A surface runoff cause identification method and device based on deep learning and explainability analysis

Abstract

The application discloses a kind of based on deep learning and explainability analysis surface runoff cause identification method and device, to solve the technical problem that the existing surface runoff cause identification method leads to lower identification precision.Method includes obtaining meteorological hydrological factor time series data, and meteorological hydrological factor time series data is input to based on deep learning surface runoff cause identification model;Based on deep learning surface runoff cause identification model includes one-dimensional convolution module, bidirectional long short-term memory network, based on the identification module of multi-head attention mechanism;One-dimensional convolution module is used to carry out convolution operation to meteorological hydrological factor time series data, and output local response feature sequence;Target comprehensive feature sequence is generated by bidirectional long short-term memory network to local response feature sequence executes bidirectional processing operation;Target comprehensive feature sequence is input to based on the identification of multi-head attention mechanism module and is identified, and output target surface runoff cause identification result.

Description

Technical Field

[0001] This invention relates to the field of surface runoff evolution technology, and in particular to a method and apparatus for identifying the causes of surface runoff based on deep learning and interpretability analysis. Background Technology

[0002] Surface runoff is the water flow along the surface slope after precipitation. As the main component of runoff at the outlet section of a watershed, it directly determines the flood formation process and the spatial and temporal distribution of water resources. Its dynamic changes have a decisive impact on the water cycle and ecosystem balance. This physical process is coupled with multiple factors such as topographic slope, soil permeability, vegetation cover, and precipitation characteristics. These factors generate synergistic effects through nonlinear interactions, resulting in significant spatial heterogeneity and temporal variability in the runoff generation mechanism.

[0003] Against the backdrop of global warming, the climate system is undergoing a profound reorganization in precipitation intensity, frequency, and spatial distribution. This reorganization alters the storm-runoff response relationship, resulting in a multiplier effect on the probability and magnitude of extreme runoff events. Therefore, quantitatively analyzing the runoff evolution patterns driven by both climate and underlying surface factors is not only the scientific basis for optimal water resource allocation but also a crucial prerequisite for building a resilient flood control system. The research findings can be directly transformed into the algorithmic core of flood and drought disaster risk early warning systems.

[0004] Most existing methods for identifying the causes of surface runoff employ a single LSTM (Long Short-Term Memory) network, BiLSTM (Bidirectional Long Short-Term Memory) network, or XGBoost (eXtreme Gradient Boosting) model to identify the causes of surface runoff. However, these models have limitations in balancing data fitting accuracy and the analysis of nonlinear relationships between variables, and fail to deeply analyze the interactions between meteorological, hydrological, and geomorphological factors, resulting in low identification accuracy. Summary of the Invention

[0005] This invention provides a method and apparatus for identifying the causes of surface runoff based on deep learning and interpretability analysis, which solves the technical problem of low identification accuracy caused by existing methods for identifying the causes of surface runoff.

[0006] The first aspect of this invention provides a method for identifying the causes of surface runoff based on deep learning and interpretability analysis, comprising:

[0007] Acquire time series data of meteorological and hydrological factors, and input the time series data of meteorological and hydrological factors into a surface runoff cause identification model based on deep learning; the surface runoff cause identification model based on deep learning includes a one-dimensional convolutional module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism;

[0008] The one-dimensional convolution module is used to perform convolution operations on the meteorological and hydrological factor time series data to output a local response feature sequence;

[0009] The local response feature sequence is processed bidirectionally by the bidirectional long short-term memory network to generate the target comprehensive feature sequence.

[0010] The target's comprehensive feature sequence is input into the recognition module based on the multi-head attention mechanism for recognition, and the target's surface runoff cause recognition result is output.

[0011] Optionally, the one-dimensional convolution module includes a one-dimensional convolutional layer, a random deactivation layer, and a residual connection layer; the step of using the one-dimensional convolution module to perform convolution operations on the meteorological and hydrological factor time series data to output a local response feature sequence includes:

[0012] The one-dimensional convolutional layer is used to perform one-dimensional convolution on the time series data of the meteorological and hydrological factors, and outputs a one-dimensional convolutional feature sequence.

[0013] The one-dimensional convolutional feature sequence is input into the random deactivation layer to perform a random deactivation operation, thereby generating a random deactivation feature sequence.

[0014] The random inactivation feature sequence and the meteorological and hydrological factor time series data are added together through the residual connection layer to generate a local response feature sequence.

[0015] Optionally, the bidirectional long short-term memory network includes a bidirectional long short-term memory network layer, a normalization layer, and a residual connection layer; the step of performing bidirectional processing operations on the local response feature sequence through the bidirectional long short-term memory network to generate the target comprehensive feature sequence includes:

[0016] The local response feature sequence is used as the input to the bidirectional long short-term memory network layer, and a bidirectional temporal feature sequence is output.

[0017] The bidirectional temporal feature sequence is standardized using the standardization layer to output a standardized temporal feature sequence;

[0018] The standardized temporal feature sequence and the local response feature sequence are added together through the residual connection layer to output the target comprehensive feature sequence.

[0019] Optionally, the recognition module based on the multi-head attention mechanism includes a multi-head attention mechanism layer, a flattening layer, and a fully connected layer; the step of inputting the target's comprehensive feature sequence into the recognition module based on the multi-head attention mechanism for recognition and outputting the target surface runoff cause recognition result includes:

[0020] The target's comprehensive feature sequence is used as the input to the multi-head attention mechanism layer, and a multi-dimensional tensor is output.

[0021] The multidimensional tensor is flattened by the flattening layer to output a one-dimensional tensor;

[0022] The one-dimensional tensor is mapped using the fully connected layer to generate the target surface runoff cause identification result.

[0023] Optionally, the training process of the deep learning-based surface runoff causation identification model specifically includes:

[0024] Multiple meteorological and hydrological factor time series data for model training are obtained, and each of the meteorological and hydrological factor time series data for model training is input into the initial deep learning-based surface runoff cause identification model, and the training surface runoff cause identification results corresponding to each of the meteorological and hydrological factor time series data for model training are output.

[0025] Calculate the total sum of squares based on multiple training surface runoff causation identification results;

[0026] The weights corresponding to the actual runoff observations are output based on the actual runoff observations corresponding to the training surface runoff cause identification results using a two-segment weighting method and a three-segment weighting method.

[0027] Based on the identification results of the surface runoff causes of each training exercise, the actual observed values ​​of each runoff, and the corresponding weights, calculate the weighted sum of squared residuals.

[0028] Substitute the weighted residual sum of squares and the total sum of squares into the preset loss function and take the derivative to output the model gradient;

[0029] The model parameters of the initial deep learning-based surface runoff causation identification model are updated using the model gradient to determine the intermediate deep learning-based surface runoff causation identification model, and the number of model updates is counted in real time.

[0030] Determine whether the number of model updates has reached the preset training count threshold;

[0031] If so, then the feature importance score is calculated based on the multiple training surface runoff cause identification results output by the intermediate deep learning-based surface runoff cause identification model.

[0032] Based on the feature importance score, the feature contribution percentage is calculated;

[0033] The intermediate deep learning-based surface runoff cause identification model is used as the trained deep learning-based surface runoff cause identification model.

[0034] Optionally, the preset loss function is specifically:

[0035] ;

[0036] in, The loss value corresponding to the preset loss function; This is the weighted sum of squared residuals; The total sum of squares; Value 10 -8 .

[0037] A second aspect of the present invention provides a surface runoff causation identification device based on deep learning and interpretability analysis, comprising:

[0038] The acquisition module is used to acquire time series data of meteorological and hydrological factors and input the time series data of meteorological and hydrological factors into a deep learning-based surface runoff causation identification model; the deep learning-based surface runoff causation identification model includes a one-dimensional convolution module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism.

[0039] The convolution module is used to perform convolution operations on the meteorological and hydrological factor time series data using the one-dimensional convolution module, and output local response feature sequences.

[0040] The bidirectional module is used to perform bidirectional processing operations on the local response feature sequence through the bidirectional long short-term memory network to generate the target comprehensive feature sequence;

[0041] The identification module is used to input the comprehensive feature sequence of the target into the identification module based on the multi-head attention mechanism for identification, and output the identification result of the cause of surface runoff of the target.

[0042] A computer device provided in a third aspect of the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any of the preceding claims.

[0043] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any of the preceding claims.

[0044] The fifth aspect of the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any of the preceding claims.

[0045] As can be seen from the above technical solutions, the present invention has the following advantages:

[0046] The present invention provides a method for identifying the causes of surface runoff based on deep learning and interpretability analysis. First, time-series data of meteorological and hydrological factors are acquired and input into a deep learning-based surface runoff cause identification model. This model includes a one-dimensional convolutional module, a bidirectional long short-term memory network, and a recognition module based on a multi-head attention mechanism. Next, the one-dimensional convolutional module performs convolution operations on the meteorological and hydrological factor time-series data, outputting a local response feature sequence. The bidirectional long short-term memory network performs bidirectional processing on the local response feature sequence to generate a comprehensive target feature sequence. Finally, the comprehensive target feature sequence is input into the recognition module based on a multi-head attention mechanism for identification, outputting the target surface runoff cause identification result. Based on this scheme, the present invention enhances the ability to extract local mutation features through a one-dimensional convolutional module and combines the bidirectional long short-term memory network and the multi-head attention mechanism to synergistically integrate dynamic weight allocation capabilities, thereby improving recognition accuracy. Attached Figure Description

[0047] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating the steps of a surface runoff cause identification method based on deep learning and interpretability analysis provided in Embodiment 1 of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of the surface runoff causation identification model based on deep learning provided in Embodiment 1 of the present invention;

[0050] Figure 3 The flowchart illustrates the steps of training a deep learning-based surface runoff causation identification model according to Embodiment 2 of the present invention.

[0051] Figure 4 This is a comparison of the runoff simulation performance of the CNN-BiLSTM-ATT (Convolutional Neural Network-Bidirectional Long Short-Term Memory-Attention) model and the CNN-BiLSTM (Convolutional Neural Network-Bidirectional Long Short-Term Memory) model provided in Embodiment 2 of the present invention in a typical study area of ​​the Pearl River Basin.

[0052] Figure 5 This is a structural block diagram of a surface runoff causation identification device based on deep learning and interpretability analysis, provided in Embodiment 3 of the present invention. Detailed Implementation

[0053] This invention provides a method and apparatus for identifying the causes of surface runoff based on deep learning and interpretability analysis, which solves the technical problem of low identification accuracy caused by existing surface runoff cause identification methods.

[0054] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0055] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a method for identifying the causes of surface runoff based on deep learning and interpretability analysis, as provided in Embodiment 1 of the present invention.

[0056] This invention provides a method for identifying the causes of surface runoff based on deep learning and interpretability analysis, comprising:

[0057] Step 101: Obtain time series data of meteorological and hydrological factors, and input the time series data of meteorological and hydrological factors into the surface runoff cause identification model based on deep learning; the surface runoff cause identification model based on deep learning includes a one-dimensional convolutional module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism.

[0058] It should be noted that in surface runoff and flood studies, many studies focus more on the watershed outlet process rather than the regional average flow. The outlet flow process reflects the scale and intensity of floods, facilitating a more accurate assessment of flood risk. Therefore, this scheme uses the surface runoff time series at the watershed outlet as the objective (output feature) to be solved by the model. A deep learning-based surface runoff causation identification model is constructed using time series of k natural factors, such as rainfall and evaporation (potentially influencing runoff), with a length of n, as the driving factors (input features). During the modeling process, the input features of the model are the observed values ​​of several hydro-meteorological driving factors at each time t, forming a multidimensional time series matrix X of length n. t This refers to time series data of meteorological and hydrological factors; correspondingly, the model's output feature is the sub-basin outlet runoff value y at the current time t. t That is, the result of identifying the causes of target surface runoff, which can be represented as: , ,in, This represents the observed value of the k-th input feature at time t, a feature (rainfall, temperature, etc.) in the time series data of meteorological and hydrological factors. This represents the joint observation value of all input features at time t, i.e., the time series data of meteorological and hydrological factors; This is a deep learning-based model for identifying the causes of surface runoff.

[0059] Further, please refer to Figure 2 Deep learning models excel at capturing both abstract and hidden information in data, demonstrating exceptional capabilities in data-intensive analysis and mining tasks. This invention proposes a deep learning framework integrating CNN (Convolutional Neural Network), BiLSTM (Bidirectional Long Short-Term Memory), and an attention mechanism for comprehensively analyzing the impact of meteorological, hydrological, and geomorphological factors on surface runoff. Compared to traditional process-driven models, this framework achieves higher accuracy in simulating surface runoff and more accurately characterizes the nonlinear relationship between meteorological and hydrological conditions and runoff. The input layer of the CNN-BiLSTM-ATT framework consists of k-dimensional meteorological and hydrological factor time-series data (of length n). Specifically, the deep learning-based surface runoff causation identification model comprises a one-dimensional convolutional module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism. The bidirectional long short-term memory network includes a bidirectional long short-term memory network layer, a normalization layer, and a residual connection layer; the multi-head attention mechanism-based identification module includes a multi-head attention mechanism layer, a flattening layer, and a fully connected layer.

[0060] In this embodiment, the deep learning-based surface runoff causation identification model proposed in this invention, namely the deep learning model CNN-BiLSTM-ATT, effectively achieves joint mining of longitudinal (temporal dependence) and lateral (inter-variable correlation) features of the input feature sequence by integrating CNN, BiLSTM, and attention mechanisms, and enhances the ability to identify abnormal or extreme events. This structure demonstrates significant advantages in revealing the complex driving mechanisms of runoff formation and reproducing extreme runoff processes.

[0061] Step 102: Use a one-dimensional convolution module to perform convolution operations on the meteorological and hydrological factor time series data and output the local response feature sequence.

[0062] Specifically, step 102 may include the following sub-steps S21-S23:

[0063] Step S21: Use a one-dimensional convolutional layer to perform one-dimensional convolution on the meteorological and hydrological factor time series data, and output a one-dimensional convolutional feature sequence;

[0064] Step S22: Input the one-dimensional convolutional feature sequence into the random deactivation layer to perform random deactivation operation and generate a random deactivation feature sequence;

[0065] Step S23: Add the random inactivation feature sequence and the meteorological and hydrological factor time series data through the residual connection layer to generate the local response feature sequence.

[0066] It should be noted that the one-dimensional convolutional neural network (1D-CNN) module captures short-term temporal abrupt changes and spatial local response features in the input layer, such as local peaks and their durations in rainfall, outliers in temperature, and abrupt patterns in evaporation rates. The one-dimensional convolutional module includes one-dimensional convolutional layers, dropout layers (random deactivation layers), and residual layers (residual connection layers). The p filters in the one-dimensional convolutional layer slide across each channel of the input time series, capturing p features. The dropout layer prevents overfitting, while the residual layer accelerates convergence during model training. CNNs learn from data by leveraging their hierarchical layer structure, gaining a deep understanding of the layers while comprehending complex relationships within the input characteristics. Therefore, they are suitable for placement near the input layer to provide an initial understanding of time series with complex variations in natural conditions such as rainfall and temperature.

[0067] Step 103: Perform bidirectional processing on the local response feature sequence through a bidirectional long short-term memory network to generate the target comprehensive feature sequence.

[0068] Specifically, step 103 may include the following sub-steps S31-S33:

[0069] Step S31: Use the local response feature sequence as the input to the bidirectional long short-term memory network layer, and output the bidirectional temporal feature sequence;

[0070] Step S32: Standardize the bidirectional temporal feature sequence using a standardization layer and output the standardized temporal feature sequence;

[0071] Step S33: Add the standardized temporal feature sequence and the local response feature sequence through the residual connection layer to output the target comprehensive feature sequence.

[0072] It should be noted that a Bidirectional Long Short-Term Memory (BiLSTM) module further mines the general nonlinear functional relationship between historical rainfall, evaporation, vegetation richness, and current surface runoff. This module consists of BiLSTM layers, Batch Normalization (BN) layers, and residual layers. BiLSTM excels at capturing long-term dependencies in sequences, extracting 2×q features for each time step. BiLSTM is a combination of forward LSTM and backward LSTM. The bidirectional network enables recursive response training to the hidden states at the beginning and end of the sequence. This module improves training stability and enhances the ability to remember cross-seasonal and cross-annual input data, providing the model with the ability to analyze the "pre-event driving factors" of extreme events.

[0073] Step 104: Input the target's comprehensive feature sequence into the recognition module based on the multi-head attention mechanism for recognition, and output the target's surface runoff cause recognition result.

[0074] Specifically, step 104 may include the following sub-steps S41-S43:

[0075] Step S41: Use the target's comprehensive feature sequence as input to the multi-head attention mechanism layer and output a multi-dimensional tensor;

[0076] Step S42: Flatten the multidimensional tensor using a flattening layer to output a one-dimensional tensor;

[0077] Step S43: Map the one-dimensional tensor using a fully connected layer to generate the target surface runoff cause identification result.

[0078] It should be noted that multi-head attention mechanism is a type of self-attention mechanism. Through parallel processing of multiple heads, multi-head self-attention can capture more diverse features, improving the model's expressive power and generalization performance. Based on the aforementioned feature extraction, the multi-head self-attention mechanism assigns higher weights to features contributing more to surface runoff extrema. This helps improve the model's ability to express runoff extrema. The expression of the multi-head attention mechanism is shown in the following formula:

[0079] ;

[0080] ;

[0081] in, For multi-head attention functions; For query vector; The key vector; It is a value vector; For splicing; is the output of the i-th head in the multi-head attention mechanism; h is the number of heads in the multi-head attention mechanism; , , ,as well as It is a parameter matrix.

[0082] Furthermore, after processing by the multi-head attention layer, the multidimensional features will be flattened by the layer, which will then convert the multidimensional tensor into one dimension and map it to the output space by the fully connected layer, ultimately forming a surface runoff output sequence of length n, which is the target surface runoff cause identification result.

[0083] For comparison of technical effectiveness, existing technologies can be used as a reference. Current surface runoff analysis methods are mainly divided into two types. The first is data-based analysis methods, including correlation analysis, which are mainly used to assess the impacts of climate change and landscape pattern changes on runoff and sediment transport. The second method uses process-based models or water balance principles to infer the mechanisms affecting runoff changes by establishing a quantitative relationship between environmental changes and runoff. Both methods have limitations in balancing data fitting accuracy and the analysis of nonlinear relationships between variables. In addition, existing quantitative studies on the pathogenesis mechanisms of watershed runoff often lack systematicity and fail to deeply analyze the interactions between meteorological, hydrological, and geomorphological factors.

[0084] Among them, the first data-based analysis method, correlation analysis, infers the driving mechanism by examining the relationship between statistical influencing factors and runoff. This method is widely used in Pearl River Basin (PRB) studies, as it can quickly identify correlations between variables. It has been used to explore the formation mechanism of the Water Storage Deficit Index (WSDI) and to quantify the impacts of climate change and human activities on runoff and sediment transport. However, this type of method can only reflect linear statistical relationships, making it difficult to capture the complex nonlinear interactions between variables, and it cannot elucidate causal relationships.

[0085] The second approach, based on hydrological theories (such as the water balance principle) or physical models (such as SWAT (Soil and Water Assessment Tool, a distributed hydrological model at the watershed scale) and the VIC model (Variable Infiltration Capacity, a large-scale land surface model based on energy-water balance), establishes functional relationships between influencing factors and runoff. This method has successfully revealed runoff generation mechanisms in multiple watersheds, clearly presenting the interactions of variables through water balance equations or process models. For example, the impact of temperature on watershed hydrology, sediment and nutrient transport, and the quantitative impact of climate change and human activities on runoff. However, physical models struggle to fully utilize large-scale data, and the integration of theory and data is limited, hindering in-depth analysis of runoff formation mechanisms.

[0086] Furthermore, recent studies have found that deep learning models, with their multi-layer nonlinear mapping capabilities, significantly outperform traditional process models in runoff simulation accuracy. The latest research combines Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Sequence-to-Sequence (Seq2Seq) networks with preprocessing techniques such as wavelets and Empirical Mode Decomposition, achieving a correlation coefficient R0. 2The coefficient of determination can reach as high as 0.90, and the Nash efficiency coefficient (NSE) exceeds 0.80. However, these models still generally lack a unified interpretability framework, making it difficult to quantify the contributions of each driving factor and conduct cross-basin comparisons. Specifically, deep learning excels in identifying abstract features from data, surpassing traditional regression methods and overcoming their inherent limitations. Despite these advances, the "black box" nature of deep learning models still limits their interpretability. To improve model transparency and credibility, eXplainable Artificial Intelligence (XAI) technology has emerged. By combining deep learning models to uncover the nonlinear relationships between input features and output variables, and using eXplainable Artificial Intelligence (XAI) technology to explain the contribution of features to the results, it is possible to deeply analyze the triggering mechanisms of extreme hydrological events.

[0087] To address the aforementioned problems, this invention provides a surface runoff causation identification method based on deep learning and interpretability analysis, used to identify and quantify the main driving factors and their contributions to watershed surface runoff. First, high-resolution meteorological and hydrological data, including precipitation, temperature, soil moisture, evaporation, wind speed, and upstream runoff, are collected. A unified time-series database is then constructed according to sub-watershed divisions for the CNN-BiLSTM-ATT model (i.e., a deep learning-based surface runoff causation identification model). This model uses a convolutional neural network (CNN) to extract multi-scale spatial features, combines a bidirectional long short-term memory network (BiLSTM) to capture temporal dependencies, and introduces an attention mechanism (ATT) to dynamically allocate feature weights, thereby accurately characterizing complex nonlinear processes and significantly improving the simulation accuracy of extreme runoff events, specifically enhancing the feature modeling capability for responses to extreme runoff (high flood peaks, low droughts). Finally, SHAP (SHapley Additive exPlanations) interpretability analysis is introduced to quantitatively evaluate the contribution of each input factor to the model output, achieving a visual analysis of the watershed hydrological response mechanism. This invention not only reveals new mechanistic characteristics of watershed-scale hydrological processes, but also provides quantifiable technical support for flood control scheduling and refined water resource management.

[0088] In this embodiment of the invention, a method for identifying the causes of surface runoff based on deep learning and interpretability analysis is provided. First, time-series data of meteorological and hydrological factors are acquired and input into a deep learning-based surface runoff cause identification model. This model includes a one-dimensional convolutional module, a bidirectional long short-term memory network, and a recognition module based on a multi-head attention mechanism. Next, the one-dimensional convolutional module performs convolution operations on the meteorological and hydrological factor time-series data, outputting a local response feature sequence. The bidirectional long short-term memory network performs bidirectional processing on the local response feature sequence to generate a comprehensive target feature sequence. Finally, the comprehensive target feature sequence is input into the recognition module based on a multi-head attention mechanism for identification, outputting the target surface runoff cause identification result. Based on the above scheme, this invention enhances the ability to extract local mutation features through a one-dimensional convolutional module and combines the bidirectional long short-term memory network and the multi-head attention mechanism to synergistically integrate dynamic weight allocation capabilities, thereby improving recognition accuracy.

[0089] For better illustration, refer to Figure 3 The diagram illustrates the steps of training a deep learning-based surface runoff causation identification model according to Embodiment 2 of the present invention. This process may include the following steps:

[0090] Step 301: Obtain multiple meteorological and hydrological factor time series data for model training, and input each meteorological and hydrological factor time series data for model training into the initial deep learning-based surface runoff cause identification model, and output the training surface runoff cause identification results corresponding to each meteorological and hydrological factor time series data for model training.

[0091] Step 302: Calculate the total sum of squares based on the identification results of multiple training surface runoff causes.

[0092] Step 303: Using the two-segment weighting method and the three-segment weighting method, output the weights corresponding to the actual runoff observation values ​​corresponding to the runoff cause identification results of each training surface runoff.

[0093] Step 304: Calculate the weighted sum of squared residuals based on the identification results of the causes of surface runoff in each training area, the actual observed values ​​of each runoff, and the corresponding weights.

[0094] Step 305: Substitute the weighted residual sum of squares and the total sum of squares into the preset loss function and take the derivative to output the model gradient.

[0095] Step 306: Update the model parameters of the initial deep learning-based surface runoff cause identification model using model gradient, determine the intermediate deep learning-based surface runoff cause identification model, and count the number of model updates in real time.

[0096] Step 307: Determine whether the number of model updates has reached the preset training number threshold.

[0097] Step 308: If yes, then calculate the feature importance score based on the multiple training surface runoff cause identification results output by the intermediate deep learning-based surface runoff cause identification model.

[0098] Step 309: Calculate the feature contribution percentage based on the feature importance score.

[0099] Step 3010: Use the intermediate deep learning-based surface runoff cause identification model as the trained deep learning-based surface runoff cause identification model.

[0100] It should be noted that (data standardization settings) are necessary because the input features such as rainfall, soil moisture, and temperature, and the output features of surface runoff, have different dimensions and orders of magnitude. Therefore, before model training, the input and output features need to be standardized to speed up the model solution and reduce the possibility of incorrectly judging the importance of features.

[0101] Furthermore, (sliding window partitioning of input and output data) this invention employs a sliding window method to partition the training and validation sets. Each sliding window has a fixed size (e.g., a sequence containing a stride). Within the window, data is divided into training and validation sets in a fixed proportion (e.g., the first 80% is the training set, and the remainder is the validation set). The window slides along the time series with a fixed stride, forming multiple independent windows sequentially. In this way, the training and validation data within each window do not overlap at all. Compared to the traditional method of partitioning by the overall proportion (e.g., 80% training, 20% validation), the sliding window method has the following advantages: 1) The data between each sliding window is independent, and the training and validation sets are partitioned separately within each window, ensuring that the training and validation data do not overlap, avoiding data leakage, and improving the reliability of the model; 2) The event intensity of hydrological time series (e.g., rainfall and runoff) varies in different years. The sliding window method, through multiple independent partitions, increases the consistency of data distribution between the training and validation sets, which helps to improve the generalization ability of the model.

[0102] Furthermore, to improve the model's ability to fit extreme runoff values, this invention proposes two types of adaptive weighted loss functions based on the standard R² loss. A threshold is set according to the magnitude of the observed runoff value, and weights are assigned to the residuals in different intervals to highlight the training effect during the peak phase. Let P... i O represents the simulated runoff value at step size i (the result of training surface runoff cause identification). iLet be the runoff observation value (actual runoff observation value) corresponding to the i-th training surface runoff cause identification result, and n be the total number of meteorological and hydrological factor time series data used for model training. and These are the global average values ​​of the simulated runoff and the global average values ​​of the observed runoff, respectively. The formula for calculating the Sum of Squares Total (SST) is as follows:

[0103] ;

[0104] Furthermore, for the two-stage weighted method, runoff is divided into two categories: "peak" and "non-peak," with higher weights assigned to the peak interval to enhance the model's focus on extreme runoff. For the three-stage weighted method, runoff is divided into three levels: "low," "medium," and "high," each with different weights, to help the model comprehensively capture the runoff generation mechanism throughout the entire process.

[0105] Furthermore, the weights can be selected using two methods: hard segmentation or Sigmoid smoothing. 1) Hard weights: The weights are determined directly based on the threshold, and the weights change between different levels. 2) Soft weights: The threshold difference is first input into the Sigmoid function, and then mapped to the specified weight range to achieve a continuous and smooth transition.

[0106] When using the two-segment weighting method, the calculation formulas for hard weights and soft weights are as follows:

[0107] ;

[0108] ;

[0109] Where q is the runoff threshold calculated in percentiles; For high weight, Low weight; This refers to the Sigmoid function. Since the Sigmoid function takes values ​​of (0, 1), Directly through this function, then It was eventually compressed into the [0,1] interval. hour, ; hour, The Sigmoid function has a non-zero derivative over the entire real number domain, which not only makes the weights... arrive Smooth changes between these parameters help the model converge stably during backpropagation.

[0110] Similarly, the formulas for calculating the hard and soft weights in the three-segment weighted average are as follows:

[0111] ;

[0112] ;

[0113] In the three-stage weighted average, greater than The third tier of high weight, As medium weight, It has a low weight.

[0114] This is a parameter controlling the steepness of the sigmoid transition, used to adjust the rate at which the weights shift from one level to another. When the value is large, The transition between adjacent thresholds q is steeper, and the weights are closer to hard segmentation; when When the value is smaller, the weight transition is smoother. In practical applications, The value can be selected from 5 to 20 depending on the actual needs.

[0115] Furthermore, based on the above steps, the weight w corresponding to each actual runoff observation is obtained. i Used to calculate the weighted sum of squared residuals : .

[0116] Furthermore, the weighted sum of squared residuals and the total sum of squares are substituted into the preset loss function and differentiated to output the model gradient, thereby updating the model parameters. The preset loss function is as follows:

[0117] ;

[0118] in, The loss value corresponding to the preset loss function; This is the weighted sum of squared residuals; The total sum of squares; Value 10 -8 .

[0119] Furthermore, if the number of model updates reaches a preset training threshold, it indicates that the model has completed training. Since deep learning models are often considered "black boxes" due to their complex network structures, it is difficult to directly understand how input features affect surface runoff. Therefore, this invention employs Expected Gradients (EG) to analyze model behavior. EG is an improved version of Integrated Gradients (IG)—IG, as a feature attribution method based on Aumann-Shapley values ​​(extending Shapley values ​​from game theory to infinite-player scenarios), is widely used for interpretability analysis of deep learning models. This method establishes a correlation between daily surface runoff sequences and specific input features by calculating the gradient of the model output with respect to the input.

[0120] Furthermore, the baseline input refers to a specific data point in the feature space, which serves as a reference point or starting point to determine the contribution of each feature. By calculating the gradient integral along the path from the baseline input (α=0) to the actual input (α=1), the contribution of each input feature to the prediction result can be quantified. However, the integral gradient method (IG) has two main limitations: it requires manual selection of the baseline input and has high computational cost. To address these issues, the expected gradient method (EG) proposed by Erion et al. uses random sampling to approximate the baseline selection problem. The EG method assumes that the training dataset follows a basic distribution D and randomly samples the baseline input from it. Then, it integrates the gradient of all possible baseline inputs x' (according to the distribution density function p) D The contribution of feature i is calculated by weighting the factors. The following is the specific calculation method for the EG feature importance score:

[0121] ;

[0122] in, is the expected gradient score of feature i, which quantifies the average marginal contribution of the i-th variable / feature in the input vector x to the model output f(x); the baseline input x' is a "neutral" point selected from the input samples that represents the normal distribution (D) of the input features, such as a sample point randomly drawn from a region close to the center of the input data for statistical analysis / cluster analysis or calculation of the mean or median. i For each step size, the input (from samples, true values) for a specific feature i (rainfall, temperature, etc.) contains... In the integral formula, f(x) is the trained model (i.e., the trained deep learning-based surface runoff cause identification model). These are the model input values ​​used to calculate the gradient. The interpolation coefficients are used to construct an integration path by gradually and linearly varying the input of f(x) from the baseline x' to the target x, before integration. The denominator is x. i This means that the calculation starts from x. i 'To x i The change; the independent variable in the molecule does not contain i, which means that the input of the model f(x) includes all inputs, involving linear interpolation of all features.

[0123] Furthermore, the SHapley AdditiveexPlanations (SHAP) package is used for computation when constructing the runoff causation analysis program in this invention. This method is applicable to various types of neural networks and can provide post-hoc interpretation of the model output. Using the research employed in this invention combined with simulation data from this study area, SHAP can quantify the contribution of each input feature i (such as rainfall, soil moisture, temperature, upstream inflow, etc.) to the predicted runoff at the watershed outlet at each time step t. This time-by-time contribution attribution approach not only helps to deepen the understanding of runoff formation mechanisms but also enables the model to perform detailed analyses of extreme runoff events such as sudden and drastic changes. By identifying the dynamic evolution of the dominant driving factors during extreme events, this method provides strong support for revealing the formation pathways and key controlling factors of extreme hydrological processes.

[0124] In practical applications, users often need to compare and analyze the causal mechanisms of runoff across multiple watersheds or sub-watersheds. However, due to differences in the catchment area of ​​each sub-watershed, the magnitudes of their outlet runoff and the contributions of corresponding input factors are not directly comparable, thus limiting the accurate identification of causal differences between regions. To address this issue, this invention introduces a normalized attribution processing strategy: first, the absolute value of the contribution of each input feature within each sub-watershed is taken and its mean is calculated; then, its proportion in the total contribution is determined. This method not only standardizes the contribution across watersheds of different scales, but also enhances the transferability and regional comparative analysis capabilities of the method across different watersheds, helping to reveal the similarities and differences in driving mechanisms between regions. The formula for calculating the proportion of feature contribution is as follows:

[0125] ;

[0126] in, This represents the percentage contribution of the i-th input feature relative to all other features over the entire time series. This method multiplies the ratio of the cumulative contribution of each feature to the total contribution by 100% to obtain an intuitive "percentage" form, which facilitates comparison and visualization between different features and different regions. is the total number of input features; n is the length of the time series.

[0127] Furthermore, for deep learning models to extract meaningful information from sequences, a key prerequisite is that the model must possess reasonable and stable prediction accuracy. During the model training phase, this invention can also use Mean Absolute Error (MAE) as the objective function. For the trained model, MAE and Root Mean Square Error (RMSE) are used together to measure the deviation between predicted and actual surface runoff values, and the Coefficient of Determination (R²) and Nash-Sutcliffe Efficiency (NSE) are combined to evaluate the model's goodness of fit.

[0128] ;

[0129] It is worth mentioning that if the number of model updates does not reach the preset training number threshold, the intermediate deep learning-based surface runoff cause identification model is used as the new initial deep learning-based surface runoff cause identification model, and the process jumps to step 301 until the number of model updates reaches the preset training number threshold. The intermediate deep learning-based surface runoff cause identification model determined when the number of model updates reaches the preset training number threshold is used as the trained deep learning-based surface runoff cause identification model, and the feature contribution ratio is output.

[0130] Furthermore, in terms of runoff, the Pearl River, as China's second-largest river, profoundly impacts regional ecosystems, sediment transport, and deltaic landform evolution through its runoff variations. However, research on the driving factors of runoff changes in the Pearl River Basin remains insufficient; previous work has largely focused on the spatiotemporal trends of meteorological and hydrological elements, lacking a quantitative characterization of the runoff mechanisms under the coupling effects of climate change and complex geomorphology. To verify the effectiveness of this invention, using the Pearl River Basin as an example, the basin is divided into several sub-units based on topographic relief and river network structure, and its runoff confluence characteristics are systematically analyzed and comparative experiments are conducted.

[0131] Specifically, this invention develops a runoff causation analysis framework that integrates the CNN-BiLSTM-ATT deep learning model with interpretable artificial intelligence (XAI) technology. While achieving accurate modeling of complex nonlinear relationships, it significantly improves the interpretability of the model and the ability to identify the driving mechanisms of extreme hydrological events.

[0132] For a multi-module integrated deep learning framework: This application constructs a multi-module integrated deep learning model framework CNN-BiLSTM-ATT, integrating the advantages of CNN in extracting local features of sequences, the advantages of BiLSTM in extracting long-term dependencies of time series, and the advantages of multi-head attention mechanism in identifying extreme runoff contribution factors, achieving high-precision and robust runoff simulation. Each module has a clear task division. In simulating runoff in multiple sub-catchments of the Pearl River Basin, the model exhibits extremely strong fitting ability and cross-regional adaptability. Taking the study area as an example, the model's average Nash efficiency coefficient (NSE) and coefficient of determination (R²) are both 0.93 in 22 sub-catchments, with some sub-catchments having NSE and R² values ​​as high as 0.97, and more than 40% of the sub-catchments having an NSE value exceeding 0.95, significantly outperforming current mainstream models such as LSTM, BiLSTM, and Seq2Seq. This result verifies the simulation performance and stability advantages of this model architecture in large-scale, complex hydrological systems. Figure 4 As shown, Figure 4 The distribution of (b) root mean square error (RMSE), (c) mean absolute error (MAE), (d) Nash efficiency coefficient (NSE) and (e) coefficient of determination (R²).

[0133] To enhance the expressive power of extreme runoff responses by introducing a multi-head attention mechanism: This invention effectively improves the simulation accuracy of extreme runoff events by introducing a multi-head attention mechanism into the model structure. The multi-head attention module can dynamically identify and emphasize the variables in the input features that are most sensitive to flood peaks or extreme drought responses, improving the model's adaptability to abnormal fluctuations. Compared with the CNN-BiLSTM structure (a contrast model without multi-head attention), the root mean square error (RMSE) is reduced and extreme value predictions are more accurate, such as... Figure 4 As shown.

[0134] To construct a "deep learning + XAI" framework for explaining runoff driving mechanisms, a data-driven framework for analyzing hydrological driving mechanisms is proposed, combining deep learning models with interpretable artificial intelligence (XAI) technology. This framework uses data-driven methods to mine the nonlinear relationships between meteorological and hydrological variables (such as rainfall, temperature, and upstream runoff) and surface runoff, and quantifies the influence of characteristic factors through the SHAP method, thereby revealing the dynamic processes of different driving factors at various times. This process is not only applicable to the explanation of general runoff processes, but also capable of finely characterizing the triggering paths and dominant mechanisms of extreme events, such as the causal analysis of flood peaks or the formation process of drought minimum flows. Compared with traditional correlation or process model / hydrological equation simulation methods, this framework balances high time series accuracy, achieves quantitative attribution, and has the advantage of method transferability. This approach can overcome the limitations of traditional methods, more accurately reveal the complex hydrological driving mechanisms of the Pearl River Basin, and provide scientific support for hydrological disaster prediction and water resource management.

[0135] Compared to existing single LSTM, BiLSTM, or XGBoost models, this invention, for the first time, synergistically integrates the local mutation feature extraction capability of CNNs, the bidirectional long-term memory capability of BiLSTMs, and the dynamic weight allocation capability of multi-head attention mechanisms in its network architecture. This allows for the simultaneous capture of high-frequency, low-frequency, short-term, long-term, and extreme runoff response features, significantly improving both prediction accuracy and robustness. Furthermore, the invention introduces interpretable AI methods such as Normalized SHAP, which can quantify the relative contributions of multiple driving factors to runoff on a time-by-time basis, breaking through the "black box" limitation of traditional models and enabling mechanistic tracing and cross-basin comparable analysis of extreme flood peaks and drought events. Therefore, this invention not only improves runoff simulation performance but also provides a transparent and transferable mechanistic explanation framework, laying a more reliable technical foundation for flood control scheduling and water resource optimization management.

[0136] In this embodiment of the invention, the model architecture incorporates a multi-head attention mechanism to specifically enhance the feature modeling capability for responses to extreme runoff events (high flood peaks, low droughts). The invention also normalizes the contribution calculation, supporting direct comparison of runoff formation mechanisms across multiple watersheds. Furthermore, the simulation accuracy is higher, its adaptability is stronger, and its reproduction accuracy for high-flow events is higher.

[0137] Please see Figure 5 , Figure 5 This is a structural block diagram of a surface runoff causation identification device based on deep learning and interpretability analysis, provided in Embodiment 3 of the present invention.

[0138] This invention provides a surface runoff causation identification device based on deep learning and interpretability analysis, comprising:

[0139] The acquisition module 501 is used to acquire time series data of meteorological and hydrological factors and input the time series data of meteorological and hydrological factors into the surface runoff cause identification model based on deep learning; the surface runoff cause identification model based on deep learning includes a one-dimensional convolution module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism.

[0140] Convolution module 502 is used to perform convolution operations on meteorological and hydrological factor time series data using a one-dimensional convolution module, and output local response feature sequences.

[0141] Bidirectional module 503 is used to perform bidirectional processing operations on the local response feature sequence through a bidirectional long short-term memory network to generate the target comprehensive feature sequence;

[0142] The identification module 504 is used to input the comprehensive feature sequence of the target into the identification module based on the multi-head attention mechanism for identification, and output the identification result of the cause of surface runoff of the target.

[0143] Furthermore, the one-dimensional convolution module includes a one-dimensional convolutional layer, a random deactivation layer, and a residual connection layer; the convolution module 502 is specifically used for:

[0144] One-dimensional convolutional layers are used to perform one-dimensional convolution on the time series data of meteorological and hydrological factors, and one-dimensional convolutional feature sequences are output.

[0145] One-dimensional convolutional feature sequences are input into a random deactivation layer to perform random deactivation operations, generating random deactivation feature sequences.

[0146] By adding the random inactivation feature sequence and the time series data of meteorological and hydrological factors through the residual connection layer, a local response feature sequence is generated.

[0147] Furthermore, the bidirectional long short-term memory network includes a bidirectional long short-term memory network layer, a normalization layer, and a residual connection layer; the bidirectional module 503 is specifically used for:

[0148] The local response feature sequence is used as the input to the bidirectional long short-term memory network layer, and the output is a bidirectional temporal feature sequence.

[0149] A normalization layer is used to normalize the bidirectional temporal feature sequence, and a normalized temporal feature sequence is output.

[0150] The standardized temporal feature sequence and the local response feature sequence are added together by the residual connection layer to output the target comprehensive feature sequence.

[0151] Furthermore, the recognition module based on the multi-head attention mechanism includes a multi-head attention mechanism layer, a flattening layer, and a fully connected layer; the recognition module 504 is specifically used for:

[0152] The target's comprehensive feature sequence is used as input to the multi-head attention mechanism layer, and a multi-dimensional tensor is output.

[0153] A one-dimensional tensor is output by flattening a multidimensional tensor using a flattening layer.

[0154] A fully connected layer is used to map a one-dimensional tensor to generate the target surface runoff cause identification results.

[0155] In one optional device embodiment, it further includes:

[0156] Multiple meteorological and hydrological factor time series data for model training are obtained, and each meteorological and hydrological factor time series data for model training is input into the initial deep learning-based surface runoff cause identification model. The training surface runoff cause identification results corresponding to each meteorological and hydrological factor time series data for model training are output.

[0157] Calculate the total sum of squares based on multiple training surface runoff cause identification results;

[0158] The two-segment weighting method and the three-segment weighting method are used to output the weights corresponding to the actual runoff observation values ​​corresponding to the runoff cause identification results of each training surface runoff.

[0159] Based on the identification results of the causes of surface runoff in each training area, the actual observed values ​​of each runoff, and the corresponding weights, the weighted sum of squared residuals is calculated.

[0160] Substitute the weighted residual sum of squares and the total sum of squares into the preset loss function and take the derivative to output the model gradient;

[0161] Model gradients are used to update the model parameters of the initial deep learning-based surface runoff causation identification model, determine the intermediate deep learning-based surface runoff causation identification model, and count the number of model updates in real time.

[0162] Determine whether the number of model updates has reached the preset training count threshold;

[0163] If so, then the feature importance score is calculated based on the multiple training surface runoff cause identification results output by the intermediate deep learning-based surface runoff cause identification model.

[0164] Calculate the feature contribution percentage based on the feature importance score;

[0165] The intermediate deep learning-based surface runoff cause identification model is used as the trained deep learning-based surface runoff cause identification model.

[0166] Furthermore, the pre-defined loss function is as follows:

[0167] ;

[0168] in, The loss value corresponding to the preset loss function; This is the weighted sum of squared residuals; The total sum of squares; Value 10 -8 .

[0169] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0170] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any of the above embodiments.

[0171] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any of the above embodiments.

[0172] This invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any of the above embodiments.

[0173] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0174] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the causes of surface runoff based on deep learning and interpretability analysis, characterized in that, include: Acquire time series data of meteorological and hydrological factors, and input the time series data of meteorological and hydrological factors into a surface runoff cause identification model based on deep learning; the surface runoff cause identification model based on deep learning includes a one-dimensional convolutional module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism; The one-dimensional convolution module is used to perform convolution operations on the meteorological and hydrological factor time series data to output a local response feature sequence; The local response feature sequence is processed bidirectionally by the bidirectional long short-term memory network to generate the target comprehensive feature sequence. The target's comprehensive feature sequence is input into the multi-head attention mechanism-based recognition module for recognition, and the target's surface runoff cause recognition result is output. The training process of the deep learning-based surface runoff causation identification model is as follows: Multiple meteorological and hydrological factor time series data for model training are obtained, and each of the meteorological and hydrological factor time series data for model training is input into the initial deep learning-based surface runoff cause identification model, and the training surface runoff cause identification results corresponding to each of the meteorological and hydrological factor time series data for model training are output. Calculate the total sum of squares based on multiple training surface runoff causation identification results; The weights corresponding to the actual runoff observations are output based on the actual runoff observations corresponding to the training surface runoff cause identification results using a two-segment weighting method and a three-segment weighting method. Based on the identification results of the surface runoff causes of each training exercise, the actual observed values ​​of each runoff, and the corresponding weights, calculate the weighted sum of squared residuals. Substitute the weighted residual sum of squares and the total sum of squares into the preset loss function and take the derivative to output the model gradient; The model parameters of the initial deep learning-based surface runoff causation identification model are updated using the model gradient to determine the intermediate deep learning-based surface runoff causation identification model, and the number of model updates is counted in real time. Determine whether the number of model updates has reached the preset training count threshold; If so, then the feature importance score is calculated based on the multiple training surface runoff cause identification results output by the intermediate deep learning-based surface runoff cause identification model. Based on the feature importance score, the feature contribution percentage is calculated; The intermediate deep learning-based surface runoff cause identification model is used as the trained deep learning-based surface runoff cause identification model. The preset loss function is specifically as follows: ； in, The loss value corresponding to the preset loss function; This is the weighted sum of squared residuals; The total sum of squares; Value 10 -8 .

2. The surface runoff causation identification method based on deep learning and interpretability analysis according to claim 1, characterized in that, The one-dimensional convolutional module includes a one-dimensional convolutional layer, a random deactivation layer, and a residual connection layer; The step of using the one-dimensional convolution module to perform convolution operations on the meteorological and hydrological factor time series data to output a local response feature sequence includes: The one-dimensional convolutional layer is used to perform one-dimensional convolution on the time series data of the meteorological and hydrological factors, and outputs a one-dimensional convolutional feature sequence. The one-dimensional convolutional feature sequence is input into the random deactivation layer to perform a random deactivation operation, thereby generating a random deactivation feature sequence. The random inactivation feature sequence and the meteorological and hydrological factor time series data are added together through the residual connection layer to generate a local response feature sequence.

3. The surface runoff causation identification method based on deep learning and interpretability analysis according to claim 1, characterized in that, The bidirectional long short-term memory network includes a bidirectional long short-term memory network layer, a normalization layer, and a residual connection layer; The step of performing bidirectional processing on the local response feature sequence through the bidirectional long short-term memory network to generate the target comprehensive feature sequence includes: The local response feature sequence is used as the input to the bidirectional long short-term memory network layer, and a bidirectional temporal feature sequence is output. The bidirectional temporal feature sequence is standardized using the standardization layer to output a standardized temporal feature sequence; The standardized temporal feature sequence and the local response feature sequence are added together through the residual connection layer to output the target comprehensive feature sequence.

4. The surface runoff causation identification method based on deep learning and interpretability analysis according to claim 1, characterized in that, The recognition module based on the multi-head attention mechanism includes a multi-head attention mechanism layer, a flattening layer, and a fully connected layer; The step of inputting the target's comprehensive feature sequence into the multi-head attention-based recognition module for recognition and outputting the target's surface runoff cause recognition result includes: The target's comprehensive feature sequence is used as the input to the multi-head attention mechanism layer, and a multi-dimensional tensor is output. The multidimensional tensor is flattened by the flattening layer to output a one-dimensional tensor; The one-dimensional tensor is mapped using the fully connected layer to generate the target surface runoff cause identification result.

5. A surface runoff causation identification device based on deep learning and interpretability analysis, applied to the surface runoff causation identification method based on deep learning and interpretability analysis as described in claim 1, characterized in that, include: The acquisition module is used to acquire time series data of meteorological and hydrological factors and input the time series data of meteorological and hydrological factors into a deep learning-based surface runoff causation identification model; the deep learning-based surface runoff causation identification model includes a one-dimensional convolution module, a bidirectional long short-term memory network, and an identification module based on a multi-head attention mechanism. The convolution module is used to perform convolution operations on the meteorological and hydrological factor time series data using the one-dimensional convolution module, and output local response feature sequences. The bidirectional module is used to perform bidirectional processing operations on the local response feature sequence through the bidirectional long short-term memory network to generate the target comprehensive feature sequence; The identification module is used to input the comprehensive feature sequence of the target into the identification module based on the multi-head attention mechanism for identification, and output the identification result of the cause of surface runoff of the target.

6. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the surface runoff causation identification method based on deep learning and interpretability analysis as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the surface runoff causation identification method based on deep learning and interpretability analysis as described in any one of claims 1-4.

8. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the surface runoff causation identification method based on deep learning and interpretability analysis as described in any one of claims 1-4.

Images