has a uniform resolution of 30m. For dynamic time-series feature data (cumulative precipitation), the original spatial resolution is 1 km. In the subsequent classification and initial screening stage, in order to construct a one-dimensional feature vector of the target point, bilinear interpolation is used to interpolate the rainfall data within a 1 km grid around the target location to a 30 m grid at that point; in the regression prediction stage, the dynamic data maintains the original 1 km resolution and is not downscaled or resampled to avoid introducing pseudo-precision.

[0023] (II) Definition of Dynamic Time Series Feature Data For each grid point, its dynamic temporal feature data is constructed as follows: Based on the hysteresis effect of urban hydrological response, the historical time backtracking window is set to the past 12 hours, and divided at 1-hour intervals to obtain 12 time steps. At each time step, cumulative precipitation at multiple time scales is extracted as features, including 1-hour, 2-hour, 3-hour, 6-hour, and 24-hour cumulative rainfall, and the number of feature channels is determined to be C=5. Therefore, the dynamic temporal feature data of each grid point can be represented as a two-dimensional array: 12 (time steps) × 5 (number of feature channels). This data only describes the temporal variation of rainfall at that point and does not contain any spatial neighborhood information.

[0024] (III) Constructing Local Spatial Windows and Static Spatial Tensors Centered on the water accumulation monitoring station (i.e., the target predicted location), a local spatial window with a spatial size of 7×7 grid is constructed outward. This window size effectively captures the confluence characteristics of micro-topography while avoiding the introduction of excessive irrelevant noise.

[0025] Based on this window, the static attributes of all grid points are extracted from the static spatial feature data and organized into a static spatial tensor with dimensions of 7 (height) × 7 (width) × 6 (number of static feature channels).

[0026] It should be noted that, since the spatial resolution of the static feature data is 30m, the actual geographic range corresponding to this 7×7 grid window is 210m × 210m. This window size was determined through hydrological sensitivity analysis, which can effectively capture local runoff characteristics such as micro-topographical runoff and road network constraints, while avoiding the introduction of excessive spatial noise.

[0027] (iv) Constructing a dynamic spacetime tensor by integrating space and time Arrange the dynamic temporal feature data (each point has a dimension of 12×4) of each grid point within the aforementioned local spatial window according to their spatial location, and add a spatial dimension to obtain the dynamic spatiotemporal tensor. This tensor contains both temporal evolution and spatial distribution information, and its dimensions are: 12 (time steps) × 7 (height) × 7 (width) × C (number of feature channels). In this embodiment, C=4, corresponding to the cumulative rainfall in 1 hour, 3 hours, 6 hours, and 24 hours, so the dynamic spatiotemporal tensor dimension is 12×7×7×4.

[0028] It should be noted that, unlike the static branch, the dynamic feature data (cumulative precipitation) in the dynamic spatiotemporal tensor retains its original 1km resolution. Therefore, the actual geographical range corresponding to this 7×7 grid window is 7km × 7km. The use of a larger-scale spatial window is a design based on the urban hydrological response mechanism: the spatial variation of the rainfall process is gradual, and a 1km resolution is sufficient to characterize the regional rainfall distribution; while the 7km window can effectively capture the delayed contribution of runoff from the upstream area surrounding the target location to the formation of water accumulation at the target point.

[0029] In summary, although both the static and dynamic branches use a 7×7 window, their physical coverage differs: the static branch focuses on local micro-topography (210m×210m), while the dynamic branch senses the regional rainfall field (7km×7km). This differentiated window design avoids the pseudo-accuracy problem caused by forcibly downscaling large-scale rainfall data to a smaller grid.

[0030] In one embodiment, to satisfy the input constraints of different levels of models in the cascaded architecture, the original tensor is subjected to differentiated feature assembly: For the initial classification screening: extract the latest time step from the dynamic time series feature data, that is, the dynamic features of the current time center grid (target prediction position) (dimension is 1×C, C=5 in this embodiment), and 6-dimensional static spatial feature data, and concatenate the two into a one-dimensional feature vector. In this embodiment, the total dimension is 11. For the regression model: retain the 7×7×6 static space tensor and the 12×7×7×4 dynamic spatiotemporal tensor, which will be used as inputs for the subsequent static and dynamic branches, respectively.

[0031] S102. Initial screening and diversion of waterlogging risk based on classification model The constructed one-dimensional concatenated feature vector is then input into a pre-trained classification model.

[0032] In this embodiment, the classification model uses a one-dimensional feature vector as input, which is composed of two concatenated parts: Static component: Extract 6-dimensional static spatial feature data (slope, proportion of impervious surface, distance from river, population density, GDP density, road network density); Dynamic component: Extract rainfall features from the current time and target prediction location in the dynamic time series feature data, specifically including the cumulative rainfall over the past 1 hour, 2 hours, 3 hours, 6 hours, and 24 hours (5 dimensions in total).

[0033] The aforementioned one-dimensional feature vector has a total dimension of 11. The classification model employs a lightweight model such as random forest, resulting in extremely low computational complexity. The core purpose of this design is to implement an early termination mechanism: for samples judged as having non-significant water accumulation, which accounts for the vast majority of real-world scenarios, a safe baseline value is directly output without constructing and calculating complete static / dynamic spatiotemporal tensors, nor is it necessary to call subsequent two-branch deep learning regression models. If the classification model were to handle large tensors like the regression model, each sample would require extensive convolutional or recurrent neural network computations, completely negating the computational efficiency advantage of early termination and failing to meet the timeliness requirements of real-time early warning.

[0034] The classifier is trained as follows: The maximum observed surface water depth is extracted, and a preset physical limit of 0.15 meters is used as the standard for label mapping: samples with a water depth < 0.15 meters are marked as 0 (non-significant water accumulation), and samples ≥ 0.15 meters are marked as 1 (significant water accumulation). Instead of using the conventional 0.5 probability threshold, the model iterates through all predicted probabilities using cross-validation, plots a Precision-Recall curve, and searches for the optimal decision threshold that maximizes the F1-Score. For example, the optimal threshold obtained on the validation set is 0.72 (this value varies depending on the dataset).

[0035] The classifier outputs the probability p of belonging to the positive class for each grid cell to be predicted. If p < the optimal judgment threshold, it is judged as a non-significant water accumulation area, the system triggers the early termination mechanism, directly outputs the preset non-water accumulation depth value of 0 meters, and does not perform subsequent complex tensor operations on it; if p ≥ the threshold, it is judged as a significant water accumulation area, the grid index is recorded, and the complete static tensor (7×7×6) and dynamic spatiotemporal tensor (12×7×7×4) corresponding to the grid cell are allowed to enter the static and dynamic branches of the regression model.

[0036] S103. Refined water depth regression prediction based on a dual-branch deep learning architecture Combination Figure 1The diagram shown is of the CNN-ConvLSTM neural network architecture. This step only processes the positive samples selected in the previous step.

[0037] Before performing bi-branch regression prediction, two pre-trained independent scalers are invoked to normalize each channel of the 7×7×6 static tensor and each channel of the 12×7×7×4 dynamic spatiotemporal tensor, respectively, to eliminate gradient oscillations caused by different physical dimensions (such as slope and rainfall). The normalized data are then input into the static branch and the dynamic branch, respectively.

[0038] (1) Static spatial feature extraction branch Combination Figure 1 The content within the orange dashed box will be explained in detail: The input is a 7×7×6 static tensor. Initial spatial features are extracted first through a 3×3 two-dimensional convolutional layer (32 filters, ReLU activation). Then, the input enters the multi-scale spatial encoding module. This module uses three convolutional kernels with different receptive fields in parallel: 1×1 convolution, 3×3 convolution, and 3×3 dilated convolution with a dilation rate of 2. Each branch of the convolutional kernel has 8 filters. The output features of each convolutional kernel are concatenated along the channel dimension and then batch normalized. They then enter the channel attention SE module: each 2D feature map undergoes global average pooling to compress it into a global scalar; it then passes through two fully connected layers: the first layer compresses the number of channels with ReLU activation, and the second layer restores the number of channels with Sigmoid activation, generating channel weight coefficients; these coefficients are then multiplied channel by channel of the input feature map. In one embodiment, the above multi-scale spatial encoding-channel attention enhancement process is executed twice in series to further extract deep spatial features, ultimately outputting the spatial features.

[0039] (2) Dynamic spatiotemporal feature extraction branch Combination Figure 1 The content within the blue dashed box will be explained in detail: The input is a 12×7×7×4 dynamic spatiotemporal tensor. This is fed into a two-layer stacked 2D convolutional long short-term memory network (ConvLSTM2D), with its hidden state feature dimension configured to 16. The first ConvLSTM2D layer extracts spatial features from each frame while preserving the temporal dimension, outputting a feature tensor that fuses spatiotemporal information, and then performs batch normalization. The second ConvLSTM2D layer performs the final integration of the entire time series information, folding the temporal dimension and outputting a 2D feature map retaining only the spatial and channel dimensions (7×7×16) as the spatiotemporal features.

[0040] (3) Feature cascade and regression output Combination Figure 1 The content on the right side will be explained in detail: The spatial features from the static branch and the spatiotemporal features from the dynamic branch are concatenated along the channel dimension to form a feature cube containing complete spatiotemporal context. The concatenated tensor is then passed through a 3×3 convolutional layer (64 channels, Swish activation function) and batch normalization, followed by a 3×3 convolutional layer (128 channels, Swish) with a stride of 2 for spatial dimensionality reduction. A global average pooling layer flattens the tensor into a one-dimensional vector, and a Dropout layer is added to prevent overfitting. Finally, after passing through a hidden layer (64 neurons, Swish activation), it is connected to the final output layer to output a single water depth prediction value.

[0041] To eliminate invalid computational artifacts caused by sensor noise and surface disturbances, a forced denoising physical threshold of 0.01 meters can be introduced: when the predicted water depth after inverse normalization is lower than this threshold, it is forcibly corrected to 0.

[0042] Through the aforementioned differentiated design, this invention can filter out a large number of non-waterlogged samples (early withdrawal) with relatively low computational cost, and only activate the refined regression model for a few high-risk samples, thereby meeting the timeliness requirements of real-time early warning while maintaining prediction accuracy. Specifically, the static branch and the dynamic branch employ different spatial window ranges (210m vs 7km) and different rainfall time scales, respectively adapting to the needs of their respective physical mechanisms, achieving an optimal balance between computational efficiency and prediction accuracy.

[0043] (4) Model training During regression model training, Huber Loss is explicitly used as the core loss function: squared loss is used for samples with residual absolute value less than δ (δ=1.0 in this example), and linear loss is used for samples with residual absolute value ≥ δ, thereby effectively suppressing the huge gradient oscillations caused by outliers in extreme rainstorm scenarios.

[0044] The AdamW optimizer is used to execute the backpropagation algorithm, with an initial learning rate of 0.0001 and an explicit weight decay coefficient of 1e-6. Meanwhile, to address the gradient explosion problem that easily occurs in deep time series recurrent networks during backpropagation, a gradient pruning mechanism is introduced, forcibly limiting the upper limit of the gradient norm of each layer's parameters to 5.0.

[0045] S104. Dynamically map the predicted water depth to a disaster risk classification based on a tiered physical defense threshold: When the water depth is below 0.15m, it is classified as Level 1 safe (it should be noted that for predicted points that are classified as non-significant water accumulation and directly output as the safety benchmark value (0 meters), they automatically fall into the range of water depth below 0.15 meters in the final risk classification, and are therefore classified as Level 1 safe). Water depth between 0.15m and 0.5m is classified as Level 2 moderate risk; water depth greater than or equal to 0.5m is classified as Level 3 severe waterlogging.

[0046] Meanwhile, the population density scalar is read from the central grid of the static feature tensor and multiplied by the discretized hazard level parameter to generate a weighted population exposure index.

[0047] Finally, the metadata coordinates, target prediction time, predicted absolute water depth, hazard level, and weighted exposure are packaged and encapsulated into a structured concurrent early warning sequence matrix, which is then output to the downstream disaster prevention and mitigation emergency response terminal.

[0048] To verify the effectiveness of the logical branching assignment mechanism based on physical magnitude differences and the technical effects of multi-scale spatial coding and physical constraint time backtracking of the present invention, this application compared the cascaded architecture with the commonly used single end-to-end direct regression model in the prior art.

[0049] (I) Performance Comparison of Cascaded Architecture and Direct Regression Model Combination Figure 2 As shown, this experiment compares the predictive performance of the cascaded architecture of this invention with that of a single end-to-end direct regression model in a real urban environment. The test dataset contains a large number of non-flooding samples (zero-value samples) and flooding samples of different intensities.

[0050] Figure 2 Figures (a) and (c) show the performance and error metrics of the two strategies on the full test set. The performance metrics include the Nash efficiency coefficient (NSE) and the Pearson correlation coefficient (Pearson r), while the error metrics include the root mean square error (RMSE) and the mean absolute error (MAE). As can be seen from the figures, on the full test set, due to the interference of a large number of zero-value samples, the performance difference between the two models is not significant. The single direct regression model even performs reasonably well on some metrics, because the large number of easily predictable zero-value samples inflates the overall evaluation metrics.

[0051] However, regarding the core needs of actual urban flooding prevention, namely the prediction accuracy for severe urban flooding scenarios (actual observed water depth ≥ 0.15m), such as... Figure 2 As shown in (b) and 2(d), the performance of the two models diverged significantly. Figure 2 (b) shows that, on samples of severe urban flooding, the NSE and Pearson r index of the cascaded architecture of this invention are significantly better than the single direct regression model. Figure 2 (d) shows that the RMSE of the cascaded architecture of this invention decreased by about 29% compared to the single direct regression model (from about 0.38m to less than 0.27m), and the MAE also decreased significantly accordingly.

[0052] The above comparison fully demonstrates that the present invention eliminates random interference from shallow water by using a pre-classifier, enabling the subsequent dedicated regression sub-model to focus on fitting the flow patterns in deep water areas. This effectively overcomes the technical challenge of severely underestimating extreme high water levels in existing technologies and significantly improves the prediction accuracy in deep water areas.

[0053] (II) Comparison of prediction curves for typical urban flooding events Combination Figure 3 As shown, this experiment selected four typical waterlogging events and compared the prediction results of the method of this invention (CNN-ConvLSTM, red line) with those of the existing technologies of multiple linear regression (MLR, green line) and long short-term memory network (LSTM, blue line). The actual observed water depth is represented by the black line.

[0054] This reveals that when faced with a rapid rise in water levels caused by sudden heavy rainfall, the MLR model (green line) completely fails to capture the nonlinear trend, and its predicted curve deviates significantly from the actual value. While the LSTM model (blue line) can learn certain temporal patterns, its predicted curve shows obvious stagnation and peak clipping after the actual water level reaches the 0.5m to 0.8m range, completely losing its ability to track high-risk water levels. Its predicted peak value is significantly lower than the actual peak value, and there is a significant delay in the peak arrival time. In contrast, the prediction results of the method of this invention (red line) are relatively close to the actual observed values ​​(black line). Specifically, this invention successfully captured and predicted extreme flood peaks exceeding 1.5m, with the time lag error of peak arrival controlled within a very small range. This indicates that by introducing a multi-scale spatial coding module and a physically constrained time backtracking mechanism, this invention significantly improves the model's ability to perceive sudden runoff accumulation, not only solving the problem of traditional models' inaccurate measurement of extreme values ​​but also providing valuable lead time for emergency disaster prevention in high-risk urban areas.

[0055] like Figure 4 As shown, based on the same inventive concept, this invention also discloses an urban waterlogging prediction system based on a classification-regression cascade architecture, used to execute any urban waterlogging prediction method based on a classification-regression cascade architecture, including: The data acquisition module is used to acquire multi-source heterogeneous spatiotemporal data of the target prediction location. The multi-source heterogeneous spatiotemporal data includes static feature data for characterizing the spatial properties of the underlying surface, and dynamic time-series feature data for characterizing the changes of the rainfall process over time. The classification and screening module is used to construct the target point feature vector corresponding to the target prediction location based on multi-source heterogeneous spatiotemporal data, input the target point feature vector into the classification model, and obtain the judgment result of whether the target prediction location belongs to a significant water accumulation area. If not, it outputs the preset safety benchmark value; if yes, it triggers the regression prediction module. The regression prediction module embeds a dual-branch deep learning regression model, which is used to construct a static spatial tensor within a local spatial window and a dynamic spatiotemporal tensor within the local spatial window covering a preset time backtracking window, based on multi-source heterogeneous spatiotemporal data and with the target prediction location as the center. The static branch extracts spatial topological constraint features and the dynamic branch extracts the temporal lag effect of the rainfall process, and the fusion is used to generate the predicted value of water accumulation depth. The output module is used to output the predicted water depth.

[0056] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0057] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting urban waterlogging based on a classification-regression cascade architecture, characterized in that, include: S1. Obtain multi-source heterogeneous spatiotemporal data of the target prediction location. The multi-source heterogeneous spatiotemporal data includes static feature data for characterizing the spatial properties of the underlying surface, and dynamic time-series feature data for characterizing the changes of the rainfall process over time. S2. Construct a target point feature vector corresponding to the target prediction location based on the multi-source heterogeneous spatiotemporal data, input the target point feature vector into the classification model, and obtain the determination result of whether the target prediction location belongs to a significant water accumulation area; S3. If the determination result is negative, output a preset safety benchmark value as the predicted water depth at the target prediction location. S4. If the determination result is yes, then based on the multi-source heterogeneous spatiotemporal data, with the target prediction location as the center, construct a static spatial tensor within a local spatial window, and a dynamic spatiotemporal tensor within the local spatial window that covers a preset time backtracking window. The static spatial tensor and the dynamic spatiotemporal tensor are respectively input into the static branch and the dynamic branch of the pre-trained two-branch deep learning regression model; The static branch is used to extract spatial topological constraint features, and the dynamic branch is used to extract the temporal lag effect of the rainfall process; the output features of the two branches are fused to generate the predicted water depth at the target prediction location.

2. The method according to claim 1, characterized in that, The static spatial feature data includes terrain slope, impervious surface ratio, distance from river, population density, GDP density, and road network density; the dynamic temporal feature data includes cumulative precipitation at multiple preset time scales.

3. The method according to claim 1, characterized in that, The classification model is a random forest classifier, and its training process includes: binarizing the actual observed water depth with a preset physical boundary to generate hard labels; traversing the predicted probabilities output by the classification model through cross-validation, plotting the Precision-Recall curve, and finding the probability threshold that maximizes the F1-Score as the classification decision threshold.

4. The method according to claim 1, characterized in that, The static spatial tensor and the dynamic spatiotemporal tensor are constructed as follows: The local spatial window is a 7×7 grid window centered on the target prediction location, and the static spatial tensor is composed of static spatial feature data within the window; the preset time backtracking window is 12 time steps sampled at 1-hour intervals, with the current time as the endpoint, and the dynamic spatiotemporal tensor is composed of dynamic temporal feature data within the 7×7 grid window and at the 12 time steps.

5. The method according to claim 1, characterized in that, The static branch performs the following steps: The extracted static spatial tensor is subjected to two-dimensional convolution to extract initial spatial features; Multi-scale spatial encoding is performed on the initial spatial features to extract features from different receptive fields; Channel attention weighting is applied to the multi-scale spatially encoded features to output spatial features; The multi-scale spatial encoding uses parallel 1×1 convolutional kernels, 3×3 convolutional kernels, and 3×3 dilated convolutional kernels with an inflation rate of 2. The output feature maps of each convolutional kernel are concatenated along the channel dimension and used as the input for channel attention weighting. The channel attention weighting compresses the features of each channel through global average pooling, generates channel weight coefficients through two fully connected layers, and multiplies the weight coefficients with the input feature map channels of the channel attention weighting module to achieve feature recalibration.

6. The method according to claim 1, characterized in that, The dynamic branch performs the following steps: The extracted dynamic spatiotemporal tensor is input into a two-layer stacked two-dimensional convolutional long short-term memory network for processing; The first layer extracts the spatial features of each time step and retains the time dimension, outputs a three-dimensional tensor, and then connects to a batch normalization layer. The second layer integrates information from all time steps and collapses the time dimension to output spatiotemporal features.

7. The method according to claim 1, characterized in that, The process of fusing the output features of the two branches to generate a predicted water depth at the target location specifically includes: The spatial features output by the static branch and the spatiotemporal features output by the dynamic branch are concatenated along the channel dimension. The concatenated features are then flattened into a one-dimensional vector by passing them through a convolutional layer, a global average pooling layer, a Dropout layer, and a fully connected layer, and mapped to a dimensionless predicted value. The dimensionless predicted value is then denormalized to obtain the predicted water depth.

8. The method according to claim 1, characterized in that, The dual-branch deep learning regression model uses HuberLoss as the loss function during training, employs the AdamW optimizer to perform backpropagation, and uses a gradient pruning mechanism to prune the gradient norm, with the upper limit of the gradient norm set to 5.

0.

9. A city waterlogging prediction system based on a classification-regression cascade architecture, used to execute any one of the city waterlogging prediction methods based on a classification-regression cascade architecture as described in claims 1-8, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous spatiotemporal data of the target prediction location. The multi-source heterogeneous spatiotemporal data includes static feature data for characterizing the spatial properties of the underlying surface, and dynamic time-series feature data for characterizing the changes of the rainfall process over time. The classification and screening module is used to construct the target point feature vector corresponding to the target prediction location based on the multi-source heterogeneous spatiotemporal data, input the target point feature vector into the classification model, obtain the determination result of whether the target prediction location belongs to a significant water accumulation area, if not, output the preset safety benchmark value, if yes, trigger the regression prediction module; The regression prediction module, which embeds a dual-branch deep learning regression model, is used to construct a static spatial tensor within a local spatial window and a dynamic spatiotemporal tensor covering a preset time backtracking window within the local spatial window, based on the multi-source heterogeneous spatiotemporal data and with the target prediction location as the center. The static branch extracts spatial topological constraint features, and the dynamic branch extracts the temporal lag effect of the rainfall process. After fusion, the predicted value of water accumulation depth is generated. The output module is used to output the predicted water depth.

Images