Intelligent short-time flood response forecasting method for complex basin based on satellite precipitation perception

Abstract

The application provides a complex basin short-time flood response intelligent prediction method based on satellite precipitation perception, and relates to the technical field of hydrological meteorological monitoring and disaster prevention and mitigation. The method comprises the following steps: acquiring satellite precipitation data of a target basin observed by a multi-channel satellite, and performing pretreatment; inputting the pretreated satellite precipitation data into a short-time precipitation prediction model to output a short-time satellite precipitation prediction result; and inputting the short-time satellite precipitation prediction result into a semi-distributed hydrological model to simulate a flood response process at a basin scale and output predicted flood process data. The application solves the precision problem of short-time precipitation prediction of a complex basin and improves the timeliness and stability of flood prediction.

Description

Technical Field

[0001] This invention relates to the field of hydrological and meteorological monitoring and disaster prevention and mitigation technology, and in particular to an intelligent forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing. Background Technology

[0002] Accurate and timely short-term precipitation forecasts are a key constraint for reliable flood forecasting, especially in large lake basins. Uncertainties in precipitation input are often rapidly amplified within a short period through watershed runoff and lake regulation, directly impacting the accuracy of peak flood magnitude and timing forecasts. However, traditional flood forecasting systems heavily rely on surface rain gauge observations, which are spatially unevenly distributed within watersheds—sparsely distributed in mountainous areas and relatively densely distributed in plains—making it difficult to effectively characterize the fine-scale precipitation structure triggered by strong convection and orographic lifting. This observational bottleneck not only leads to systematic biases in areal rainfall estimation but also allows precipitation uncertainties within the short forecast period to be directly transmitted and amplified in flood forecast results, becoming a critical challenge restricting the accuracy and timeliness of flood warnings in complex watersheds.

[0003] Existing short-term precipitation forecasting methods are mainly based on numerical weather prediction (NWP) models and their related physical frameworks. These methods involve complex computational processes and lengthy inference times, typically requiring several hours or even longer to complete a single forecast. Precipitation nowcasting methods based on traditional extrapolation algorithms such as optical flow are computationally efficient, but they often rely on assumptions about the stationarity of precipitation patterns, leading to a significant decrease in forecast accuracy under complex weather conditions. In recent years, the development of satellite remote sensing technology has provided new possibilities for overcoming these limitations. With the increasing availability of such high spatiotemporal resolution satellite data, machine learning (ML) and deep learning (DL) methods have gradually become important technical paths for satellite precipitation nowcasting. However, existing deep learning-based precipitation nowcasting research primarily focuses on radar or idealized observation data. When these models are directly applied to multi-channel, highly correlated observation data from geostationary meteorological satellites, they face the following challenges: 1) They are prone to significant feature redundancy, weakening the effective representation of key precipitation characteristics. 2) Traditional models tend to degrade in areas with complex terrain, and their fixed convolutional kernels are difficult to adapt to the nonlinear cloud evolution triggered by topographic uplift.

[0004] Furthermore, in flood forecasting practice, short-term precipitation nowcasting is not the ultimate goal. Its core value lies in whether it can effectively drive hydrological or hydrodynamic models as an external forcing condition, thereby forming a reliable simulation of the flood evolution process. In traditional flood forecasting practice, precipitation prediction is separated from the hydrological response process, lacking physical consistency feedback and verification, leading to the accumulation and amplification of errors during transmission. In complex watersheds with large lakes and rivers coupled, the combined effects of lake regulation, river confluence, and topographic effects further enhance the sensitivity of the hydrodynamic response to the spatiotemporal structure of precipitation, making the practice of substituting the accuracy of precipitation prediction for the reliability of flood forecasting significantly limited. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes an intelligent forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing. This method generates short-term satellite precipitation prediction results through a lightweight attention-enhanced LA-UNet model, which is then embedded into a hydrological model to quantify the propagation and amplification effects of precipitation prediction errors in the flood response process. This effectively suppresses the cumulative effect of errors in runoff calculation and significantly improves the usability and stability of flood forecasts.

[0006] This invention provides, in one aspect, a smart forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing, comprising:

[0007] Acquire satellite precipitation data of the target watershed from multi-channel satellite observations and perform preprocessing;

[0008] The preprocessed satellite precipitation data is input into a short-term precipitation prediction model, which outputs short-term satellite precipitation prediction results; wherein, the short-term precipitation prediction model includes:

[0009] The perception layer is used to organize the preprocessed satellite precipitation data into a single-channel raster time series tensor.

[0010] The cognitive layer is constructed based on the Lightweight Attention-Based U-Net model, which uses the single-channel raster temporal tensor as input to generate the short-term satellite precipitation prediction results.

[0011] The short-term satellite precipitation prediction results are input into a semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, and output predicted flood process data.

[0012] In one embodiment of the present invention, the construction process of the perception layer includes:

[0013] The preprocessed satellite precipitation data is organized in a single-channel raster time series format, with each time step corresponding to a quantitative precipitation estimation precipitation field, characterizing the continuous evolution of precipitation.

[0014] The final result is a single-channel raster time series tensor with dimensions T×H×W×C, where C=1 indicates single-channel input, T is the time series length, H represents the spatial height dimension of the precipitation field, and W represents the spatial width dimension of the precipitation field.

[0015] In one embodiment of the present invention, a lightweight attention-enhanced LA-UNet model is introduced to construct a cognitive layer, comprising:

[0016] Based on the standard U-Net model, channel-space dual-dimensional attention modules are embedded in multiple network layers at different scales of the encoder, and all standard convolutional layers in the encoder and decoder are replaced with depthwise separable convolutional layers to obtain the lightweight attention-enhanced LA-UNet model.

[0017] The dual-dimensional attention module is used to extract highly correlated multi-channel features from the single-channel raster temporal tensor for adaptive recalibration during feature downsampling and semantic abstraction.

[0018] Among them, the channel-space dual-dimensional attention module compresses the spatial dimension through global average pooling to generate channel attention weights, highlighting radiation and cloud system information channels that are closely related to precipitation formation and evolution;

[0019] In terms of spatial attention, a spatial attention weight map is generated through a depthwise separable convolutional layer to focus on the dominant precipitation belt, enhance the features of the dominant precipitation belt, and weaken the feature responses of non-dominant precipitation belts; wherein, the dominant precipitation belt includes strong convective cloud clusters and topographically lifted precipitation belts.

[0020] In one embodiment of the present invention, the method further includes: selecting a benchmark model and performing comparative training with the lightweight attention-enhanced LA-UNet model, wherein the benchmark model includes a standard U-Net model, a UNet++ model, and a ConvLSTM model, comprising:

[0021] Based on multiple precipitation thresholds, the spatial distribution accuracy of precipitation prediction results from various models under multiple short-term forecast lead times is evaluated, including:

[0022] The channel-space dual-dimensional attention module is embedded into the standard U-Net model and the UNet++ model respectively to form the U-NetCBAM model and the UNet++CBAM model with attention mechanism;

[0023] The performance of the U-NetCBAM model with attention mechanism, UNet++CBAM model, standard U-Net model, UNet++ model, ConvLSTM model, and the lightweight attention-enhanced LA-UNet model were compared. Based on low, medium, and high precipitation thresholds, the precipitation prediction results of each model under multiple short-term forecast lead times were verified.

[0024] The accuracy of precipitation spatial distribution of each model is evaluated based on the evaluation indicators of probability detection rate, false alarm rate, critical success index and F1 score.

[0025] In one embodiment of the present invention, the process of selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model for training further includes:

[0026] Evaluating the feature redundancy of different models at different network layers includes:

[0027] Extract the feature map tensors of the standard U-Net model, UNet++ model and the lightweight attention-enhanced LA-UNet model at shallow, middle and deep layers, and flatten the features of each channel into a one-dimensional vector;

[0028] Calculate the Pearson correlation coefficient between channels, construct the channel correlation matrix, and analyze the differences in deep feature representation capabilities among the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model.

[0029] In one embodiment of the present invention, the process of selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model for training further includes:

[0030] Selecting typical regions of different categories within the target watershed, a regional comparative analysis of each model is conducted in these typical regions, including:

[0031] The gradient-weighted class activation mapping method is used to visualize the typical regions of different categories that the standard U-Net model, UNet++ model and lightweight attention-enhanced LA-UNet model focus on in the short-term forecast timeliness task.

[0032] The output of the feature generation layer before the final convolutional layer of each model is selected as the analysis object, and attention heatmaps are generated for comparison in different typical regions of the target watershed; wherein, the different typical regions include plains, lakeside transition zones and mountain uplift zones.

[0033] In one embodiment of the present invention, the process of selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model for training further includes:

[0034] Regional comparative analysis of each model was conducted in typical regions of the different categories, including:

[0035] Evaluate whether the outputs of different models meet the basic physical and statistical constraints of hydrological simulation on rainfall input, including:

[0036] The short-term precipitation prediction results output by the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model are compared and evaluated.

[0037] Hourly precipitation prediction results and cumulative precipitation processes for typical rainfall events were calculated for the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model, respectively.

[0038] Under precipitation scenarios dominated by typical regions of different categories, the prediction results indicators of each model are calculated under multiple short-term forecast lead times, and the error characteristics of the models are compared and analyzed. The prediction results indicators include: mean absolute error, correlation coefficient, and mean relative root mean square error.

[0039] In one embodiment of the present invention, the short-term satellite precipitation prediction results are input into a semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, and the predicted flood process data is output, including:

[0040] The preprocessed satellite precipitation data and the short-time satellite precipitation prediction results are input together into the semi-distributed hydrological model;

[0041] Using the measured flood process data from hydrological stations within the target watershed, the runoff generation and confluence parameters of the semi-distributed hydrological model are calibrated and validated; wherein, the measured flood process data includes: measured peak flow, measured total flood volume, and measured peak occurrence time for each time period;

[0042] The semi-distributed hydrological model outputs the predicted flood process data under multiple forecast timeframes. The predicted flood process data includes: predicted peak flow, predicted total flood volume, and predicted peak time for each time period.

[0043] The river channel calculation process of the semi-distributed hydrological model follows the law of conservation of mass, and the governing equation is shown in equation (1):

[0044] (1),

[0045] In the formula, and O These represent inflow and outflow respectively. S Represents the potential retained water volume of a sub-basin. t For time parameters;

[0046] Based on the acquired basic geographic data of the target watershed, the boundaries of sub-watersheds are divided within the target watershed, and key topographic features such as river network, watershed slope, and slope length are extracted to divide several sub-watersheds; wherein, the basic geographic data includes topography, land use type, and soil data;

[0047] The runoff generation process of each sub-basin was simulated, and the infiltration excess rainfall was calculated using the SCS curve number method, as shown in equation (2):

[0048] (2),

[0049] In the formula, P e This refers to rainfall exceeding the infiltration limit. P Total rainfall, This is the initial loss amount. S max This represents the maximum potential water retention capacity.

[0050] The confluence process of each sub-basin was simulated. The Clark unit hydrograph method was used at the sub-basin scale, and the Muskingan method was used at the river network scale to calculate the river channel storage, as shown in equation (3):

[0051] (3),

[0052] In the formula, The storage time constant; These are the Muskingan weighting coefficients, with values ​​ranging from 0 to 1. ≤0.5 is used to reflect the translation and flattening characteristics of flood waves in the river channel.

[0053] In one embodiment of the present invention, the method further includes: constructing a multi-index evaluation system to evaluate the short-term precipitation prediction model and the semi-distributed hydrological model, comprising:

[0054] The probability detection rate, false alarm rate, critical success index, and F1 score are used as the first evaluation indicators to assess the accuracy of the spatial distribution prediction of precipitation in the short-term precipitation prediction model.

[0055] The short-term precipitation prediction accuracy of the short-term precipitation prediction model is evaluated based on the measured rainfall data from rain gauge stations and the short-term satellite precipitation prediction results, using mean absolute error, correlation coefficient, and mean relative root mean square error as the second evaluation index.

[0056] The relative error of peak flow, the relative error of total flood volume, the peak occurrence time error, and the Nash efficiency coefficient are used as the third evaluation index. The measured flood process data of hydrological stations in the target watershed are used as a reference benchmark and compared with the predicted flood process data output by the semi-distributed hydrological model to evaluate the flood process simulation accuracy of the semi-distributed hydrological model.

[0057] The multi-indicator evaluation system includes a first evaluation indicator, a second evaluation indicator, and a third evaluation indicator.

[0058] In one embodiment of the present invention, preprocessing of satellite precipitation data includes:

[0059] The satellite precipitation data were subjected to geometric correction, orthorectification, and spatial mapping and terrain correction.

[0060] The corrected satellite precipitation data is uniformly projected onto a geographic coordinate system grid with a preset resolution, and then subjected to numerical normalization and time synchronization processing. The corrected satellite precipitation data is then regionally cropped.

[0061] After removing invalid and outlier values ​​from the regionally cropped satellite precipitation data, it is divided into training, validation, and test sets according to a preset ratio.

[0062] As can be seen from the above solutions, the advantages of the present invention are:

[0063] This invention provides an intelligent forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing. The method preprocesses satellite precipitation data from a target watershed observed by multiple satellites and inputs it into a short-term precipitation prediction model, outputting short-term satellite precipitation prediction results. This model is built on a lightweight attention-enhanced LA-UNet model, which can adaptively identify fine precipitation structures such as strong convective clouds and orographic lifting zones, providing high-precision short-term precipitation input for flood forecasting. Then, the short-term satellite precipitation prediction results are input into a semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, outputting predicted flood process data. This method systematically quantifies the propagation and amplification effects of short-term precipitation prediction errors in the flood response process, effectively suppressing the cumulative effect of errors in runoff generation and confluence calculations. It solves the accuracy problem of short-term precipitation prediction in complex watersheds and significantly improves the accuracy and reliability of short-term flood warnings in complex watersheds through an error suppression mechanism, enhancing the timeliness and stability of flood forecasts. Attached Figure Description

[0064] Figure 1 A schematic diagram of the overall process of a smart forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing, provided by an embodiment of the prior art, is shown.

[0065] Figure 2 A schematic diagram showing the comparison of nowcast precipitation forecasts from multiple models is presented.

[0066] Figure 3 The feature similarity matrix and box plot between shallow and deep channels in the multi-model diagram are shown.

[0067] Figure 4 This diagram illustrates a visual comparison of gradient-weighted class activation mappings in a multi-model precipitation nowcasting task.

[0068] Figure 5 The diagram shows the nowcast precipitation process of multiple models under different short-term forecast lead times. (a)-(d) represent hourly precipitation data for different monitoring points using the standard U-Net model, with (a) indicating hourly precipitation data for the first monitoring point, (b) indicating hourly precipitation data for the second monitoring point, (c) indicating hourly precipitation data for the third monitoring point, and (d) indicating cumulative precipitation data for the third monitoring point. (e)-(h) represent hourly precipitation data for different monitoring points using the UNet++ model, with (e) indicating the first monitoring point... Hourly precipitation data for point (f) shows hourly precipitation data for the second monitoring point, (g) shows hourly precipitation data for the third monitoring point, and (h) shows cumulative precipitation data for the third monitoring point; (i)-(l) are hourly precipitation data for different monitoring points using the LA-UNet model of this invention, (i) shows hourly precipitation data for the first monitoring point, (j) shows hourly precipitation data for the second monitoring point, (k) shows hourly precipitation data for the third monitoring point, and (l) shows cumulative precipitation data for the third monitoring point;

[0069] Figure 6 A schematic diagram comparing the precipitation nowcasting process of multiple models under different short-term forecast lead times is shown. Detailed Implementation

[0070] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0071] In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0072] refer to Figure 1 As shown, Figure 1 This diagram illustrates the overall process of an intelligent forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing, provided by an embodiment of the present invention.

[0073] A smart forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing includes the following steps:

[0074] Step S1: Obtain satellite precipitation data of the target watershed from multi-channel satellite observations and perform preprocessing.

[0075] Step S2: Input the preprocessed satellite precipitation data into the short-term precipitation prediction model and output the short-term satellite precipitation prediction results.

[0076] Step S3: Input the short-term satellite precipitation prediction results into the semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, and output the predicted flood process data.

[0077] In this embodiment, satellite precipitation data of a target watershed, such as a large lake basin or other complex watershed, is acquired. This satellite precipitation data serves as the primary input for short-term precipitation nowcasting, fed into a short-term precipitation prediction model. The short-term satellite precipitation prediction results are then input into a semi-distributed hydrological model to simulate a watershed-scale flood response process, outputting predicted flood event data. This method uses satellite precipitation data as its source, constructing an integrated coupling chain between precipitation prediction and hydrological response. The short-term satellite precipitation prediction results effectively constrain the flood evolution process, improving the accuracy of short-term satellite precipitation prediction and the timeliness of flood forecasting.

[0078] In one embodiment, in step S1, satellite precipitation data of the target watershed is acquired and preprocessed. Specifically, the satellite precipitation data undergoes geometric correction, orthorectification, and spatial mapping and topographic correction. Then, the corrected satellite precipitation data is uniformly projected onto a geographic coordinate system grid of a preset resolution, such as a 0.01°×0.01° latitude and longitude grid in the WGS84 geographic coordinate system, and numerical normalization and time synchronization are performed. Regional clipping is then applied to the corrected satellite precipitation data. It should be noted that the satellite precipitation data used in this embodiment is a publicly available dataset from the Fengyun Satellite Remote Sensing Data Service Network, located at https: / / satellite.nsmc.org.cn / DataPortal / cn / home / index.html, and the data does not contain sensitive information.

[0079] The regionally cropped satellite precipitation data undergoes quality screening to ensure data usability and consistency. To address null values ​​and outliers in the satellite data, pixels marked as invalid in the original data and observations exceeding physically reasonable ranges are uniformly identified as null values ​​and removed. Finally, the quality-screened satellite precipitation data is divided into training, validation, and test sets according to a preset ratio, for example, an 8:1:1 ratio.

[0080] In addition, in some embodiments, hourly measured rainfall data from rain gauge stations within the target watershed are collected. For a small number of missing data periods, a spatial weighting method using nearby rain gauge stations is used for interpolation. It should be noted that the hourly measured rainfall data from rain gauge stations used in these embodiments is a publicly available dataset, sourced from the Lake-Watershed Branch Center of the National Earth System Science Data Center, at https: / / lake.geodata.cn / index.html. This data does not contain sensitive information.

[0081] Measured flood event data within the target watershed are acquired for selecting typical flood events to calibrate model parameters and evaluate performance. The measured flood event data includes: measured peak discharge, measured total flood volume, measured peak time, and measured discharge sequence. The measured flood event data used in this embodiment of the invention is a publicly available dataset sourced from the Lake-Watershed Branch Center of the National Earth System Science Data Center, at https: / / lake.geodata.cn / index.html. This data does not contain sensitive information.

[0082] Basic geographic data, including topography, land use, and soil data, for the target area were collected and organized for subsequent hydrological model construction and parameter setting. For example, topographic data was obtained from the 30m resolution SRTM DEM product provided by the U.S. Geological Survey. Land use data was sourced from the 2020 30m resolution product released by the Resource and Environmental Science Data Center of the Chinese Academy of Sciences. Based on research needs, land use types were categorized into five classes: cultivated land, forest land, grassland, construction land, and water area, to characterize runoff generation characteristics under different underlying surface conditions. Soil data was obtained from the World Soil Database product released by the Food and Agriculture Organization of the United Nations, from which key attributes such as soil texture and soil thickness were extracted to estimate soil water storage capacity and infiltration characteristics. It should be noted that the basic geographic data used for the above-mentioned topography, land use, and soil are publicly available datasets, including those from the U.S. Geological Survey (USGS), https: / / earthexplorer.usgs.gov / , the Resource and Environmental Science Data Platform, https: / / www.resdc.cn / , and the World Soil Database (FAO), https: / / www.fao.org / soils-portal / data-hub / soil-maps-and-databases. The data does not contain sensitive information.

[0083] Furthermore, three typical study areas were selected within the target watershed. Considering the degree of topographic relief and the dominant precipitation mechanisms such as convective precipitation and orographic lifting, the target watershed was divided into a plain area, a lakeside transition area, and a mountainous lifting area.

[0084] In one embodiment, in step S2, the short-term precipitation prediction model is constructed based on the Lightweight Attention-Based U-Net (LA-UNet) model, using the LA-UNet model as the core to balance computational efficiency with the spatial focusing capability of key precipitation structures. The short-term precipitation prediction model includes a perception layer and a cognitive layer. The perception layer organizes the preprocessed satellite precipitation data into a single-channel raster time-series tensor; the cognitive layer, constructed based on the Lightweight Attention-Based U-Net model, uses the single-channel raster time-series tensor as input and focuses on characterizing the complex nonlinear relationship between satellite precipitation data and the spatial distribution of precipitation to generate the short-term satellite precipitation prediction result.

[0085] Specifically, for the perception layer, the preprocessed satellite precipitation data is organized in a single-channel raster time series format. Each time step corresponds to a quantitative precipitation estimation precipitation field, representing the continuous evolution of precipitation. Finally, a single-channel raster time series tensor with dimensions T×H×W×C is formed, where C=1 indicates a single-channel input, T is the time series length, H and W represent the spatial dimensions of the precipitation field within the target watershed, H (Height) represents the spatial height dimension of the precipitation field, i.e., the number of rows in the raster data; W (Width) represents the spatial width dimension of the precipitation field, i.e., the number of columns in the raster data.

[0086] For the cognitive layer, a lightweight attention-enhanced LA-UNet model is introduced. Specifically, based on the standard U-Net model, channel-space dual-dimensional attention modules are embedded in multiple scale network layers of the encoder, and all standard convolutional layers in the encoder and decoder are replaced with depthwise separable convolutional layers to obtain the lightweight attention-enhanced LA-UNet model. The standard U-Net model adopts a symmetrical encoder-decoder structure, fusing multi-scale spatial features through skip connections. The encoder consists of multiple sets of convolutional layers + pooling layers, which gradually reduce the spatial resolution through downsampling while doubling the number of feature channels to extract abstract semantic features. A typical structure is: each set contains two 3×3 standard convolutions + ReLU activation, followed by 2×2 max pooling to complete downsampling. The decoder restores the spatial resolution through upsampling and combines the high-resolution features of the encoder to generate fine segmentation results. The structure of each set is: transposed convolution upsampling, concatenated with the corresponding layer feature map of the encoder, and then passed through two 3×3 standard convolutions + ReLU activation. Skip connections directly transmit high-resolution features from each layer of the encoder to the corresponding layer of the decoder, compensating for detail loss during upsampling and improving segmentation accuracy. In this embodiment, based on the standard U-Net model, a channel-space dual-dimensional attention module (CBAM) is embedded in multiple network layers at various scales of the encoder, and all standard convolutional layers of the encoder and decoder are replaced with depthwise separable convolutional layers to obtain the lightweight attention-enhanced LA-UNet model. This dual-dimensional attention module is used to extract highly correlated multi-channel features from the single-channel raster temporal tensor for adaptive recalibration during feature downsampling and semantic abstraction. Specifically, the channel-space dual-dimensional attention module compresses the spatial dimension through global average pooling to generate channel attention weights, highlighting radiation and cloud information channels closely related to precipitation formation and evolution. In terms of spatial dimension attention, a spatial attention weight map is generated through depthwise separable convolutional layers, focusing on the dominant precipitation band, strengthening its features, and weakening the feature responses of non-dominant precipitation bands; the dominant precipitation band includes strong convective cloud clusters and orographically lifted precipitation bands. In addition, the model replaces the traditional standard convolution with depthwise separable convolution, which reduces the coupling strength between channels while maintaining spatial continuity of features at different scales.

[0087] In one embodiment, to evaluate the performance advantages of the lightweight attention-enhanced LA-UNet model in satellite precipitation data-driven precipitation nowcasting, a benchmark model is further selected for comparative training with the lightweight attention-enhanced LA-UNet model. The benchmark models include the standard U-Net model, the UNet++ model, and the ConvLSTM model. The standard U-Net model employs a symmetric encoder-decoder structure, fusing multi-scale spatial features through skip connections, and is used to evaluate precipitation representation capabilities without introducing an explicit attention mechanism. The UNet++ model introduces nested and dense skip connections between nodes on top of the U-Net model and achieves deep supervision through a loss function, used to compare the differences between "dense connection-driven feature reuse" and the lightweight attention-enhanced strategy of this invention. The convolutional recursive temporal model ConvLSTM explicitly characterizes the temporal evolution of the precipitation field through a convolutional recursive structure, used to evaluate the difference in forecast performance between explicit temporal recursion and the lightweight attention-enhanced LA-UNet model of this invention. All benchmark models and the lightweight attention-enhanced LA-UNet model of this invention use the same loss function, optimization strategy and training configuration during training. Only the impact of differences in network structure on precipitation nowcasting performance is compared.

[0088] In one specific implementation, a benchmark model and the lightweight attention-enhanced LA-UNet model are selected for comparative training. This includes: evaluating the spatial distribution accuracy of precipitation prediction results for each model across multiple short-term forecast periods based on multiple precipitation thresholds. Specifically, to explore the role of the attention mechanism in precipitation prediction, a channel-spatial dual-dimensional attention module is embedded into the standard U-Net model and the UNet++ model, respectively, forming an attention-enhanced U-NetCBAM model and a UNet++CBAM model. Then, the performance of the attention-enhanced U-NetCBAM model, the UNet++CBAM model, the standard U-Net model, the UNet++ model, the ConvLSTM model, and the lightweight attention-enhanced LA-UNet model are compared together. Based on precipitation thresholds of varying intensities—low (e.g., light rain at 1.0 mm / h), medium (e.g., moderate rain at 5.0 mm / h), and high (e.g., heavy rain at 10.0 mm / h)—the precipitation prediction results of each model were validated across multiple short-term forecast periods, such as the predictions for the next 1, 2, and 3 hours. These precipitation thresholds can be determined according to meteorological industry standards for grading precipitation intensity, but this invention is not limited to them. Finally, the spatial distribution accuracy of each model was evaluated using probabilistic detectivity (POD), false alarm rate (FAR), critical success index (CSI), and F1 score. Figure 2 As shown, Figure 2 This diagram illustrates a comparison of precipitation nowcasts using various models, including the U-NetCBAM model with attention mechanism, the UNet++CBAM model, the standard U-Net model, the UNet++ model, the ConvLSTM model, and the lightweight attention-enhanced LA-UNet model of this invention. It shows the spatial distribution simulation results of precipitation calculated using satellite precipitation data based on satellite remote sensing, employing these models. The diagram only includes numerical precipitation information (represented by color bars on the right) and does not contain any sensitive geographical information such as geographic coordinates, place names, or administrative divisions, nor does it reflect any geographic boundaries or topographic features. As can be seen from the diagram, at the 15-minute forecast stage, all models can capture the overall development trend of the precipitation system well; however, as the forecast lead time increases, the uncertainty of the model predictions gradually accumulates, and spatial ambiguity intensifies. By 30 minutes, some models showed loss of precipitation details. At 60 minutes, while the standard U-Net and UNet++ models could still roughly locate precipitation areas, the edge structures and the core of heavy precipitation were significantly degraded. Notably, simply adding the CBAM attention module directly to the standard U-Net and UNet++ models did not improve performance; instead, it led to significant performance degradation under most precipitation threshold conditions. This indicates that the attention mechanism does not provide universal benefits for all network structures, and its effectiveness highly depends on its co-design with the overall feature modeling path. The ConvLSTM model showed the most significant performance degradation across different forecast timeframes, especially under high precipitation threshold scenarios, where its spatial localization ability and ability to preserve heavy precipitation structures were clearly insufficient. In contrast, the lightweight attention-enhanced LA-UNet model of this invention performed better across all forecast timeframes and precipitation thresholds. Although it also faced detail decay issues in long-term predictions, its overall spatial consistency and morphological restoration capabilities were significantly superior to the comparative models.

[0089] In one specific implementation, the selection of a benchmark model and the lightweight attention-enhanced LA-UNet model for comparative training also includes: evaluating the feature redundancy of different models at different network layers. Specifically, given that the ConvLSTM model has shown significantly inferior overall performance compared to convolutional models across various forecast lead times and precipitation thresholds, the comparative analysis focuses only on three types of models: the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model. By extracting the feature map tensors of the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model at shallow, mid, and deep layers, the features of each channel are flattened into a one-dimensional vector; then, the Pearson correlation coefficient between channels is calculated to construct a channel correlation matrix, and the differences in deep feature representation capabilities among the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model are analyzed. Figure 3 As shown, Figure 3 The figure shows a comparison of feature similarity matrices and box plots between shallow and deep channels for the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model of this invention. As can be seen from the figure, the lightweight attention-enhanced LA-UNet model retains a certain degree of channel correlation in the shallow layer (cbam1). As the network deepens, feature redundancy is significantly reduced in the middle layer (cbam3), while in the deep bottleneck layer (cbam5), correlation almost completely disappears except for the diagonal lines. This indicates that the CBAM attention mechanism effectively suppresses the accumulation of redundant features through adaptive recalibration of channel and spatial responses. In contrast, although the standard U-Net also shows a trend of gradually decreasing channel correlation from shallow to deep layers, its optimization is significantly limited. In the deep bottleneck layer (down4), a large number of significantly positively correlated channels can still be observed. Figure 3 (Light red band). The UNet++ model exhibits a large area of ​​strong positive correlation in the shallow layer (conv00), and this high correlation is not weakened as the network deepens. Instead, it is further strengthened in the deep bottleneck layer (conv40), indicating that the nested dense skip connections of the UNet++ model lead to excessive feature fusion in this task, which weakens the model's ability to learn discriminative semantic features.

[0090] In one specific implementation, the selection of a benchmark model and the lightweight attention-enhanced LA-UNet model for comparative training further includes: selecting typical regions of different categories within the target watershed and performing regional comparative analysis on each model within these typical regions. Specifically, a gradient-weighted class activation mapping method is used to visualize the typical regions of different categories that the standard U-Net model, UNet++ model, and lightweight attention-enhanced LA-UNet model focus on in the short-term forecast timeliness task; the output of the feature generation layer before the final convolutional layer of each model is selected as the analysis object, and attention heatmaps are generated and compared in the typical regions of different categories within the target watershed; wherein, the typical regions of different categories include plains, lakeside transition zones, and mountainous uplift zones. Figure 4 As shown, Figure 4 A visual comparison diagram of gradient-weighted class activation mappings for the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model of this invention is shown in the precipitation nowcasting task. Figure 4 The ad in the figure shows the actual precipitation values, standard U-Net model, UNet++ model, and lightweight attention-enhanced LA-UNet model predictions for typical region I (i.e., plain and valley area) in turn. Figure 4 The eh section sequentially displays the actual precipitation values ​​for typical region II (i.e., the lakeside transition zone), the predictions from the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model; Figure 4 The il columns sequentially display the actual precipitation values ​​for typical region III (i.e., the mountain uplift zone), the predictions from the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model. Figure 4The `mp` section sequentially displays the actual precipitation values, standard U-Net model, UNet++ model, and lightweight attention-enhanced LA-UNet model predictions for the composite region combining typical regions II and III. In typical region I, the lightweight attention-enhanced LA-UNet model accurately focuses on the high-albedo precipitation core, achieving stable positioning of the precipitation center. In typical region II, the attention distribution of the lightweight attention-enhanced LA-UNet model clearly delineates the banded precipitation structure formed by the movement of the weather system. In typical region III, the lightweight attention-enhanced LA-UNet model exhibits a highly sensitive and compact response zone to localized strong convection triggered by orographic uplift, highly consistent with the topographic gradient and precipitation intensity distribution. Particularly in the composite region of types II and III, where precipitation is driven by the coupling forcing of the lakeshore effect and orographic uplift, the lightweight attention-enhanced LA-UNet model demonstrates robust sensing capabilities, and its activation heatmap accurately locates convective cells at complex boundaries. In contrast, both the standard U-Net model and the UNet++ model exhibited generalized responses and blurred boundaries in the aforementioned scenarios, making it difficult to accurately capture key spatial features of precipitation under multi-factor coupling conditions. Verification showed that the lightweight attention-enhanced LA-UNet model demonstrated a more concentrated, stable spatial attention distribution that highly correlated with actual precipitation patterns across various typical study areas.

[0091] In one specific implementation, a benchmark model is selected and compared with the lightweight attention-enhanced LA-UNet model for training. This also includes evaluating the computational efficiency and resource utilization of various models, including the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model. Specifically, computational complexity analysis is performed on different models, including the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model. Verification shows that the lightweight attention-enhanced LA-UNet model has significantly fewer parameters and a smaller model size than the standard U-Net model and the UNet++ model, with an overall size approximately one-eighth that of the standard U-Net model. In terms of GFLOPs (Gross Floating-Point Operations) and inference latency, the lightweight attention-enhanced LA-UNet model exhibits the lowest performance among the three models, demonstrating good real-time prediction capabilities. Further resource utilization analysis shows that the peak GPU memory requirement of the lightweight attention-enhanced LA-UNet model is significantly lower than the baseline model, with memory usage reduced by nearly half compared to U-Net. Verification has shown that the lightweight attention-enhanced LA-UNet model of this invention has higher computational efficiency than the benchmark model, the lowest resource consumption, and good real-time prediction capabilities.

[0092] In one specific implementation, the selection of a benchmark model and the lightweight attention-enhanced LA-UNet model for comparative training further includes: evaluating whether the outputs of different models meet the basic physical and statistical constraints of hydrological simulation on rainfall input. Specifically, the short-term precipitation prediction results output by the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model are compared and evaluated; the hourly precipitation prediction results and cumulative precipitation processes for typical rainfall events are calculated for the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model, respectively; under precipitation scenarios dominated by different types of typical regions, the prediction result indicators of each model are calculated under multiple short-term forecast lead times, and the model error characteristics are compared and analyzed, wherein the prediction result indicators include: mean absolute error, correlation coefficient, and mean relative root mean square error. Figure 5 As shown, Figure 5 The diagram shows the nowcast precipitation events for various models, including the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model, at short-term forecast lead times of 1 hour, 2 hours, and 3 hours. Figure 5 In the first image, (a)-(d) represent hourly precipitation data for different monitoring points using the standard U-Net model. (a) shows hourly precipitation data for the first monitoring point, (b) shows hourly precipitation data for the second monitoring point, (c) shows hourly precipitation data for the third monitoring point, and (d) shows cumulative precipitation data for the third monitoring point. In the second image, (e)-(h) represent hourly precipitation data for different monitoring points using the UNet++ model. (e) shows hourly precipitation data for the first monitoring point, (f) shows hourly precipitation data for the second monitoring point, (g) shows hourly precipitation data for the third monitoring point, and (h) shows cumulative precipitation data for the third monitoring point. In the third image, (i)-(l) represent hourly precipitation data for different monitoring points using the LA-UNet model of this invention. (i) shows hourly precipitation data for the first monitoring point, (j) shows hourly precipitation data for the second monitoring point, (k) shows hourly precipitation data for the third monitoring point, and (l) shows cumulative precipitation data for the third monitoring point. Different models exhibit significant differences in terms of rainfall peak characterization, temporal alignment accuracy, and precipitation magnitude preservation. Under a 1-hour forecast period, the lightweight attention mechanism LA-UNet model proposed in this invention (…) Figure 5 The hourly precipitation forecast data of the U-Net model showed the highest agreement with the observed data, with its forecast curve exhibiting a high degree of consistency with the observations in terms of peak precipitation time and precipitation remission period, and exhibiting the smallest time offset. In contrast, the standard U-Net model ( Figure 5 The AC model systematically underestimates some of the peak precipitation, while the UNet++ model ( Figure 5 The hourly forecasts for (e.g.) fluctuate significantly and lack stability.

[0093] Regarding the prediction of cumulative precipitation, the cumulative precipitation curve of the LA-UNet model with the lightweight attention mechanism proposed in this invention ( Figure 5 The cumulative precipitation curves of the l-model best match the observed values ​​across all forecast periods, indicating its strongest constraint on total precipitation. In contrast, the standard U-Net model ( Figure 5 d) tends to underestimate cumulative precipitation at certain stages, UNet++ model ( Figure 5 The cumulative curve of the h) model exhibits more significant instability, failing to meet the stability requirements of hydrologically driven inputs. Different models show significant differences in rainfall peak characterization, time-series alignment accuracy, and precipitation magnitude preservation. The lightweight attention-enhanced LA-UNet model best matches the observed hourly precipitation forecast at a 1-hour lead time, with its prediction curve maintaining high consistency with observations at the rainfall peak and during rainfall interruptions, exhibiting minimal time shift. In contrast, the standard U-Net model systematically underestimates some rainfall peaks, while the UNet++ model shows larger fluctuations in hourly rainfall predictions, indicating relatively insufficient stability. The differences between models are even more significant at the level of cumulative precipitation. The lightweight attention-enhanced LA-UNet model of this invention shows the best fit between its cumulative precipitation curves and observed values ​​across all lead times, indicating its strongest constraint on total precipitation. In comparison, the standard U-Net model underestimates cumulative precipitation in some stages, while the UNet++ model's cumulative curve exhibits more significant instability. Figure 6 A comparative diagram of the flood nowcasting process for various models under different lead times is shown. Figure 6 The diagram illustrates the prediction performance of the standard U-Net model, the UNet++ model, and the LA-UNet model of this invention under short-term fluctuation scenarios. Figure 6 The middle df diagram illustrates the prediction performance of the standard U-Net model, the UNet++ model, and the LA-UNet model of this invention in the late rising scenario. Figure 6The diagram illustrates the forecasting performance of the standard U-Net model, the UNet++ model, and the LA-UNet model of this invention under small-scale flood scenarios. Looking at three different flood scenarios (short-term fluctuation type, late-stage rising type, and small-scale process), the forecasting performance of all models slightly decreases with the extension of the lead time. However, the lightweight attention-enhanced LA-UNet model of this invention shows better fit between its forecast curve and the observed flood (black solid line) in all scenarios, especially under long-term lead times of 2h and 3h, where its forecast results more closely follow the fluctuations of the actual process, while the forecast curves of the standard U-Net model and the UNet++ model show more significant deviations from the observed values. Simultaneously, the correspondence between changes in rainfall input (grey bars) and the flood process also reflects the simulation ability of each model for the rainfall-flood response relationship. The lightweight attention-enhanced LA-UNet model is also more accurate in fitting this response. Overall, the method of this invention demonstrates superior stability and accuracy in flood nowcasting for different flood scenarios.

[0094] Finally, the training process is compared between the above-mentioned benchmark model and the lightweight attention-enhanced LA-UNet model of the present invention. A short-term precipitation prediction model is constructed based on the lightweight attention-enhanced LA-UNet model. The preprocessed satellite precipitation data is input into the model, and the short-term satellite precipitation prediction results are output.

[0095] In one embodiment, in step S3, the short-term satellite precipitation prediction results are further input into a semi-distributed hydrological model. The semi-distributed hydrological model serves as the physical layer of the overall architecture of the short-term precipitation prediction model-hydrological model, simulates and generates a watershed-scale flood response process, and outputs predicted flood process data.

[0096] Specifically, the preprocessed satellite precipitation data and the short-term satellite precipitation prediction results are input into the semi-distributed hydrological model, such as the HEC-HMS (The Hydrologic Engineering Center's-Hydrologic Modeling System) hydrological model. This is used to transform the satellite precipitation nowcast results into a watershed-scale flood response process, realizing the physical constraints of precipitation extrapolation information on runoff formation and flood evolution.

[0097] The river channel calculation process of the semi-distributed hydrological model follows the law of conservation of mass to ensure the physical consistency of the simulation. The governing equation is shown in equation (1):

[0098] (1),

[0099] In the formula, andO These represent inflow and outflow respectively. S Represents the potential retained water volume of a sub-basin. t This is a time parameter.

[0100] Then, based on the acquired basic geographic data of the target watershed, the boundaries of sub-watersheds are divided in the target watershed, and key topographic features such as river network, watershed slope, and slope length are extracted to divide several sub-watersheds; wherein, the basic geographic data includes topography, land use type, and soil data.

[0101] Specifically, the runoff generation process of each sub-basin is simulated, and the infiltration excess rainfall is calculated using the SCS (Soil Conservation Service) curve number method, as shown in equation (2):

[0102] (2),

[0103] In the formula, P e This refers to rainfall exceeding the infiltration limit. P Total rainfall, This is the initial loss amount; The maximum potential water retention capacity is estimated from soil data.

[0104] The above-mentioned SCS curve number method calculation formula has been widely verified to be applicable to natural watershed runoff calculation, and the initial loss calculation formula... ,in This is the initial loss amount. S The potential retained water volume of the sub-basin is represented by a coefficient of 0.2, which is a commonly used empirical coefficient in the field of hydrology.

[0105] The confluence process of each sub-basin was simulated. The Clark unit hydrograph method was used at the sub-basin scale, and the Muskingan method was used at the river network scale to calculate the river storage capacity, as shown in equation (3):

[0106] (3),

[0107] In the formula, S For the potential retained water volume of the sub-basin, The storage time constant is initially estimated using topographic data using the Muskingan method. These are the Muskingan weighting coefficients, with values ​​ranging from 0 to 1. ≤0.5 is used to reflect the translation and flattening characteristics of flood waves in the river channel.

[0108] Furthermore, by driving the runoff generation and runoff calculation of each sub-basin through spatial weight allocation, a dynamic coupling relationship from precipitation nowcasting to flood response simulation is constructed.

[0109] Using the measured flood event data from hydrological stations within the target watershed, the runoff generation and confluence parameters of the semi-distributed hydrological model are calibrated and validated. The measured flood event data includes: measured peak flow, measured total flood volume, and measured peak time for each time period; the measured peak flow for all time periods is summarized into a measured flow sequence. The semi-distributed hydrological model outputs predicted flood event data for multiple forecast lead times. The predicted flood event data includes: predicted peak flow, predicted total flood volume, predicted peak time, and the predicted peak flow for all time periods summarized into a predicted flow sequence. Table 1 shows the HEC-HMS model construction and parameter calibration process in an example.

[0110] Table 1 HEC-HMS Model Construction and Parameter Calibration

[0111]

[0112] Furthermore, in one embodiment, a multi-index evaluation system is constructed to evaluate the short-term precipitation prediction model and the semi-distributed hydrological model.

[0113] Specifically, as described in the training process of the aforementioned short-term precipitation prediction model, the probability detection rate, false alarm rate, critical success index, and F1 score are used as the first evaluation indicators to assess the accuracy of the spatial distribution prediction of precipitation in the short-term precipitation prediction model. As shown in equations (4)-(7), equations (4)-(7) respectively refer to the probability detection rate (… POD ), false alarm rate ( FAR Key Success Index ( CSI )and F1 Fraction.

[0114] (4),

[0115] (5),

[0116] (6),

[0117] (7),

[0118] In the formula, TP This indicates the number of grid cells where both the measured rainfall data and the predicted rainfall result exceed the rainfall threshold. FP This indicates the number of grid cells where the predicted precipitation exceeds the precipitation threshold, but the actual measured rainfall data does not. FN This indicates the number of grid cells where the measured rainfall data exceeds the precipitation threshold, but the predicted rainfall does not. This precipitation threshold can be determined based on meteorological industry standards for grading precipitation intensity, but this invention is not limited to this.

[0119] The F1 score is a composite score that considers both probability and detection rate. POD ) and false alarm rate ( FAR Evaluation metrics are used to balance the accuracy and precision of forecasts.

[0120] With mean absolute error ( ), correlation coefficient ( ), mean relative root mean square error ( As the second evaluation index, the short-term precipitation prediction accuracy of the short-term precipitation prediction model is evaluated based on the measured rainfall data from rain gauge stations and the short-term satellite precipitation prediction results. Specifically, it is shown in equations (8)-(10).

[0121] (8),

[0122] (9),

[0123] (10)

[0124] in This is actual rainfall data measured at rain gauge stations. This is a short-term satellite precipitation forecast result. , These represent the average of measured rainfall data for each time period and the average of short-term satellite precipitation forecasts for each time period, respectively. , , N The number of samples is denoted as 'sample size'. The closer the correlation coefficient is to 1, the smaller the mean absolute error and mean relative root mean square error, the higher the prediction accuracy of the short-term precipitation prediction model, and the more accurate the satellite precipitation data obtained.

[0125] Using the relative error of peak flow ( ), relative error of total flood volume ( Peak occurrence time error () Nash efficiency coefficient ( NSE The measured flood process data from hydrological stations within the target watershed are used as the third evaluation index. This data is compared with the predicted flood process data output by the semi-distributed hydrological model to assess the accuracy of the flood process simulation of the semi-distributed hydrological model. Specifically, this is shown in equations (11)-(14).

[0126] (11),

[0127] (12)

[0128] (13)

[0129] (14)

[0130] in , These are the measured total flood volume and the predicted total flood volume, respectively. , These are the measured peak time and the predicted peak time, respectively. This represents the average measured peak flow rate for each time period. The relative error of the peak flow rate is ( ). ), relative error of total flood volume ( The smaller the absolute value, the smaller the peak occurrence time error. The closer to 0, NSE The closer the value is to 1, the better the forecast result.

[0131] The multi-indicator evaluation system includes a first evaluation indicator, a second evaluation indicator, and a third evaluation indicator.

[0132] In summary, the intelligent forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing provided by this invention constructs a coupled model of precipitation prediction and hydrological response. This achieves both accurate prediction of short-term satellite precipitation in complex watersheds and effective constraint of precipitation information on flood evolution. Specifically, the satellite precipitation data of the target watershed observed by multi-channel satellites is preprocessed and input into the short-term precipitation prediction model, outputting short-term satellite precipitation prediction results. The short-term precipitation prediction model is built based on a lightweight attention-enhanced LA-UNet model, which can adaptively identify fine precipitation structures such as strong convective clouds and orographic lifting zones, providing high-precision short-term precipitation input for flood forecasting. Then, the short-term satellite precipitation prediction results are input into a semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, outputting predicted flood process data. This method systematically quantifies the propagation and amplification effects of short-term precipitation forecasting errors during flood response, thereby effectively suppressing the cumulative effect of errors in runoff generation and confluence calculations. It not only solves the accuracy problem of short-term precipitation forecasting in complex watersheds, but also significantly improves the accuracy and reliability of short-term flood warnings in complex watersheds such as large lake basins and mountain-plain composite watersheds through error suppression mechanisms. It also enhances the availability and stability of flood forecasts, providing a systematic solution that balances physical consistency, computational efficiency, and interpretability.

[0133] It should be noted that the scope of the methods in the embodiments of the present invention is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be applied, omitted, or combined. In addition, features described with reference to certain examples may be combined in other examples.

[0134] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these modifications are within the protection scope of the present invention.

Claims

1. A smart forecasting method for short-term flood response in complex watersheds based on satellite precipitation sensing, characterized in that, Include: Acquire satellite precipitation data of the target watershed from multi-channel satellite observations and perform preprocessing; The preprocessed satellite precipitation data is input into a short-term precipitation prediction model, which outputs short-term satellite precipitation prediction results; wherein, the short-term precipitation prediction model includes: The perception layer is used to organize the preprocessed satellite precipitation data into a single-channel raster time series tensor. The construction process of the perception layer includes: The preprocessed satellite precipitation data is organized in a single-channel raster time series format, with each time step corresponding to a quantitative precipitation estimation precipitation field, characterizing the continuous evolution of precipitation. The final result is a single-channel raster temporal tensor with dimensions T×H×W×C, where C=1 indicates single-channel input, T is the time series length, H represents the spatial height dimension of the precipitation field, and W represents the spatial width dimension of the precipitation field. The cognitive layer is constructed based on the lightweight attention-enhanced LA-UNet model, which uses the single-channel raster temporal tensor as input to generate the short-term satellite precipitation prediction results. Construct a cognitive layer, including: Based on the standard U-Net model, channel-space dual-dimensional attention modules are embedded in multiple network layers at different scales of the encoder, and all standard convolutional layers in the encoder and decoder are replaced with depthwise separable convolutional layers to obtain the lightweight attention-enhanced LA-UNet model. The dual-dimensional attention module is used to extract highly correlated multi-channel features from the single-channel raster temporal tensor for adaptive recalibration during feature downsampling and semantic abstraction. Among them, the channel-space dual-dimensional attention module compresses the spatial dimension through global average pooling to generate channel attention weights, highlighting radiation and cloud system information channels that are closely related to precipitation formation and evolution; In terms of spatial attention, a spatial attention weight map is generated through a depthwise separable convolutional layer to focus on the dominant precipitation belt, enhance the features of the dominant precipitation belt, and weaken the feature responses of non-dominant precipitation belts; wherein, the dominant precipitation belt includes strong convective cloud clusters and topographically lifted precipitation belts. The short-term satellite precipitation prediction results are input into a semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, and output predicted flood process data.

2. The method according to claim 1, characterized in that, It also includes: selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model for training, wherein the baseline model includes the standard U-Net model, the UNet++ model, and the ConvLSTM model, and includes: Based on multiple precipitation thresholds, the spatial distribution accuracy of precipitation prediction results from various models under multiple short-term forecast lead times is evaluated, including: The channel-space dual-dimensional attention module is embedded into the standard U-Net model and the UNet++ model respectively to form the U-NetCBAM model and the UNet++CBAM model with attention mechanism; The performance of the U-NetCBAM model with attention mechanism, UNet++CBAM model, standard U-Net model, UNet++ model, ConvLSTM model, and the lightweight attention-enhanced LA-UNet model were compared. Based on low, medium, and high precipitation thresholds, the precipitation prediction results of each model under multiple short-term forecast lead times were verified. The accuracy of precipitation spatial distribution of each model is evaluated based on the evaluation indicators of probability detection rate, false alarm rate, critical success index and F1 score.

3. The method according to claim 2, characterized in that, The training process involves selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model, and also includes: Evaluating the feature redundancy of different models at different network layers includes: Extract the feature map tensors of the standard U-Net model, UNet++ model and the lightweight attention-enhanced LA-UNet model at shallow, middle and deep layers, and flatten the features of each channel into a one-dimensional vector; Calculate the Pearson correlation coefficient between channels, construct the channel correlation matrix, and analyze the differences in deep feature representation capabilities among the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model.

4. The method according to claim 2, characterized in that, The training process involves selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model, and also includes: Selecting typical regions of different categories within the target watershed, a regional comparative analysis of each model is conducted in these typical regions, including: The gradient-weighted class activation mapping method is used to visualize the typical regions of different categories that the standard U-Net model, UNet++ model and lightweight attention-enhanced LA-UNet model focus on in the short-term forecast timeliness task. The output of the feature generation layer before the final convolutional layer of each model is selected as the analysis object, and attention heatmaps are generated for comparison in different typical regions of the target watershed; wherein, the different typical regions include plains, lakeside transition zones and mountain uplift zones.

5. The method according to claim 2, characterized in that, The training process involves selecting a baseline model and comparing it with the lightweight attention-enhanced LA-UNet model, and also includes: Regional comparative analysis of each model was conducted in typical regions of the different categories, including: Evaluate whether the outputs of different models meet the basic physical and statistical constraints of hydrological simulation on rainfall input, including: The short-term precipitation prediction results output by the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model are compared and evaluated. Hourly precipitation prediction results and cumulative precipitation processes for typical rainfall events were calculated for the standard U-Net model, the UNet++ model, and the lightweight attention-enhanced LA-UNet model, respectively. Under precipitation scenarios dominated by typical regions of different categories, the prediction results indicators of each model are calculated under multiple short-term forecast lead times, and the error characteristics of the models are compared and analyzed. The prediction results indicators include: mean absolute error, correlation coefficient, and mean relative root mean square error.

6. The method according to claim 1, characterized in that, The short-term satellite precipitation prediction results are input into a semi-distributed hydrological model to simulate and generate a watershed-scale flood response process, outputting predicted flood process data, including: The preprocessed satellite precipitation data and the short-time satellite precipitation prediction results are input together into the semi-distributed hydrological model; Using the measured flood process data from hydrological stations within the target watershed, the runoff generation and confluence parameters of the semi-distributed hydrological model are calibrated and validated; wherein, the measured flood process data includes: measured peak flow, measured total flood volume, and measured peak occurrence time for each time period; The semi-distributed hydrological model outputs the predicted flood process data under multiple forecast timeframes. The predicted flood process data includes: predicted peak flow, predicted total flood volume, and predicted peak time for each time period. The river channel calculation process of the semi-distributed hydrological model follows the law of conservation of mass, and the governing equation is shown in equation (1): （1）， In the formula, and O These represent inflow and outflow respectively. S Represents the potential retained water volume of a sub-basin. t For time parameters; Based on the acquired basic geographic data of the target watershed, the boundaries of sub-watersheds are divided within the target watershed, and key topographic features such as river network, watershed slope, and slope length are extracted to divide several sub-watersheds; wherein, the basic geographic data includes topography, land use type, and soil data; The runoff generation process of each sub-basin was simulated, and the infiltration excess rainfall was calculated using the SCS curve number method, as shown in equation (2): （2）， In the formula, P e This refers to rainfall exceeding the infiltration limit. P Total rainfall, This is the initial loss amount. This represents the maximum potential water retention capacity. The confluence process of each sub-basin was simulated. The Clark unit hydrograph method was used at the sub-basin scale, and the Muskingan method was used at the river network scale to calculate the river channel storage, as shown in equation (3): （3）， In the formula, The storage time constant; These are the Muskingan weighting coefficients, with values ​​ranging from 0 to 1. ≤0.5 is used to reflect the translation and flattening characteristics of flood waves in the river channel.

7. The method according to claim 6, characterized in that, It also includes: constructing a multi-index evaluation system to evaluate the short-term precipitation prediction model and the semi-distributed hydrological model, including: The probability detection rate, false alarm rate, critical success index, and F1 score are used as the first evaluation indicators to assess the accuracy of the spatial distribution prediction of precipitation in the short-term precipitation prediction model. The short-term precipitation prediction accuracy of the short-term precipitation prediction model is evaluated based on the measured rainfall data from rain gauge stations and the short-term satellite precipitation prediction results, using mean absolute error, correlation coefficient, and mean relative root mean square error as the second evaluation index. The relative error of peak flow, the relative error of total flood volume, the peak occurrence time error, and the Nash efficiency coefficient are used as the third evaluation index. The measured flood process data of hydrological stations in the target watershed are used as a reference benchmark and compared with the predicted flood process data output by the semi-distributed hydrological model to evaluate the flood process simulation accuracy of the semi-distributed hydrological model. The multi-indicator evaluation system includes a first evaluation indicator, a second evaluation indicator, and a third evaluation indicator.

8. The method according to claim 1, characterized in that, Preprocessing of satellite precipitation data includes: The satellite precipitation data were subjected to geometric correction, orthorectification, and spatial mapping and terrain correction. The corrected satellite precipitation data is uniformly projected onto a geographic coordinate system grid with a preset resolution, and then subjected to numerical normalization and time synchronization processing. The corrected satellite precipitation data is then regionally cropped. After removing invalid and outlier values ​​from the regionally cropped satellite precipitation data, it is divided into training, validation, and test sets according to a preset ratio.

Images