Interpretability evaluation method and system for hydrological and meteorological deep learning forecasting models

By conducting interpretability evaluation on the deep learning forecasting model for hydrology and meteorology, and utilizing methods such as expected gradient, gradient weight mapping, and layer-by-layer correlation propagation, the opacity problem of deep learning models was solved, thereby improving the interpretability and predictive reliability of the models.

CN118657225BActive Publication Date: 2024-11-19SUN YAT SEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411141699.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-19
Estimated Expiration
2044-08-20

AI Technical Summary

Technical Problem

The lack of transparency in deep learning models for hydrological and meteorological forecasting leads to a lack of interpretability, affecting the reliability of model predictions and their practical application.

Method used

Multiple methods, including the expected gradient method, class activation mapping of gradient weights, and layer-by-layer correlation propagation, are used to interpret the deep learning forecast model for hydrology and meteorology. By acquiring target feature data and optimizing the model, the interpretability of the model is improved.

Benefits of technology

It improves the reliability and robustness of deep learning models, enhances the credibility of model prediction results, and overcomes the interpretability challenge of black box models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118657225B_ABST
    Figure CN118657225B_ABST
Patent Text Reader

Abstract

The present disclosure provides an interpretability evaluation method and system for a hydrological and meteorological deep learning forecast model, including: obtaining target feature data from a pre-built hydrological forecast database; determining a hydrological and meteorological deep learning forecast model based on the target feature data; predicting and interpreting the hydrological and meteorological deep learning forecast model based on an expected gradient method to obtain first prediction and interpretation data; predicting and interpreting the hydrological and meteorological deep learning forecast model based on a gradient weighted class activation mapping method to obtain second prediction and interpretation data; predicting and interpreting the hydrological and meteorological deep learning forecast model based on a layer-by-layer correlation propagation method to obtain third prediction and interpretation data; and optimizing the hydrological and meteorological deep learning forecast model based on the first prediction and interpretation data, the second prediction and interpretation data, and the third prediction and interpretation data. Thus, the reliability and robustness of the model are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the technical field of hydrological and meteorological forecasting, and in particular, to an interpretability evaluation method and system applicable to a hydrological and meteorological deep learning forecasting model. Background Art

[0002] Hydrometeorological forecasts provide estimates of future basin conditions based on current and future meteorological conditions and hydrological observations. Typical forecasts include precipitation, evaporation, runoff, and soil moisture, which interact with each other and together constitute the natural cycle of water. Hydrometeorological forecasts have a wide range of applications, including predicting extreme weather events such as early warning of floods and droughts; optimizing water resource scheduling, hydropower station operations, long-term water resource planning, and achieving precision agricultural irrigation and watershed management; and can also be used to assess the impact of climate change on hydrological systems.

[0003] Due to the nonlinear characteristics of the atmosphere at altitude, accurate prediction of hydrometeorological elements faces certain difficulties. In order to improve the accuracy of hydrometeorological forecasts, extensive research has been conducted in the main directions, namely model improvement and data-driven. Physics-based models adopt the mechanism of the hydrological cycle process and are highly interpretable, such as the Weather Research and Forecasting Model (WRF). Despite the transparency of physics-based models, their accuracy is relatively low, which is the main challenge facing hydrometeorological forecasts.

[0004] With the continuous development of big data and the continuous improvement of computer computing power in recent years, deep learning technology relying on huge data volume and high computing power has begun to be applied to the field of hydrometeorological forecasting. Scholars have used convolutional neural networks and long short-term memory neural networks based on deep learning to improve local precipitation forecasts to the 3-hour level. At the same time, deep learning also has a good response to extreme precipitation. Short-term predictions of evapotranspiration can be improved in accuracy by applying machine learning and deep learning. However, deep learning is often regarded as an opaque black box. Although these artificial intelligence methods can provide more accurate predictions of hydrometeorological variables, they face substantial challenges in interpretation due to their black box characteristics. Explainable artificial intelligence aims to promote communication between artificial intelligence models and human understanding and attempts to build trust in opaque models. Due to the need to explain the decisions made by the model, the lack of interpretability has become an obstacle to the further application of black box models. Summary of the invention

[0005] The embodiments described in this article provide a method and system for evaluating the interpretability of a hydrological and meteorological deep learning forecasting model, which overcomes the above-mentioned problems.

[0006] In a first aspect, according to the content of the present disclosure, a method for evaluating the interpretability of a hydrological and meteorological deep learning forecasting model is provided, comprising:

[0007] Obtain target characteristic data from a pre-built hydrological forecast database;

[0008] Determine a hydrological and meteorological deep learning forecasting model based on target feature data. The hydrological and meteorological deep learning forecasting model is used to predict regional soil moisture data through regional precipitation data.

[0009] Based on the expected gradient method, the hydrological and meteorological deep learning forecast model is predicted and interpreted to obtain the first prediction and interpretation data;

[0010] The class activation mapping method based on gradient weights is used to predict and explain the hydrological and meteorological deep learning forecast model to obtain the second prediction and explanation data;

[0011] Based on the layer-by-layer correlation propagation method, the hydrological and meteorological deep learning forecast model is predicted and interpreted to obtain the third prediction and interpretation data;

[0012] Based on the first prediction explanation data, the second prediction explanation data and the third prediction explanation data, the hydrological and meteorological deep learning forecasting model is optimized.

[0013] Optionally, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the expected gradient method to obtain first prediction and interpretation data, including:

[0014] Inputting the first precipitation data into the hydrological and meteorological deep learning forecasting model to obtain first soil moisture prediction data, wherein the first precipitation data corresponds to a baseline precipitation and a target precipitation;

[0015] Determining local gradient characterization data corresponding to the first precipitation data based on a baseline precipitation and a target precipitation corresponding to the first precipitation data, and the first soil moisture prediction data;

[0016] Based on the local gradient characterization data corresponding to the first precipitation data, first prediction interpretation data is determined, and the first prediction interpretation data is used to describe the expected gradient characterization of the first soil moisture prediction data relative to the first precipitation data.

[0017] Optionally, determining the first prediction interpretation data based on the local gradient representation data corresponding to the first precipitation data includes:

[0018] Determining a characteristic weight representation of a baseline precipitation amount corresponding to the first precipitation data;

[0019] First prediction interpretation data is determined based on a feature weight representation of a baseline precipitation amount corresponding to the first precipitation data and local gradient characterization data corresponding to the first precipitation data.

[0020] Optionally, a class activation mapping method based on gradient weights is used to predict and interpret the hydrological and meteorological deep learning forecast model to obtain second prediction and interpretation data, including:

[0021] Inputting the second precipitation data into the hydrological and meteorological deep learning forecasting model to obtain second soil moisture prediction data;

[0022] Determine corresponding weights of a plurality of feature maps based on convolution gradient data of the second soil moisture prediction data relative to a preset convolution layer in the hydrological and meteorological deep learning forecasting model;

[0023] Each feature map is weighted and combined based on multiple feature map weights to obtain second prediction explanation data.

[0024] Optionally, performing a weighted combination on each feature map based on multiple feature map weights to obtain second prediction explanation data includes:

[0025] Performing a weighted combination on each feature map based on a plurality of feature map weights to determine feature heat map data, wherein the feature heat map data includes: an image region where the second precipitation data is associated with the second soil moisture prediction data;

[0026] Image region recognition is performed on the characteristic heat map data to obtain second prediction interpretation data.

[0027] Optionally, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the layer-by-layer correlation propagation method to obtain third prediction and interpretation data, including:

[0028] Inputting the third precipitation data into the hydrological and meteorological deep learning forecasting model to obtain third soil moisture prediction data;

[0029] Obtaining a correlation score of the third soil moisture prediction data relative to each output layer neuron in the hydrological and meteorological deep learning forecasting model, where the correlation score is used to describe the correlation between the data feature points included in the third precipitation data and the third soil moisture prediction data;

[0030] Based on the relevance score, a third predictive explanatory data is determined.

[0031] Optionally, also include:

[0032] Based on the saliency map method, the hydrological and meteorological deep learning forecast model is predicted and explained to obtain the fourth prediction and explanation data;

[0033] Based on the integral gradient method, the hydrological and meteorological deep learning forecast model is predicted and interpreted to obtain the fifth prediction and interpretation data;

[0034] Based on the fourth prediction interpretation data and the fifth prediction interpretation data, the hydrological and meteorological deep learning forecasting model is optimized.

[0035] Optionally, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on a saliency map method to obtain fourth prediction and interpretation data, including:

[0036] Inputting the fourth precipitation data into the hydrological and meteorological deep learning forecasting model to obtain fourth soil moisture prediction data, the fourth precipitation data including: precipitation data at multiple moments;

[0037] Obtaining time gradient data of the fourth soil moisture prediction data relative to precipitation data at each moment;

[0038] Determine time gradient heat map data based on time gradient data of the fourth soil moisture prediction data relative to precipitation data at each moment, wherein the time gradient heat map data includes: an image region where the fourth precipitation data is associated with the fourth soil moisture prediction data;

[0039] Image region recognition is performed on the time gradient heat map data to obtain fourth prediction interpretation data.

[0040] Optionally, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the integral gradient method to obtain fifth prediction and interpretation data, including:

[0041] Inputting the fifth precipitation data into the hydrological and meteorological deep learning forecasting model to obtain the fifth soil moisture prediction data, wherein the fifth precipitation data corresponds to the baseline precipitation and the target precipitation;

[0042] Determining the integrated gradient data corresponding to the fifth precipitation data based on the baseline precipitation and the target precipitation corresponding to the fifth precipitation data;

[0043] determining, based on the integrated gradient data corresponding to the fifth precipitation data, an influence score of the fifth precipitation data relative to the fifth soil moisture prediction data;

[0044] Based on the impact score, a fifth predictive explanatory data is determined.

[0045] In a second aspect, according to the content of the present disclosure, a system for evaluating the interpretability of a hydrological and meteorological deep learning forecasting model is provided, comprising:

[0046] An acquisition module is used to obtain target characteristic data from a pre-built hydrological forecast database;

[0047] A first determination module is used to determine a hydrological and meteorological deep learning forecasting model based on target feature data, and the hydrological and meteorological deep learning forecasting model is used to predict regional soil moisture data through regional precipitation data;

[0048] The second determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the expected gradient method to obtain first prediction and interpretation data;

[0049] A third determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on a gradient weighted class activation mapping method to obtain second prediction and interpretation data;

[0050] The fourth determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the layer-by-layer correlation propagation method to obtain third prediction and interpretation data;

[0051] The optimization module is used to optimize the hydrological and meteorological deep learning forecasting model based on the first prediction explanation data, the second prediction explanation data and the third prediction explanation data.

[0052] The interpretability evaluation method of the hydrological and meteorological deep learning forecast model provided in the embodiment of the present application obtains target feature data from a pre-constructed hydrological forecast database; determines the hydrological and meteorological deep learning forecast model based on the target feature data, and the hydrological and meteorological deep learning forecast model is used to predict regional soil moisture data through regional precipitation data; predicts and interprets the hydrological and meteorological deep learning forecast model based on the expected gradient method to obtain the first prediction interpretation data; predicts and interprets the hydrological and meteorological deep learning forecast model based on the class activation mapping method of the gradient weight to obtain the second prediction interpretation data; predicts and interprets the hydrological and meteorological deep learning forecast model based on the layer-by-layer correlation propagation method to obtain the third prediction interpretation data; and optimizes the hydrological and meteorological deep learning forecast model based on the first prediction interpretation data, the second prediction interpretation data, and the third prediction interpretation data. In this way, by interpreting the model afterwards, the characteristic effect of the hydrological and meteorological deep learning forecast model is evaluated, which is convenient for effectively improving the reliability and robustness of the model.

[0053] The above description is only an overview of the technical solution of the embodiment of the present application. In order to more clearly understand the technical means of the embodiment of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the embodiment of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly described below. It should be noted that the drawings described below only relate to some embodiments of the present disclosure, rather than limiting the present disclosure, wherein:

[0055] Figure 1 It is a flow chart of an interpretability evaluation method of a hydrological and meteorological deep learning forecasting model provided in an embodiment of the present disclosure.

[0056] Figure 2 It is a structural diagram of an interpretability evaluation system for a hydrological and meteorological deep learning forecasting model provided in an embodiment of the present disclosure.

[0057] It should be noted that the elements in the drawings are schematic and not drawn to scale. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work also fall within the scope of protection of the present disclosure.

[0059] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by a person skilled in the art to which the subject matter of the present disclosure belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the specification and the relevant art, and will not be interpreted in an idealized or overly formal form unless otherwise explicitly defined herein. As used herein, a statement that two or more parts are "connected" or "coupled" together shall mean that the parts are joined together directly or through one or more intermediate components.

[0060] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase "embodiments" in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0061] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists, A and B exist at the same time, and B exists. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. Terms such as "first" and "second" are only used to distinguish one component (or a part of a component) from another component (or another part of a component).

[0062] In the description of the present application, unless otherwise specified, “plurality” means more than two (including two), and similarly, “plurality groups” means more than two (including two).

[0063] Hydrometeorological forecasts are forecasts of the state and possible impact of hydrometeorological elements such as precipitation, evaporation, soil moisture, and runoff in a water body, a watershed, or a station / grid in the future hydrological cycle. Hydrometeorological forecasts at specific locations show some common characteristics, which are mainly manifested in the following aspects: 1) Time series consistency: water flow and soil moisture content show significant time lag effects, which means that their changes do not occur immediately, but gradually appear over time. 2) The significant role of meteorological factors: Fluctuations in precipitation, sunshine duration, and average temperature have an important impact on hydrometeorological forecasts. Changes in these meteorological elements will directly or indirectly affect the hydrological cycle, thereby affecting the forecast results. 3) Energy exchange between land surface and atmosphere: Changes in hydrometeorological elements essentially reflect the exchange process of water vapor and heat between the surface and the atmosphere. This exchange is a core component of the surface water and heat cycle and is crucial to understanding the hydrometeorological system. In particular, soil moisture plays a key role in the surface water cycle because it can absorb and retain water. Therefore, accurate prediction of soil moisture is not only crucial for understanding hydrometeorological processes, but also indirectly indicates potential changes in other hydrological elements. Historical studies have confirmed that soil moisture has a significant impact on hydrological processes such as precipitation, runoff and evaporation in surface model simulations.

[0064] In practical hydrometeorological forecasting, the lack of interpretability can cause the following obstacles:

[0065] Difficulty understanding the inner workings of the model reduces trust in deep learning predictions in the hydrometeorological field. Lack of interpretability makes it difficult to explain the key influencing factors of the model on specific hydrological processes, limiting the practical application of model predictions. It may lead to misunderstandings of model predictions, making it difficult for decision makers to accept and adopt the application of deep learning technology in the hydrometeorological field. It hinders scientists and engineers from improving the model because of the lack of detailed information to understand how deep learning processes hydrometeorological data. Opaque models make it impossible to meet the needs for transparency and interpretability in regulatory and policy requirements, affecting the promotion and adoption of deep learning technology.

[0066] In order to overcome the interpretability challenge of black box models, scholars have also developed many technologies, including intrinsically interpretable models and post-hoc interpretation methods. After training with a large amount of data, black box models can be considered as more optimized forecasting models. Post-hoc interpretation methods are expected to overcome the interpretability challenge of black box models.

[0067] In this embodiment, the predicted hydrometeorological elements focus on soil moisture because soil moisture plays a key role in the global water cycle, which determines the utilization of water resources and the flow of surface energy. The retention and change of soil moisture are not only affected by soil texture (such as sandy and clay components) and vegetation cover, but also have a high sensitivity to meteorological conditions such as rainfall, air temperature and soil temperature. Since the dynamic change process of soil moisture is complex and highly nonlinear, it makes it a challenge to accurately predict it.

[0068] Therefore, this embodiment selects soil moisture as the hydrometeorological forecast object, and the selected explanatory variable is single-day precipitation. However, this embodiment is not limited to the application of soil moisture forecast, nor is it limited to explaining the impact of single-day precipitation on soil moisture.

[0069] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0070] Figure 1 is a flow chart of an interpretability evaluation method for a hydrological and meteorological deep learning forecasting model provided by an embodiment of the present disclosure, such as Figure 1 As shown in Figure 1, the specific process of the interpretability evaluation method of the hydrological and meteorological deep learning forecasting model includes:

[0071] S110, acquiring target characteristic data from a pre-built hydrological forecast database.

[0072] Among them, the hydrological forecast database includes: regional historical database and forecast factor database.

[0073] The data included in the hydrological forecast database may be from the ERA5-Land dataset, such as total precipitation, U wind component at 10 m height, V wind component, 2 m temperature, net solar radiation and net thermal radiation, and the features may be selected through a priori studies.

[0074] The selected ERA5-Land data covers the entire domestic area, including not only soil moisture data, but also a series of related meteorological variables, such as precipitation, temperature, wind speed and radiation. The domestic geographical and climatic conditions are diverse, and the interpretable forecast of soil moisture is of great significance to agriculture and the environment.

[0075] The ERA5-Land dataset is a global land application reanalysis dataset developed by a country's medium-term weather forecast center and is part of the ERA5 climate reanalysis project. The temporal resolution of the data is daily, and the regional historical database and forecast factor database include but are not limited to the following characteristics of related variables. Time series: Soil Moisture (SM); Meteorological forcing: Total Precipitation (TP), 10 m U-windcomponent (U), 10 m V-wind component (V), 2 mTemperature (TA); Radiative flux: Net Surface Solar Radiation (SR), Net Surface Thermal Radiation (TR).

[0076] The above data and some static variables (i.e., target feature data) can be used to input the hydrological and meteorological deep learning forecasting model. The target feature data are shown in Table 1.

[0077] Table 1 The 12 input features of the random forest forecast model and their descriptions

[0078]

[0079] S120. Determine a hydrological and meteorological deep learning forecasting model based on the target feature data, where the hydrological and meteorological deep learning forecasting model is used to predict regional soil moisture data through regional precipitation data.

[0080] Among them, a black box model for hydrological and meteorological factor forecasting can be established, that is, a hydrological and meteorological deep learning forecasting model. The black box models available include any of the deep learning models such as convolutional neural network, fully connected network, recurrent neural network, convolutional long short-term memory recursive neural network, etc. The first 70% of the time series in the data set is used as the training set, and the last 30% is used as the test set to statistically calculate the forecast error.

[0081] The deep learning model established in this embodiment is a convolutional long short-term memory neural network, and the first 70% of the time series data in the data set is used as a training set, and the last 30% is used as a test set. Convolutional Long Short-Term Memory Neural Network (Convolutional Long Short-Term Memory, Conv-LSTM) is a deep learning model that combines Convolutional Neural Network (CNN) and Long Short-Term Memory Neural Network, which is specifically used to process time series data with spatial correlation. It is suitable for the processing of soil moisture data in this embodiment, and its data not only changes over time, but also has spatial continuity. The model structure includes, input layer: the input is a five-dimensional tensor; convolutional LSTM layer: extract the spatial features of the input data and process the sequence data, capture the long-term dependencies in the time series, update the internal state through three gates and a unit state, and learn long-term dependencies; batch normalization layer and Dropout layer: normalize the output of the intermediate layer during training to improve the stability and performance of training, while reducing overfitting; output layer: output prediction results according to task requirements.

[0082] The statistical prediction errors include R2 (R-square, R2), mean absolute error (MAE), and root mean square error (RMSE). For details, please refer to formulas (1)-(3).

[0083] (1)

[0084] (2)

[0085] (3)

[0086] in, n Indicates the number of data points of soil moisture that need to be forecasted, Indicates Observed values ​​of soil moisture, Indicates The forecast value of each hydrological and meteorological element.

[0087] The deep learning model provided in this embodiment is a Conv-LSTM model, which can extract spatial features and capture long-term dependencies well. Compared with ordinary linear regression models and LSTM, Conv-LSTM shows better performance in the test of soil moisture prediction.

[0088] The dataset contains 1096 days of data from January 1, 2016 to December 31, 2018, with a spatial resolution of 9 kilometers. The first 70% of the data is used as the training set, and the last 30% of the data is used as the test set.

[0089] The statistical forecast errors are R2, MAE and RMSE. In this embodiment, the Conv-LSTM model is used for forecasting, and the best forecast result is determined by comparing two sets of data. Experiment 1 includes dynamic meteorological variables and static variables as original input data, and Experiment 2 only includes dynamic meteorological variables as original input data. The R2, MAE and RMSE of the experiment are shown in Table 2.

[0090] Table 2 Comparison of the impact of different input data sets on prediction performance

[0091]

[0092] The forecast error indicators of Experiment 1 are 0.92, 0.028m3 / m3 and 0.040m3 / m3 respectively, and the forecast error indicators of Experiment 2 are 0.84, 0.039m3 / m3 and 0.055m3 / m3 respectively. The two groups of experiments show that incorporating dynamic meteorological variables and static variables into input data can improve the forecast quality of soil moisture, so this embodiment adopts the solution of Experiment 1.

[0093] S130. Predict and interpret the hydrological and meteorological deep learning forecasting model based on the expected gradient method to obtain first prediction and interpretation data.

[0094] Among them, the expected gradient method selects the target model (i.e., the hydrological and meteorological deep learning forecasting model), input examples, and the potential distribution sampled from the training set as the benchmark input, calculates the model output, calculates the gradient, and integrates the gradient to obtain the EG (Expected Gradients) score. In this way, the impact of each input feature on the model forecast can be analyzed, providing an explanation for the hydrological model forecast results.

[0095] Expected Gradient is a feature attribution method for deep learning models that determines the importance of each feature to the prediction by computing the expected value of the model output relative to the input features. This method extends the concept of Integrated Gradients (IG) by randomly sampling samples from the dataset and computing the gradients of these samples, avoiding the need to choose a specific baseline input. The expected gradient satisfies a series of important interpretability axioms such as completeness and implementation invariance, which means that the sum of feature attributions is equal to the difference between the model output value and the baseline output value, and the same features will receive the same attribution for different implementations of the model. In addition, the calculation of the expected gradient can be effectively approximated by sampling methods, which makes it efficient for regularization during training, thereby improving the interpretability and performance of the model.

[0096] In some embodiments, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the expected gradient method to obtain first prediction and interpretation data, including:

[0097] The first precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain the first soil moisture prediction data, wherein the first precipitation data corresponds to a baseline precipitation and a target precipitation; based on the baseline precipitation and the target precipitation corresponding to the first precipitation data, and the first soil moisture prediction data, the local gradient characterization data corresponding to the first precipitation data is determined; based on the local gradient characterization data corresponding to the first precipitation data, the first prediction interpretation data is determined, and the first prediction interpretation data is used to describe the expected gradient characterization of the first soil moisture prediction data relative to the first precipitation data.

[0098] For example, input the original data (i.e., the first precipitation data), and forward propagate to obtain the predicted soil moisture data (i.e., the first soil moisture prediction data). A baseline value is proposed for the original precipitation data, and this baseline value is obtained from the distribution of the data, as shown in formula (4).

[0099] (4)

[0100] In formula (4), x is the target input, x ' is from the data distribution D The background samples extracted from α is the interpolation parameter drawn from the uniform distribution U(0, 1).

[0101] For features i Each original input x i and baseline input x i 'For example, this method takes the original input x iand the corresponding baseline input x i 'Build a straight line path between and accumulate the input gradient value of each feature along this path. f represents the output function of the model, i Represents the index of the feature, through the model output function f On input x Along the way from x 'arrive x The average gradient of the path is used to calculate the feature i The attributed value of .

[0102] The local gradient characterization data corresponding to the first precipitation data can be shown in the following formula (5).

[0103] (5)

[0104] In formula (5), the feature of interest in the original data is selected x i , x i ' is the baseline value of the corresponding feature, indicating the impact of the feature on the result when it is "no effect" or "missing"; Indicates the interpolation path i The local gradient of each feature represents the change in model output when a single feature changes. The local gradient is integrated. x i - x i ' is the difference in input feature change, representing the change from the baseline input to the target input.

[0105] In some embodiments, determining the first prediction interpretation data based on the local gradient representation data corresponding to the first precipitation data includes:

[0106] Determine a characteristic weight representation of the baseline precipitation corresponding to the first precipitation data; and determine first prediction interpretation data based on the characteristic weight representation of the baseline precipitation corresponding to the first precipitation data and the local gradient characterization data corresponding to the first precipitation data.

[0107] Among them, the characteristic weight of the baseline precipitation corresponding to the first precipitation data can be expressed by P D Function implementation, P D The weight function represents all possible baseline values ​​of the selected feature. The weighted expected value of all possible baseline values ​​is used as the baseline value for the final integration. The integral result is the EG score of the selected feature, that is, the final contribution value of the selected feature to the forecast result.

[0108] Based on the characteristic weight representation of the baseline precipitation corresponding to the first precipitation data and the local gradient characterization data corresponding to the first precipitation data, the first prediction interpretation data is determined, as shown in formula (6).

[0109] (6)

[0110] In formula (6), D represents the potential distribution sampled from the training dataset, and the baseline value of the selected feature follows this distribution D .

[0111] Traverse the pixel-by-pixel features of the original precipitation data, obtain the contribution score of each feature to the model soil moisture prediction result, generate a heat map, and understand the overall contribution of precipitation to the final output.

[0112] Compared with the interpretation of integral gradient, the interpretation of expected gradient avoids the problems that may arise from improper baseline value setting and further reduces the occurrence of extreme values. This method highlights the sensitivity of soil moisture to precipitation in the middle and lower reaches of the Yangtze River. This method also satisfies the two important axioms of sensitivity and implementation invariance, making the attribution results more reliable and consistent.

[0113] S140. Predict and interpret the hydrological and meteorological deep learning forecasting model based on the gradient weighted class activation mapping method to obtain second prediction and interpretation data.

[0114] Among them, the gradient weighted class activation map uses the gradient information flowing into the last convolutional layer of the deep learning model neural network to perform global average pooling, calculate the weight of each channel, and generate a heat map by weighted summing up the activation values ​​of the convolutional layer through the weights. After passing through the activation function and interpolating to the input image size, the saliency map is finally superimposed on the original input image to intuitively display the model's focus area on the input image.

[0115] Gradient-weighted Class Activation Mapping (Grad-CAM) is a method for explaining the decisions of convolutional neural networks. Its core principle is to use the gradient information of the target category to generate a visual representation of important areas in the image. Specifically, Grad-CAM calculates the gradient of the target category score generated by the input image in the last convolutional layer, and then globally averages these gradients to obtain the weights of each feature map. These weights reflect the contribution of the feature map to the importance of the target category; then, these weighted feature maps are combined and a rough heat map is generated through the ReLU activation function. The heat map highlights the image areas that are critical to the predicted soil moisture. This heat map can provide an intuitive visual explanation to help understand how deep learning models focus on specific parts of the image to make classification decisions.

[0116] In some embodiments, a class activation mapping method based on gradient weights is used to predict and interpret the hydrological and meteorological deep learning forecast model to obtain second prediction and interpretation data, including:

[0117] The second precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain second soil moisture prediction data; based on the convolution gradient data of the second soil moisture prediction data relative to the preset convolution layer in the hydrological and meteorological deep learning forecasting model, the corresponding multiple feature map weights are determined; based on the multiple feature map weights, each feature map is weightedly combined to obtain second prediction interpretation data.

[0118] Among them, the target convolution layer of the deep learning model is selected, and forward propagation is performed to obtain soil moisture prediction, that is, the second soil moisture prediction data is obtained. The gradient of the prediction result (that is, the second soil moisture prediction data) to the last convolution layer is calculated to obtain the feature map weight, as shown in the following formula (7).

[0119] (7)

[0120] In formula (7), i and j are image dimensions, y c is the original score of the prediction result before the activation function, A ij k is the feature map of the last convolutional layer, α k c is the importance weight of feature map k for category c, and Z is the length multiplied by the width of the feature layer.

[0121] In some embodiments, weighted combination of each feature map based on multiple feature map weights to obtain second prediction explanation data includes:

[0122] Based on multiple feature map weights, each feature map is weighted combined to determine feature heat map data, which includes: an image area where the second precipitation data is associated with the second soil moisture prediction data; image area recognition is performed on the feature heat map data to obtain second prediction interpretation data.

[0123] Among them, the weighted feature maps are combined, and these weights reflect the contribution of the feature maps to the importance of soil moisture. After combination, a rough heat map is generated, which highlights the image areas where precipitation has an important impact on soil moisture. For details, see formula (8).

[0124] (8)

[0125] In formula (8), is the Grad-CAM map of predicted soil moisture. The ReLU function ensures that the values ​​in the map are non-negative. Represents the weighted sum of all feature maps; A The feature layer representing the output of the last convolutional layer, k Represents the feature layer A k channels, c represents the category, A k Represents the channel in feature layer A k data, α c k represent A k The weight of .

[0126] In the source areas of the Yangtze River and the Yellow River, the daily precipitation data has the most significant impact on soil moisture prediction. The Grad-CAM method achieves a visual presentation of the key areas of the deep learning model prediction results by mapping gradient information to the class activation map. In the case discussed, the method can provide an intuitive explanation of soil moisture prediction for a specific model and input data while maintaining the original network architecture. In the generated heat map, high-value areas identify areas where single-day precipitation has a significant impact on soil moisture prediction, while low-value areas may indicate low soil moisture or limited changes. For this embodiment, the method can better capture the impact of the source areas of the Yangtze River and the Yellow River on soil moisture prediction during the model prediction process. Compared with traditional saliency maps and Class Activation Mapping (CAM) methods, Grad-CAM shows higher credibility and robustness, providing a reliable means for the explanatory analysis of deep learning models.

[0127] S150. Predict and interpret the hydrological and meteorological deep learning forecasting model based on the layer-by-layer correlation propagation method to obtain third prediction and interpretation data.

[0128] Among them, the Layer-wise Relevance Propagation (LPR) method explains the prediction by decomposing the classification decisions of the DNN (Deep Neural Network) model. This method establishes a local redistribution rule for each neuron in the network, backpropagates from the output layer to the input layer, and applies these rules during the backpropagation process to produce pixel-level decomposition, verify the model strategy, and provide a clearer explanation for the forecast.

[0129] Layer-by-layer relevance propagation is an algorithm used to explain the predictions of deep neural networks. The basic principle is to back-propagate the relevance score of each neuron to its input neurons by establishing a local redistribution rule. This process starts from the output layer and passes forward layer by layer until it reaches the input layer. During the back-propagation process, specific mathematical rules (such as the "alpha-beta" rule) are used to ensure the conservation of relevance, that is, the sum of the relevance scores of neurons in each layer is equal to the sum of the previous layer. In this way, LRP is able to quantify the contribution of input data (such as pixels in an image) to the final prediction result, thereby revealing why the model makes a specific decision.

[0130] In some embodiments, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the layer-by-layer correlation propagation method to obtain third prediction and interpretation data, including:

[0131] The third precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain the third soil moisture prediction data; the correlation score of the third soil moisture prediction data relative to each output layer neuron in the hydrological and meteorological deep learning forecasting model is obtained, and the correlation score is used to describe the correlation between the data feature points included in the third precipitation data and the third soil moisture prediction data; based on the correlation score, the third prediction explanatory data is determined.

[0132] For example, the original data (i.e., the third precipitation data) is input, and the soil moisture prediction result (i.e., the third soil moisture prediction data) is obtained through forward propagation.

[0133] After obtaining the soil moisture prediction results, the neuron activation value of the output layer is initialized as the correlation score, as shown in formula (9).

[0134] (9)

[0135] In formula (9), R j Representative j The correlation score of neurons, a j Represents thej These scores represent the contribution of each output neuron to the final prediction.

[0136] Applying the redistribution rule, starting from the output layer L and working backwards to the input layer, the goal of LRP is to decompose the prediction function f(x) into the sum of the input variables, that is, to find a set of correlation scores R p , so that ∑p Rp = f(x). These relevance scores represent the contribution of each pixel to the prediction result. For each layer in the network, LRP uses a local redistribution rule to calculate the relevance score of each neuron. This rule is usually based on the "alpha-beta" rule. Then for the lth layer and the l The redistribution of neurons between +1 layers follows the rule shown in formula (10).

[0137] (10)

[0138] In formula (10), R i is the lth layer of neurons i The relevance score, x i is the lth layer of neurons i The activation value of ω ij From the first layer to the l +1 layer weight, R j It is l +1 layer of neurons j The correlation score of each layer must be calculated, and the sum of the correlation scores between layers must be conserved.

[0139] After back propagation, in the input layer, the relevance score of each neuron can be directly associated with the input pixel value, and the obtained R represents the pixel-by-pixel relevance score as shown in formula (11).

[0140] (11)

[0141] By aggregating the correlation scores of each pixel, we can intuitively show which parts of the input data have a greater impact on the model prediction results.

[0142] From a geographical point of view, the precipitation data in the source areas of the Yangtze River and the Yellow River on that day has the most significant impact on the soil moisture forecast results. The LRP method calculates the correlation score of neurons in the output layer back-propagated to the input layer to obtain the impact of each feature in the input data (such as pixel-by-pixel daily precipitation) on the final forecast result. The correlation score can intuitively show the contribution of the original pixel and express it visually.

[0143] S160. Based on the first prediction explanation data, the second prediction explanation data and the third prediction explanation data, the hydrological and meteorological deep learning forecasting model is optimized.

[0144] Among them, the hydrological and meteorological deep learning forecasting model is optimized through the first prediction and explanation data, which facilitates the parameter adjustment of the hydrological and meteorological deep learning forecasting model.

[0145] In this embodiment, the target feature data is obtained from a pre-constructed hydrological forecast database; a hydrological and meteorological deep learning forecast model is determined based on the target feature data, and the hydrological and meteorological deep learning forecast model is used to predict regional soil moisture data through regional precipitation data; the hydrological and meteorological deep learning forecast model is predicted and interpreted based on the expected gradient method to obtain the first prediction interpretation data; the hydrological and meteorological deep learning forecast model is predicted and interpreted based on the gradient weighted class activation mapping method to obtain the second prediction interpretation data; the hydrological and meteorological deep learning forecast model is predicted and interpreted based on the layer-by-layer correlation propagation method to obtain the third prediction interpretation data; based on the first prediction interpretation data, the second prediction interpretation data and the third prediction interpretation data, the hydrological and meteorological deep learning forecast model is optimized. In this way, by interpreting the model afterwards, the characteristic effect of the hydrological and meteorological deep learning forecast model is evaluated, which is convenient for effectively improving the reliability and robustness of the model.

[0146] In some embodiments, it also includes:

[0147] Based on the saliency map method, the hydrological and meteorological deep learning forecasting model is predicted and interpreted to obtain the fourth prediction and interpretation data; based on the integral gradient method, the hydrological and meteorological deep learning forecasting model is predicted and interpreted to obtain the fifth prediction and interpretation data; based on the fourth prediction and interpretation data and the fifth prediction and interpretation data, the hydrological and meteorological deep learning forecasting model is optimized.

[0148] Among them, the saliency map method calculates the derivative of the original input image data at the corresponding pixel position based on the output result of the last layer of the model to obtain the gradient of each element of the input layer. Using this gradient map, a saliency visualization map is obtained. The higher the significance of the input element, the brighter the color appears in the map. By comparing it with the original input mapping, the key features that the model focuses on can be obtained.

[0149] The integral gradient method selects the target model (i.e., the hydrological and meteorological deep learning forecasting model), input examples and benchmark inputs, calculates the model output, calculates the gradient, and integrates the gradient to obtain the IG score. This can analyze the impact of each input feature on the model forecast and provide an explanation for the hydrological model forecast results.

[0150] In some embodiments, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the saliency map method to obtain fourth prediction and interpretation data, including:

[0151] The fourth precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain fourth soil moisture prediction data, which includes: precipitation data at multiple moments; obtaining time gradient data of the fourth soil moisture prediction data relative to the precipitation data at each moment; determining time gradient heat map data based on the time gradient data of the fourth soil moisture prediction data relative to the precipitation data at each moment, the time gradient heat map data includes: image areas where the fourth precipitation data is correlated with the fourth soil moisture prediction data; performing image area recognition on the time gradient heat map data to obtain fourth prediction interpretation data.

[0152] Among them, the saliency map is a deep learning model interpretability enhancement technology that aims to reveal the feature areas in the input samples that have a significant impact on the model prediction decision. The method first performs forward propagation on the selected input sample to obtain the predicted output of the model; then, the gradient of the output relative to the input is calculated by the back propagation algorithm, which reflects the sensitivity of the small changes in the input data to the model prediction results. Then, these gradient values ​​are normalized and mapped to the corresponding positions of the input data to generate a pixel-by-pixel saliency map. In the saliency map, higher gradient values ​​correspond to areas that have a greater impact on the model decision, and these areas are presented in the form of highlights in the figure. This method can provide an intuitive visual representation of the decision-making process of the model, and it helps to understand the behavior pattern of the Conv-LSTM model in this embodiment in terms of identifying features, predicting decision-making, etc.

[0153] For example, suppose that the explanation i The influence of precipitation data at each moment on the prediction result is calculated by using Conv-LSTM and the original data, and the prediction result is obtained through forward propagation, that is, the fourth prediction explanation data.

[0154] According to formula (12), the prediction result is directly calculated pixel by pixel relative to the first i The gradient of precipitation data at each moment is calculated and a heat map is generated based on the calculation results.

[0155] (12)

[0156] In formula (12), S i Indicates the first i pixel values, Represents model output F ( I )For the input image I No. iThe gradient at the pixel position.

[0157] From the perspective of geographical location, the Yangtze River Basin and the Yellow River Basin have the highest significance in terms of precipitation on that day, and the Qinghai-Tibet Plateau also shows a high significance level. The significance map provides a more intuitive explanation of the forecast results, directly reflects the relationship between the model output and the original input data, and helps to improve the credibility of the results of the hydrological forecast black box model. The calculation is relatively simple, and the results are intuitive and easy to understand.

[0158] In some embodiments, the hydrological and meteorological deep learning forecasting model is predicted and interpreted based on the integrated gradient method to obtain fifth prediction and interpretation data, including:

[0159] The fifth precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain the fifth soil moisture prediction data, wherein the fifth precipitation data corresponds to a baseline precipitation and a target precipitation; based on the baseline precipitation and the target precipitation corresponding to the fifth precipitation data, the integral gradient data corresponding to the fifth precipitation data is determined; based on the integral gradient data corresponding to the fifth precipitation data, the influence score of the fifth precipitation data relative to the fifth soil moisture prediction data is determined; based on the influence score, the fifth prediction interpretation data is determined.

[0160] Among them, the integrated gradient is a deep learning model interpretability technology, and its calculation principle is based on evaluating the contribution of input features to the model prediction results. Specifically, it builds a straight line path between the input value and a selected baseline value (usually an input without activation or features, such as a completely black image or a zero vector, in this embodiment, a zero vector), and accumulates the gradient value of each input feature along this path. By integrating these gradients, the cumulative contribution of each feature to the model output can be obtained, thereby quantifying the importance of each input feature. The integrated gradient algorithm only considers the input and output of the model and the function is differentiable everywhere, without the participation of the internal details of the model, so the integrated gradient algorithm satisfies the implementation invariance.

[0161] For example, the original data (i.e., the fifth precipitation data) is input, and the predicted soil moisture data (i.e., the fifth soil moisture prediction data) is obtained by forward propagation. A baseline value is set for the original precipitation data, and this baseline value is usually 0, so in this embodiment, 0 is set as the baseline input of the input precipitation data.

[0162] For features i Each original input x i and baseline input x i 'For example, in the original input x i and the corresponding baseline input x i', and accumulate the input gradient value of each feature along this path, as shown in formula (13).

[0163] (13)

[0164] In formula (13), the feature of interest in the original data is selected x i , x i ' is the baseline value of the corresponding feature, which indicates the impact of the feature on the result when it is "no effect" or "missing". x i - x i ' is the difference in input feature change, representing the change from the baseline input to the target input.

[0165] Traverse the pixel-by-pixel features of the original precipitation data, obtain the contribution score of each feature to the model soil moisture prediction result, generate a heat map, and understand the overall contribution of precipitation to the final output.

[0166] The explanation of the integrated gradient is relatively average. In most parts of China, the contribution of precipitation on that day to soil moisture prediction is low, but it shows a higher contribution in the middle and lower reaches of the Yangtze River and some areas in the lower reaches. Northeast China and parts of northern Xinjiang also show a high contribution. The advantage of the integrated gradient is that it satisfies the two important axioms of sensitivity and realization invariance, making the attribution results more reliable and consistent.

[0167] In this embodiment, the saliency map method has the characteristics of simple algorithm and intuitive and natural calculation, and can identify the geographical areas that have an impact on soil moisture. The integral gradient and expected gradient determine the area with the greatest impact on soil moisture by calculating the cumulative contribution. In contrast, the calculations of Grad-CAM and LRP are more complicated and are suitable for models with more neural network layers and multidimensional data input. In this embodiment, the thermal map interpretation results generated by Grad-CAM and LRP are similar, and both mark the contribution of the precipitation data of the source areas of the Yangtze River and the Yellow River on that day to soil moisture as the highest value. At the same time, it is believed that the precipitation in the middle and lower reaches of the Yangtze River on that day also has a certain impact on soil moisture.

[0168] This embodiment uses soil moisture as an example of deep learning forecasting for hydrology and meteorology. The proposed deep learning integrated interpretation method can overcome the application areas of deep learning models from the following aspects: Construct a comprehensive deep learning model interpretation framework, which can assist hydrological forecasting work and help select and understand more practical models by comparing with local hydrological physics knowledge. The importance of the data in the model is used to evaluate the credibility of the forecast results. The degree of influence of different meteorological data on the target results can be quantitatively expressed according to the actual situation, and the main processes of local hydrology can be used as a reference to help evaluate the credibility of the model and the accuracy of the forecast results.

[0169] In response to the "black box" problem of deep learning forecasting in hydrology and meteorology, this embodiment integrates several post-explanation methods suitable for deep learning models, and provides a complete deep learning post-explanatory system. It solves the internal opacity of the hydrology and meteorology deep learning model caused by the end-to-end training method, and the unclear relationship between input data and forecast results, enhances the credibility of the forecast results, and promotes the application of deep learning in the field of hydrology and meteorology.

[0170] This post-explainable system can reflect the relationship between input features and output forecasts from multiple perspectives, provide evidence from multiple aspects, and provide support conditions for forecast results, including: methods based on integral gradients, methods based on expected gradients, methods based on saliency maps, methods based on Grad-CAM, and methods based on layer-by-layer correlation propagation, which can provide global model result trustworthiness assessment for hydrological and meteorological element forecasts, and can draw clear heat maps for locations in image data that have a greater impact on deep learning forecast results. Therefore, according to the areas identified by the method, the data and models can be evaluated in combination with local hydrological and meteorological physical processes to further improve the model forecast results.

[0171] Figure 2 A schematic diagram of the structure of an interpretability evaluation system for a hydrological and meteorological deep learning forecast model provided in this embodiment, wherein the interpretability evaluation system for a hydrological and meteorological deep learning forecast model may include:

[0172] The acquisition module 210 is used to acquire target characteristic data from a pre-built hydrological forecast database.

[0173] The first determination module 220 is used to determine a hydrological and meteorological deep learning forecasting model based on target feature data, and the hydrological and meteorological deep learning forecasting model is used to predict regional soil moisture data through regional precipitation data.

[0174] The second determination module 230 is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the expected gradient method to obtain first prediction and interpretation data.

[0175] The third determination module 240 is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the gradient weighted class activation mapping method to obtain second prediction and interpretation data.

[0176] The fourth determination module 250 is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the layer-by-layer correlation propagation method to obtain third prediction and interpretation data.

[0177] The optimization module 260 is used to optimize the hydrological and meteorological deep learning forecasting model based on the first prediction explanation data, the second prediction explanation data and the third prediction explanation data.

[0178] Optionally, the second determining module 230 is specifically configured to:

[0179] The first precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain the first soil moisture prediction data, wherein the first precipitation data corresponds to a baseline precipitation and a target precipitation; based on the baseline precipitation and the target precipitation corresponding to the first precipitation data, and the first soil moisture prediction data, the local gradient characterization data corresponding to the first precipitation data is determined; based on the local gradient characterization data corresponding to the first precipitation data, the first prediction interpretation data is determined, and the first prediction interpretation data is used to describe the expected gradient characterization of the first soil moisture prediction data relative to the first precipitation data.

[0180] Optionally, the second determining module 230 is specifically configured to:

[0181] Determine a characteristic weight representation of the baseline precipitation corresponding to the first precipitation data; and determine first prediction interpretation data based on the characteristic weight representation of the baseline precipitation corresponding to the first precipitation data and the local gradient characterization data corresponding to the first precipitation data.

[0182] Optionally, the third determining module 240 is specifically configured to:

[0183] The second precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain second soil moisture prediction data; based on the convolution gradient data of the second soil moisture prediction data relative to the preset convolution layer in the hydrological and meteorological deep learning forecasting model, the corresponding multiple feature map weights are determined; based on the multiple feature map weights, each feature map is weightedly combined to obtain second prediction interpretation data.

[0184] Optionally, the third determining module 240 is specifically configured to:

[0185] Based on multiple feature map weights, each feature map is weighted combined to determine feature heat map data, which includes: an image area where the second precipitation data is associated with the second soil moisture prediction data; image area recognition is performed on the feature heat map data to obtain second prediction interpretation data.

[0186] Optionally, the fourth determining module 250 is specifically configured to:

[0187] The third precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain the third soil moisture prediction data; the correlation score of the third soil moisture prediction data relative to each output layer neuron in the hydrological and meteorological deep learning forecasting model is obtained, and the correlation score is used to describe the correlation between the data feature points included in the third precipitation data and the third soil moisture prediction data; based on the correlation score, the third prediction explanatory data is determined.

[0188] Optionally, it also includes: a fifth determination module and a sixth determination module.

[0189] The fifth determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the saliency map method to obtain fourth prediction and interpretation data.

[0190] The sixth determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on the integral gradient method to obtain the fifth prediction and interpretation data.

[0191] The optimization module 260 is also used to optimize the hydrological and meteorological deep learning forecasting model based on the fourth prediction interpretation data and the fifth prediction interpretation data.

[0192] Optionally, the fifth determining module is specifically configured to:

[0193] The fourth precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain fourth soil moisture prediction data, which includes: precipitation data at multiple moments; obtaining time gradient data of the fourth soil moisture prediction data relative to the precipitation data at each moment; determining time gradient heat map data based on the time gradient data of the fourth soil moisture prediction data relative to the precipitation data at each moment, the time gradient heat map data includes: image areas where the fourth precipitation data is correlated with the fourth soil moisture prediction data; performing image area recognition on the time gradient heat map data to obtain fourth prediction interpretation data.

[0194] Optionally, the sixth determining module is specifically configured to:

[0195] The fifth precipitation data is input into the hydrological and meteorological deep learning forecasting model to obtain the fifth soil moisture prediction data, wherein the fifth precipitation data corresponds to a baseline precipitation and a target precipitation; based on the baseline precipitation and the target precipitation corresponding to the fifth precipitation data, the integral gradient data corresponding to the fifth precipitation data is determined; based on the integral gradient data corresponding to the fifth precipitation data, the influence score of the fifth precipitation data relative to the fifth soil moisture prediction data is determined; based on the influence score, the fifth prediction interpretation data is determined.

[0196] The interpretability evaluation system of the hydrological and meteorological deep learning forecast model provided by the present disclosure integrates a variety of deep learning interpretable methods, which can be well used for post-model interpretation and evaluation of model results. The saliency map method can reflect the gradient relationship between input data and soil moisture prediction results, and is suitable for deep learning models with relatively simple structures; IG, EG and other methods can extract available information from the gradients of the middle layers of the neural network to obtain more accurate and more reliable interpretation results; and for more complex deep learning networks, Grad-CAM and LRP can help us understand the internal working mechanism of the deep model from the perspective of model decision-making based on the importance (Relevance) or contribution score of the prediction results, and achieve a further understanding of the model and prediction results.

[0197] The interpretability evaluation system of the hydrological and meteorological deep learning forecasting model provided in the present disclosure can execute the above-mentioned method embodiments. Its specific implementation principles and technical effects can be found in the above-mentioned method embodiments, and the present disclosure will not repeat them here.

[0198] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for evaluating the interpretability of a hydrological and meteorological deep learning forecasting model, characterized in that: include: Obtain target characteristic data from a pre-built hydrological forecast database; Determining a hydrological and meteorological deep learning forecasting model based on the target feature data, wherein the hydrological and meteorological deep learning forecasting model is used to predict regional soil moisture data through regional precipitation data; Based on the expected gradient method, the hydrological and meteorological deep learning forecasting model is predicted and interpreted to obtain first prediction and interpretation data; Based on the gradient weighted class activation mapping method, the hydrological and meteorological deep learning forecasting model is predicted and interpreted to obtain second prediction and interpretation data; Based on the layer-by-layer correlation propagation method, the hydrological and meteorological deep learning forecasting model is predicted and interpreted to obtain third prediction and interpretation data; Based on the first prediction explanation data, the second prediction explanation data and the third prediction explanation data, optimizing the hydrological and meteorological deep learning forecasting model; The method of predicting and interpreting the hydrological and meteorological deep learning forecasting model based on the expected gradient method to obtain first prediction and interpretation data includes: Inputting first precipitation data into the hydrological and meteorological deep learning forecasting model to obtain first soil moisture prediction data, wherein the first precipitation data corresponds to a baseline precipitation and a target precipitation; determining local gradient characterization data corresponding to the first precipitation data based on the baseline precipitation and the target precipitation corresponding to the first precipitation data, and the first soil moisture prediction data; determining first prediction interpretation data based on the local gradient characterization data corresponding to the first precipitation data, wherein the first prediction interpretation data is used to describe an expected gradient characterization of the first soil moisture prediction data relative to the first precipitation data; The gradient weight-based class activation mapping method predicts and interprets the hydrological and meteorological deep learning forecast model to obtain second prediction and interpretation data, including: Inputting the second precipitation data into the hydrological and meteorological deep learning forecasting model to obtain second soil moisture prediction data; determining corresponding multiple feature map weights based on the convolution gradient data of the second soil moisture prediction data relative to the preset convolution layer in the hydrological and meteorological deep learning forecasting model; performing weighted combination on each feature map based on the multiple feature map weights to obtain the second prediction interpretation data; The method of predicting and interpreting the hydrological and meteorological deep learning forecasting model based on the layer-by-layer correlation propagation method to obtain third prediction and interpretation data includes: Input the third precipitation data into the hydrological and meteorological deep learning forecasting model to obtain third soil moisture prediction data; obtain the correlation score of the third soil moisture prediction data relative to each output layer neuron in the hydrological and meteorological deep learning forecasting model, wherein the correlation score is used to describe the correlation between the data feature points included in the third precipitation data and the third soil moisture prediction data; based on the correlation score, determine the third prediction explanatory data.

2. The method according to claim 1, characterized in that The determining the first prediction interpretation data based on the local gradient characterization data corresponding to the first precipitation data includes: Determining a characteristic weight representation of a baseline precipitation amount corresponding to the first precipitation data; The first prediction explanation data is determined based on a feature weight representation of a baseline precipitation amount corresponding to the first precipitation data and local gradient characterization data corresponding to the first precipitation data.

3. The method according to claim 1, characterized in that The step of performing weighted combination on each feature map based on the weights of the plurality of feature maps to obtain the second prediction explanation data includes: Performing a weighted combination on each feature map based on a plurality of feature map weights to determine feature heat map data, wherein the feature heat map data includes: an image region where the second precipitation data is associated with the second soil moisture prediction data; Image region recognition is performed on the characteristic heat map data to obtain the second prediction explanation data.

4. The method according to claim 1, characterized in that: Also includes: Based on the saliency map method, the hydrological and meteorological deep learning forecasting model is predicted and interpreted to obtain fourth prediction and interpretation data; Performing prediction interpretation on the hydrological and meteorological deep learning forecasting model based on an integrated gradient method to obtain fifth prediction interpretation data; Based on the fourth prediction and explanation data and the fifth prediction and explanation data, the hydrological and meteorological deep learning forecasting model is optimized.

5. The method according to claim 4, characterized in that The method based on the saliency map performs prediction interpretation on the hydrological and meteorological deep learning forecasting model to obtain fourth prediction interpretation data, including: Inputting fourth precipitation data into the hydrological and meteorological deep learning forecasting model to obtain fourth soil moisture prediction data, wherein the fourth precipitation data includes: precipitation data at multiple moments; Obtaining time gradient data of the fourth soil moisture prediction data relative to the precipitation data at each of the moments; Determine time gradient heat map data based on time gradient data of the fourth soil moisture prediction data relative to precipitation data at each moment, wherein the time gradient heat map data includes: an image region where the fourth precipitation data is associated with the fourth soil moisture prediction data; Image region recognition is performed on the time gradient thermodynamic map data to obtain the fourth prediction interpretation data.

6. The method according to claim 4, characterized in that The method of predicting and interpreting the hydrological and meteorological deep learning forecasting model based on the integrated gradient method to obtain fifth prediction and interpretation data includes: Inputting fifth precipitation data into the hydrological and meteorological deep learning forecasting model to obtain fifth soil moisture prediction data, wherein the fifth precipitation data corresponds to a baseline precipitation and a target precipitation; Determining the integrated gradient data corresponding to the fifth precipitation data based on the baseline precipitation and the target precipitation corresponding to the fifth precipitation data; determining, based on the integrated gradient data corresponding to the fifth precipitation data, an influence score of the fifth precipitation data relative to the fifth soil moisture prediction data; Based on the influence score, the fifth predictive explanatory data is determined.

7. An interpretability evaluation system for a hydrological and meteorological deep learning forecast model, characterized in that: include: An acquisition module is used to obtain target characteristic data from a pre-built hydrological forecast database; A first determination module is used to determine a hydrological and meteorological deep learning forecasting model based on the target feature data, wherein the hydrological and meteorological deep learning forecasting model is used to predict regional soil moisture data through regional precipitation data; A second determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on an expected gradient method to obtain first prediction and interpretation data; A third determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on a gradient weighted class activation mapping method to obtain second prediction and interpretation data; A fourth determination module is used to predict and interpret the hydrological and meteorological deep learning forecasting model based on a layer-by-layer correlation propagation method to obtain third prediction and interpretation data; An optimization module, configured to optimize the hydrological and meteorological deep learning forecasting model based on the first prediction and explanation data, the second prediction and explanation data, and the third prediction and explanation data; Wherein, the second determining module is specifically used to: Inputting first precipitation data into the hydrological and meteorological deep learning forecasting model to obtain first soil moisture prediction data, wherein the first precipitation data corresponds to a baseline precipitation and a target precipitation; determining local gradient characterization data corresponding to the first precipitation data based on the baseline precipitation and the target precipitation corresponding to the first precipitation data, and the first soil moisture prediction data; determining first prediction interpretation data based on the local gradient characterization data corresponding to the first precipitation data, wherein the first prediction interpretation data is used to describe an expected gradient characterization of the first soil moisture prediction data relative to the first precipitation data; The third determination module is specifically used to: Inputting the second precipitation data into the hydrological and meteorological deep learning forecasting model to obtain second soil moisture prediction data; determining corresponding multiple feature map weights based on the convolution gradient data of the second soil moisture prediction data relative to the preset convolution layer in the hydrological and meteorological deep learning forecasting model; performing weighted combination on each feature map based on the multiple feature map weights to obtain the second prediction interpretation data; The fourth determining module is specifically used for: Input the third precipitation data into the hydrological and meteorological deep learning forecasting model to obtain third soil moisture prediction data; obtain the correlation score of the third soil moisture prediction data relative to each output layer neuron in the hydrological and meteorological deep learning forecasting model, wherein the correlation score is used to describe the correlation between the data feature points included in the third precipitation data and the third soil moisture prediction data; based on the correlation score, determine the third prediction explanatory data.