Deep learning model key index attribution analysis method and system for geoscience
By performing feature extraction and reverse gradient calculation on the output results of the deep learning model, the impact of input factors on the target characteristics is quantified, which solves the problem that deep learning models are difficult to explain causal relationships in the field of earth science, and improves the interpretability of the model and the accuracy of scientific research.
Patent Information
- Application Number
- CN202510433381.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-07-25
AI Technical Summary
Deep learning models have difficulty explaining the contribution of input factors to target variables in the field of geoscience, especially in extreme precipitation, flooding, and climate change studies, and existing methods have difficulty quantifying the impact of input factors on model output.
By extracting the output results of the deep learning model and calculating the gradient in reverse, dividing time, space and space-time objects, calculating the set of key indicators, and quantifying the impact of input factors on target characteristics, including attribution analysis of trend, variability, anomalies and extremes.
The quantitative analysis of the model output of input factors in geoscience problems is realized, which improves the interpretability of the model and provides a scientific basis for extreme weather prediction and climate change research.
Smart Images
Figure CN120372567A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the cross - technical field of earth science and artificial intelligence, and in particular to a method and system for attributing key indicators of a deep - learning model for earth science. This method extracts key indices from the output of the deep - learning model and calculates the cumulative gradient, and for earth - science problems such as extreme precipitation, floods, and climate change, analyzes the key affected regions / time periods of the output of the deep - learning model to characterize the contribution degree of input factors to the model simulation results. Background Art
[0002] Deep - learning technology has been widely applied in earth - science fields such as meteorological prediction, flood simulation, and climate - change research. However, most deep - learning models are black - box models and it is difficult to directly explain the causal relationship of model predictions. In particular, it is difficult to quantify the contribution of input factors to target variables (such as precipitation maximum value, flood - peak discharge). This limitation of lack of interpretability makes deep - learning models face challenges in credibility and usability in practical applications. Existing deep - learning interpretability methods calculate the contribution of feature maps to the final simulation results through gradient backpropagation and characterize the contribution degree. This method is often used for pixel or data - point analysis and is difficult to be directly applied to earth - science problems. For example, extreme precipitation requires identifying extreme situations and analyzing the contribution degree of input factors for extreme values. Time - series data (such as flood processes) need to extract peaks and trends from the output flow values. In the study of global climate change, it is necessary to analyze the contribution values for temperature outliers. The existing calculation of contribution degree for data points or pixel points has no meaning and application value in these fields. In fields such as extreme - weather prediction, watershed flood simulation, and climate - change assessment, scientists and engineers need to determine which input factors (such as temperature, humidity, wind speed) have the greatest impact on the prediction results.
[0003] The information disclosed in this background - art section is only intended to deepen the understanding of the overall background art of the present invention and should not be regarded as an admission or any form of implication that this information constitutes the prior art known to those skilled in the art. Summary of the Invention
[0004] The purpose of the present invention is to provide a method and system for attributing key indicators of a deep - learning model for earth science, which can conduct attribution analysis on the affected regions / time periods of model predictions and quantify the impact of input factors on the key characteristics of target features.
[0005] To achieve the above purpose, the present invention adopts the following technical solutions:
[0006] The first aspect of the present invention provides a method for attributing key indicators of a deep learning model for geoscience. The method extracts key indicators of features from the output results of any deep learning model for geoscience research tasks. Specifically, the method includes:
[0007] Step S1: Obtain the model output, screen out the objects whose affected situations need to be analyzed, that is, the time periods / regions to be studied in the model output. Divide the objects whose affected situations need to be analyzed into three types of objects: in terms of time, in terms of space, and in terms of time and space. For the research questions of the actual concerned time period, concerned region, and both simultaneously, obtain the analysis object categories to determine the specific means differences for subsequent key indicator extraction;
[0008] Step S2: According to the divided analysis object categories, further calculate the key indicators to obtain a key indicator set. Calculate the basic indicators according to all the divided analysis object categories. Calculate the time indicators, space indicators, and time-space indicators according to the divided object categories of the concerned time period, concerned region, and both simultaneously;
[0009] Step S3: Match the affected analysis method according to the key indicator set, and obtain the affected situation of the key indicator set by reverse calculating the model gradient;
[0010] Step S4: According to the affected situation, combined with the analysis object categories, obtain the key feature influence situation of each feature of the model target layer on the output. The key features include trendiness, variability, abnormality, and extremity.
[0011] Further, in step S1, obtaining the analysis object category is a human judgment process. The judgment basis is the model output data type and the actual research target. Among them: the model output data type includes one-dimensional sequence data, two-dimensional image data, and two-dimensional image sequence data. The actual research target includes the trendiness, variability, and abnormality of the research object. The research object includes important factors in the field of geoscience such as precipitation, temperature, wind speed, vegetation coverage, and pollutant concentration that can be analyzed by the method system. According to the model output data type and the actual research target, divide the objects whose affected situations need to be analyzed into three types of objects: in terms of time, in terms of space, and in terms of time and space. Extract the actual concerned time period for the one-dimensional sequence data, extract the rectangular concerned region image for the two-dimensional image data, and extract the rectangular concerned region image sequence for the two-dimensional image sequence data. The concerned time period and the rectangular concerned region are determined according to the actual task.
[0012] Further, in step S2, based on the data values in the attention period, rectangular attention area image, and rectangular attention area image sequence obtained in step S1, basic metrics, spatial metrics, temporal metrics, and spatio-temporal metrics are calculated. The basic metrics include the maximum value Max1 metric, minimum value Min1 metric, spatial mean M1 metric obtained through mean calculation, spatial coefficient of variation CV1 metric obtained through the coefficient of variation calculation process, maximum value Max2 metric, minimum value Min2 metric, temporal mean M2 obtained through mean calculation, and spatial coefficient of variation CV2 metric obtained through the coefficient of variation calculation process, which are obtained by screening the rectangular attention area image and the attention period. There are no restrictive requirements for the number of pixels in the rectangular attention area or the data length of the attention period; the spatial metrics include the regional difference D1 metric obtained through edge detection and difference calculation processes on the rectangular attention area image; the temporal metrics include the temporal difference D2 metric obtained through the Max2 - Min2 difference calculation process on the attention period and the slope B2 obtained through trend analysis and coefficient extraction; the spatio-temporal metric is the slope B1 obtained by calculating the difference through the spatial metric calculation process on the rectangular attention area image sequence and using the difference as time data points through the temporal metric calculation process. The spatial metrics, temporal metrics, and spatio-temporal metrics require that the number of edges E obtained through the edge detection process in the attention area pixels is greater than 0 or the data length L of the attention period is greater than 5.
[0013] Further, in step S3, each metric value calculated in step S2 can be expressed as the following relational expressions in the deep learning model:
[0014]
[0015] Among them, I represents any metric value calculated in step S2, x p is the input of the target contribution layer (purpose), that is, the input value of the target feature layer for the contribution of the output metric to be obtained, and g represents the relational expression between the input layer and the target contribution layer, represents the relational expression between the metric and the input layer. Among them, the relational expression g can be further expressed as:
[0016] g(x p ) = f n (f n-1 (f n-2 …(f p (x p ))));
[0017] In the above formula, f represents the non - linear relationship between the latter layer and the former layer in the deep learning model, n is the number of layers of superimposed non - linear operations, that is, the input to the output has experienced n non - linear operations, and its value is determined by the model structure. Here, n > p. When the number of layers n ≤ p, the calculation of the contribution layer has no practical significance and is not within the scope considered by this method.
[0018] Furthermore, the contribution of the target contribution layer to the said metric, that is, the partial derivative of the metric I with respect to the input of the target contribution layer can be expressed as:
[0019]
[0020] Furthermore, the affected gradient is the influence distribution of any of the said metrics affected by the target contribution layer, and the specific calculation process varies according to the differences of the metrics.
[0021] Furthermore, in step S4, classification is performed according to the value of the metric I and the value of the affected gradient and attribution analysis is carried out for the said categories. The classification includes trend analysis, variability analysis, anomaly analysis, and extremity analysis, which respectively represent trend attribution analysis, variability attribution analysis, anomaly attribution analysis, and extremity attribution analysis for the research object. Among them, the trend includes the metrics B1, B2, the variability includes the metrics CV1, CV2, the anomaly includes the metrics D1, D2, and the extremity includes the metrics Max1, Max2, Min1, Min2. The attribution analysis of the research object is to perform scenario analysis on the values of the respective metrics and the affected gradients corresponding to the metrics to obtain the influence of various characteristics of the research object by the target feature layer.
[0022] The second aspect of the present invention discloses a key index attribution analysis system for deep learning models in geoscience. The system extracts key indicators of features from the output results of any deep learning model for geoscience research tasks. Specifically, the system includes: a first processing unit configured to: obtain model output, screen out the objects to be analyzed for the affected situation, that is, the time period / region to be studied in the model output, divide the objects to be analyzed for the affected situation into three types of objects: in time, in space, and in time and space, and respectively obtain the analysis object categories for the actual concerned time period, concerned region, and research questions concerned about both at the same time, so as to determine the specific means differences for subsequent key indicator extraction; a second processing unit configured to: further calculate key indicators according to the divided analysis object categories to obtain a key indicator set, and the key indicator set includes basic indicators, time indicators, space indicators, and time-space indicators; a third processing unit configured to: match the affected analysis method according to the key indicator set, and obtain the affected situation of the key indicator set by inversely calculating the model gradient; a fourth processing unit configured to: according to the affected situation, combine the analysis object categories, and obtain the impact attribution of each feature of the model target layer on the output key characteristics, and the key characteristics include trendiness, variability, abnormality, and extremity.
[0023] Further, the first processing unit is specifically configured to obtain the analysis object category as a human judgment process, and the judgment basis is the model output data type and the actual research target. Among them: the model output data type includes one-dimensional sequence data, two-dimensional image data, and two-dimensional image sequence data, and the actual research target includes the trendiness, variability, and abnormality of the research object, and the research object includes important factors in the geoscience field such as precipitation, temperature, wind speed, vegetation coverage, and pollutant concentration that can be analyzed by the method system. According to the model output data type and the actual research target, the objects to be analyzed for the affected situation are divided into three types of objects: in time, in space, and in time and space. Respectively, the actual concerned time period is extracted from the one-dimensional sequence data, the rectangular concerned region image is extracted from the two-dimensional image data, and the rectangular concerned region image sequence is extracted from the two-dimensional image sequence data. The concerned time period and the rectangular concerned region are determined according to the actual task.
[0024] Further, the second processing unit is specifically configured to calculate basic metrics, spatial metrics, temporal metrics, and spatio-temporal metrics based on the data values in the attention period, rectangular attention area image, and rectangular attention area image sequence obtained by the first processing unit. The basic metrics include the maximum value Max1 metric, minimum value Min1 metric, spatial mean M1 metric obtained through mean calculation, spatial coefficient of variation CV1 metric obtained through the coefficient of variation calculation process, maximum value Max2 metric, minimum value Min2 metric, temporal mean M2 obtained through mean calculation, and spatial coefficient of variation CV2 metric obtained through the coefficient of variation calculation process, which are obtained by screening the rectangular attention area image and the attention period. There are no restrictive requirements for the number of pixels in the rectangular attention area or the data length of the attention period; the spatial metrics include the regional difference D1 metric obtained by the rectangular attention area image and through edge detection and difference calculation processes; the temporal metrics include the temporal difference D2 metric obtained by the attention period through the Max2 - Min2 difference calculation process and the slope B2 obtained through trend analysis and coefficient extraction; the spatio-temporal metrics are obtained by calculating the difference through the spatial metric calculation process for the rectangular attention area image sequence, and taking the difference as the time data point and obtaining the slope B1 through the temporal metric calculation process. The spatial metrics, temporal metrics, and spatio-temporal metrics require that the number of edges E obtained by the edge detection process for the pixels in the attention area is greater than 0 or the data length L of the attention period is greater than 5
[0025] Further, the third processing unit is specifically configured that each metric value calculated in the second processing unit can be expressed as the following relational expressions in the deep learning model:
[0026]
[0027] where I represents any metric value calculated in step S2, x p is the input of the target contribution layer (purpose), that is, the input value of the target feature layer for the contribution of the output metric to be obtained, and g represents the relational expression between the input layer and the target contribution layer, represents the relational expression between the metric and the input layer. Among them, the relational expression g can be further expressed as:
[0028] g(x p ) = f n (f n-1 (f n-2 …(f p (x p ))));
[0029] In the above formula, f represents the non - linear relationship between the latter layer and the former layer in the deep - learning model, and n is the number of layers of stacked non - linear operations, that is, the input to the output has experienced n non - linear operations, and its value is determined by the model structure. The n > p. When the number of layers n ≤ p, the calculation of the contribution layer has no practical significance and is not within the scope considered by this method. Further, the contribution of the target contribution layer to the index, that is, the partial derivative of the index I with respect to the input of the target contribution layer can be expressed as:
[0030]
[0031] Affected gradient is the influence distribution of any index affected by the target contribution layer, and the specific calculation process varies according to the different indexes.
[0032] Further, the fourth processing unit is specifically configured to classify according to the value of the index I and the value of the affected gradient and perform attribution analysis for the category. The classification includes trend analysis, variability analysis, anomaly analysis, and extremity analysis, which respectively represent trend attribution analysis, variability attribution analysis, anomaly attribution analysis, and extremity attribution analysis for the research object. Among them, the trend includes the indexes B1 and B2, the variability includes the indexes CV1 and CV2, the anomaly includes the indexes D1 and D2, and the extremity includes the indexes Max1, Max2, Min1, and Min2. The attribution analysis of the research object is to perform scenario analysis on the values of the respective indexes and the affected gradients corresponding to the indexes, and obtain the influence of various characteristics of the research object by the target feature layer.
[0033] Adopting the above - mentioned technical solution, the present invention has the following beneficial effects:
[0034] The technical solution provided by the present invention extracts and calculates the characteristic indexes reflecting the research object for the concerned area or time period, and further calculates the gradient in reverse to obtain the influence of the indexes by the specified feature layer, so as to selectively perform attribution analysis on important areas or important time periods. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following - described drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.
[0036] Figure 1Schematic flowchart of the method for attributing key indicators of a deep learning model for geoscience provided by an embodiment of the present invention;
[0037] Figure 2 Schematic structural diagram of the system for attributing key indicators of a deep learning model for geoscience provided by an embodiment of the present invention. Detailed implementation manners
[0038] The technical solutions of the present invention will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0039] The following details the specific implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the specific implementation manners described herein are only for explaining and illustrating the present invention and are not used to limit the present invention.
[0040] Combined with Figure 1 As shown, a first aspect of the present invention provides a method for attributing key indicators of a deep learning model for geoscience. The method extracts key indicators of features from the output results of any deep learning model for geoscience research tasks. The method specifically includes: Step S1, obtaining the model output, screening out the objects to be analyzed for the affected situation, that is, the time period / region to be studied in the model output, dividing the objects to be analyzed for the affected situation into three types of objects in terms of time, space, and time-space, respectively aiming at the research questions of the actual concerned time period, concerned region, and both concerned at the same time, obtaining the analysis object categories to determine the specific means differences for subsequent key indicator extraction; Step S2, further calculating key indicators according to the divided analysis object categories to obtain a key indicator set, calculating basic indicators according to all the divided analysis object categories, and calculating time indicators, space indicators, and time-space indicators respectively according to the divided object categories of the concerned time period, concerned region, and both concerned at the same time; Step S3, matching the affected analysis method according to the key indicator set, and obtaining the affected situation of the key indicator set by reversely calculating the model gradient; Step S4, according to the affected situation, combining the analysis object categories, obtaining the key feature influence situation of each input of the model on the output, and the key features include trend, variability, abnormality, and extremity.
[0041] According to the method of the first aspect, in step S1, the process of obtaining the analysis object category is a human judgment process, and the judgment basis is the model output data type and the actual research objective. Among them: the model output data type includes one-dimensional sequence data, two-dimensional image data, and two-dimensional image sequence data; the actual research objective includes the trend, variability, and abnormality of the research object, and the research object includes precipitation, temperature, wind speed, vegetation coverage, pollutant concentration, etc., which are important factors in the field of earth science that can be analyzed by the method system. According to the model output data type and the actual research objective, the objects to be analyzed for the affected situation are divided into three types of objects in terms of time, space, and time-space. Respectively, for the one-dimensional sequence data, extract the actual concerned time period, for the two-dimensional image data, extract the rectangular concerned area image, and for the two-dimensional image sequence data, extract the rectangular concerned area image sequence. The concerned time period and the rectangular concerned area are determined according to the actual task.
[0042] According to the method of the first aspect, in step S2, based on the data values in the concerned time period, rectangular concerned area image, and rectangular concerned area image sequence obtained in step S1, calculate the basic index, spatial index, time index, and time-space index.
[0043] The process of calculating the index in step S2 includes two branches. The first branch is the basic index branch. The basic index consists of the maximum value Max1 index, minimum value Min1 index obtained by screening the rectangular concerned area image, the spatial mean M1 index obtained by mean calculation, the spatial coefficient of variation CV1 index obtained by the coefficient of variation calculation process, and the maximum value Max2 index, minimum value Min2 index obtained by screening the concerned time period, the time mean M2 obtained by mean calculation, and the spatial coefficient of variation CV2 index obtained by the coefficient of variation calculation process. The basic index has no restrictive requirements on the number of pixels in the rectangular concerned area or the data length of the concerned time period. The second branch is the special index branch, including spatial, time, and time-space indexes. The spatial index includes the regional difference D1 index obtained by the rectangular concerned area image and the edge detection process and difference calculation process. The difference calculation process is the difference between the mean value I of the internal points of the edge obtained by the edge detection and the mean value I of the external points. mean-in and the external point mean value I mean-out of the difference I mean-in -I meaf-out ; the time index is the time difference D2 index obtained by the Max2 - Min2 calculation process for the concerned time period and the slope B2 obtained by the trend analysis process and coefficient extraction. The slope is the slope of the fitting equation obtained when fitting the data points of the concerned time period using the linear regression method. The fitting equation satisfies specific parameter values a and b such that the residual sum of squares (RSS) is minimized:
[0044]
[0045] where t represents the independent variable time; y is the dependent variable, representing the data value at the corresponding time point i; and L is the number of moments corresponding to the concerned period. When the RSS is minimized by the a and b, the following is satisfied:
[0046]
[0047] From the above two equations, the slope b of the fitting equation is:
[0048]
[0049] Generally, the time t of the model output data starts counting from 1, and the slope B2 is:
[0050]
[0051] The spatio-temporal index is obtained by calculating the difference from the image sequence of the rectangular concerned area through the spatial index calculation process, and using the difference as the time data point to obtain the slope B1 through the time index calculation process.
[0052] The calculation process of B1 is as follows:
[0053]
[0054] where y mi is the dependent variable, representing the difference obtained by passing the image at the corresponding time point i through the spatial index calculation process.
[0055] According to the method of the first aspect, for the spatial index, time index, and spatio-temporal index, it is required that the number of edges E obtained by the edge detection process for the pixels in the concerned area satisfies E > 0 or the data length L of the concerned period is greater than 5. The edge detection method can calculate the gradient by using the Sobel operator and set a threshold for edge detection, and use the contour tracking algorithm to generate a closed curve. The edge detection process is only one of the processes for extracting abnormal areas or key areas in the present invention, and other methods that can achieve the same function can be used for replacement. The edge detection process is not within the scope of protection of the present invention.
[0056] According to the method of the first aspect, in step S3, for each index value calculated in step S2, it can be expressed as the following relational expression in the deep learning model:
[0057]
[0058] where I represents any index value calculated in step S2, and x pis the input of the target contribution layer (purpose), that is, the input value of the target feature layer for the contribution of the output index to be obtained. g represents the relational expression between the input layer and the target contribution layer. represents the relational expression between the said index and the input layer. Among them, the relational expression g can be further expressed as:
[0059] g(x p ) = f n (f n-1 (f n-2 …(f p (x p ))));
[0060] In the above formula, f represents the non - linear relational expression between the latter layer and the former layer in the deep learning model. n is the number of layers of the stacked non - linear operations, that is, the input to the output has experienced n non - linear operations, and its value is determined by the model structure. The said n > p. When the number of layers n ≤ p, the calculation of the contribution layer has no practical significance and is not within the scope considered by the said method. Further, the contribution of the target contribution layer to the said index, that is, the partial derivative of the said index I with respect to the input of the target contribution layer can be expressed as:
[0061]
[0062] According to the method of the first aspect, the affected gradient is the influence distribution of the said any index affected by the target contribution layer, and the specific calculation process varies according to the difference of the said index.
[0063] Specifically, the influence of the basic index extreme value index m (including Max1, Min1, Max2, Min2) by the target contribution layer is obtained by the following process:
[0064]
[0065] Among them, m represents the position where the extreme value is located in the object of concern.
[0066] The influence of the basic index mean index mean (including M1 and M2) by the target contribution layer is obtained by the following process:
[0067]
[0068] Among them, i is any position (pixel / time - series data) in the object of interest (ooi). The object of interest includes the area of concern or the time period of concern or the image sequence of the area of concern. Z is the total number of positions included in the ooi. Finally, the derivative of the target I mean with respect to x is
[0069] The coefficient of variation CV of the basic index (including CV1 and CV2), and its corresponding influence by the target contribution layer is obtained through the following process:
[0070] g i (x p ) = f in (f n-1 (f n-2 …(f p (x p ))));
[0071]
[0072] The spatial index includes the regional difference D1 index, and its corresponding influence by the target contribution layer is obtained through the following process:
[0073]
[0074] where rin represents the region inside the edge obtained by the edge detection calculation, and rout represents the external region.
[0075] The time indexes, namely the time difference D2 index and the slope B2, and their corresponding influences by the target contribution layer are obtained through the following process:
[0076]
[0077] The influence of the spatio-temporal index slope B1 by the target contribution layer The calculation process is as follows:
[0078]
[0079] According to the method of the first aspect, in step S4, classification is performed according to the value of the index I and the value of the affected gradient , and attribution analysis is performed for the category. The classification includes trend analysis, variability analysis, abnormality analysis, and extremity analysis, which respectively represent trend attribution analysis, variability attribution analysis, abnormality attribution analysis, and extremity attribution analysis for the research object. Among them, the trend includes the indexes B1 and B2, and the corresponding attribution is the affected contribution distribution and The variability includes the indexes CV1 and CV2, and the corresponding attribution is the affected contribution distribution The abnormality includes the indexes D1 and D2, and the corresponding attribution is the affected contribution distribution and The extremeness includes indices Max1, Max2, Min1, and Min2, corresponding to the affected contribution distribution of attribution. By analyzing the affected gradients corresponding to the respective indices The high-value and low-value regions of the distribution are used to obtain the influence of various characteristics of the research object by the target feature layer.
[0080] Combined with Figure 2 As shown, the second aspect of the present invention provides a key index attribution analysis system for deep learning models in earth science. The system 100 extracts key indices of features from the output results of any deep learning model for earth science research tasks. The system specifically includes: a first processing unit 101 configured to: obtain the model output, screen out the objects to be analyzed for the affected situation, that is, the time period / region to be studied in the model output, divide the objects to be analyzed for the affected situation into three types of objects: temporal, spatial, and spatio-temporal, and respectively obtain the analysis object categories for the research questions of the actual concerned time period, concerned region, and both simultaneously concerned, so as to determine the specific means differences for subsequent key index extraction; a second processing unit 102 configured to: further calculate key indices according to the divided analysis object categories to obtain a key index set, the key index set including basic indices, time indices, spatial indices, and spatio-temporal indices; a third processing unit 103 configured to: match the affected analysis method according to the key index set, and obtain the affected situation of the key index set by reverse calculating the model gradient; a fourth processing unit 104 configured to: according to the affected situation, combined with the analysis object categories, obtain the influence of the model target layer features on the key characteristics of the output, the key characteristics including trendiness, variability, abnormality, and extremeness.
[0081] According to the system of the second aspect, the first processing unit determines the analysis object category through an artificial judgment process. The judgment basis includes the model output data type and the actual research target. Specifically, the model output data type can be one-dimensional sequence data, two-dimensional image data, or two-dimensional image sequence data; while the actual research target includes the trendiness, variability, and abnormality of the research object. Common research objects include important factors such as precipitation, temperature, wind speed, vegetation coverage, and pollutant concentration, all of which are objects that can be analyzed using the method of the present invention in the field of earth science. According to the model output data type and research target, the objects to be analyzed are divided into three categories: temporal objects, spatial objects, and spatio-temporal objects. Specifically, for one-dimensional sequence data, the actually concerned time period is extracted therefrom; for two-dimensional image data, the rectangular concerned region image is extracted; and for two-dimensional image sequence data, the rectangular concerned region image sequence is extracted. The selection of the time period and the rectangular concerned region is determined by the actual task requirements.
[0082] For the system according to the second aspect, the second processing unit calculates and generates relevant metrics based on the data values in the concerned time period, the rectangular concerned region image, and the sequence of rectangular concerned region images obtained by the first processing unit. Specifically, the basic metrics include the following: the maximum value Max1 metric, the minimum value Min1 metric obtained by screening the rectangular concerned region image, the spatial mean M1 metric obtained by calculating the mean value, the spatial coefficient of variation CV1 metric obtained through the coefficient of variation calculation process, and the maximum value Max2 metric, the minimum value Min2 metric obtained by screening the concerned time period, the time mean M2 obtained by calculating the mean value, and the spatial coefficient of variation CV2 metric obtained through the coefficient of variation calculation process. The spatial metric includes the regional difference D1 metric obtained from the rectangular concerned region image and through the edge detection process and the difference calculation process. The difference calculation process is the difference I mean-in between the mean value I mean-out of the internal points of the edge obtained by the edge detection and the mean value I mean-in -I mean-out of the external points; the time metric is the time difference D2 metric obtained by the Max2 - Min2 calculation process for the concerned time period and the slope B2 obtained through the trend analysis process and coefficient extraction; the slope is the slope of the fitting equation obtained when fitting the data points of the concerned time period using the linear regression method. The fitting equation satisfies specific parameter values a and b such that the residual sum of squares (RSS) is minimized:
[0083]
[0084] where t represents the independent variable time; y is the dependent variable representing the data value at the corresponding time point i; and L is the number of time instances corresponding to the concerned time period. When a and b minimize RSS, it satisfies:
[0085]
[0086] From the above two equations, the slope b of the fitting equation can be obtained as:
[0087]
[0088] Generally, the time t of the model output data starts counting from 1, and the slope B2 is
[0089]
[0090] The spatio-temporal metric is obtained by calculating the difference from the sequence of rectangular concerned region images through the spatial metric calculation process, and using the difference as the time data points through the time metric calculation process to obtain the slope B1.
[0091] The calculation process of B1 is as follows:
[0092]
[0093] where y mi is the dependent variable, representing the difference obtained by the image at the corresponding time point i through the spatial index calculation process.
[0094] According to the system of the second aspect, the calculation requirements of the spatial index, time index, and spatio-temporal index in the second processing unit require that the number of pixels in the region of interest satisfy the following conditions: the number of edges E obtained by the edge detection process > 0, or the data length L of the period of interest > 5.
[0095] According to the system of the second aspect, the third processing unit is specifically configured as follows: for each index value calculated in the second processing unit, the relationship in the deep learning model can be expressed by the following formula:
[0096] where I represents any index value calculated in step S2, and x p is the input of the target contribution layer (purpose), that is, the input value of the target feature layer for the contribution of the output index to be obtained, and g represents the relationship formula between the input layer and the target contribution layer. represents the relationship formula between the index and the input layer. Among them, the relationship formula g can be further expressed as:
[0097] g(x p ) = f n (f n-1 (f n-2 …(f p (x p ))));
[0098] In the above formula, f represents the non-linear relationship formula between the latter layer and the former layer in the deep learning model, n is the number of layers of superimposed non-linear operations, that is, the input to the output has experienced n non-linear operations, and its value is determined by the model structure. The n > p. When the number of layers n ≤ p, the calculation of the contribution layer has no practical significance and is not within the scope of consideration of the method. Further, the contribution of the target contribution layer to the index, that is, the partial derivative of the index I with respect to the input of the target contribution layer can be expressed as:
[0099]
[0100] According to the method of the first aspect, the affected gradient is the influence distribution of any index affected by the target contribution layer, and the specific calculation process varies according to the different indexes.
[0101] Specifically, the influence of the maximum and minimum values of the basic indicators m (including Max1, Min1, Max2, and Min2) on the target contribution layer is calculated through the following process:
[0102]
[0103] where m represents the position of the maximum or minimum value in the object of interest.
[0104] The influence of the average value of the basic indicators mean (including M1 and M2) on the target contribution layer is calculated through the following process:
[0105]
[0106] where i is any position (pixel / temporal data) in the object of interest (ooi), and the object of interest includes the area of interest, the time period of interest, or the image sequence of the area of interest. Z is the total number of positions included in the ooi, and the final calculation target is I mean The derivative of with respect to x is
[0107] The coefficient of variation CV of the basic indicators (including CV1 and CV2), and its corresponding influence on the target contribution layer is calculated through the following process:
[0108] g i (x p ) = f in (f n-1 (f n-2 …(F p (x p ))));
[0109]
[0110] The spatial indicator includes the regional difference D1 indicator, and its corresponding influence on the target contribution layer is calculated through the following process:
[0111]
[0112] where rin represents the area inside the edge obtained by the edge detection calculation, and rout represents the external area.
[0113] The time indicators, the time difference D2 indicator and the slope B2, and their corresponding influences on the target contribution layer are calculated through the following process:
[0114]
[0115] The calculation process of the influence of the spatio-temporal index slope B1 by the target contribution layer is as follows:
[0116]
[0117] According to the system of the second aspect, the fourth processing unit is specifically configured to classify according to the value of the index I and the affected gradient value, and perform attribution analysis for each category. The classification includes trend analysis, variability analysis, abnormality analysis, and extremity analysis, which respectively represent trend attribution, variability attribution, abnormality attribution, and extremity attribution for the research object.
[0118] Specifically:
[0119] The trend analysis includes indices B1 and B2, and correspondingly attributes the affected contribution distribution and
[0120] The variability analysis includes indices CV1 and CV2, and correspondingly attributes the affected contribution distribution
[0121] The abnormality analysis includes indices D1 and D2, and correspondingly attributes the affected contribution distribution and
[0122] The extremity analysis includes indices Max1, Max2, Min1, and Min2, and correspondingly attributes the affected contribution distribution
[0123] By analyzing the high-value areas and low-value areas of the affected gradient distribution corresponding to each index, the influence of the research object by the target feature layer under different characteristics (such as trend, variability, abnormality, and extremity) can be obtained.
[0124] The technical solution provided by the present invention combines the output results of the deep learning model, gradient backpropagation, edge detection, and trend regression analysis to perform affected attribution analysis on the key features in the field of earth science. This method can effectively quantify the influence of different input factors on the target features, improve the model interpretability, and provide a scientific basis for extreme weather prediction, flood assessment, and climate change research. By classifying and analyzing characteristics such as trend, variability, abnormality, and extremity, the present invention can be flexibly applied in various research tasks and provide accurate input factor influence analysis for decision-making in related fields.
[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for attributing key indicators of a deep learning model for geoscience, characterized in that, It includes the following steps: S1. Obtain the model output, screen out the objects to be analyzed for the affected situation, that is, the time period / region to be studied in the model output, and classify the objects to be analyzed for the affected situation; S2. According to the classified categories of the analysis objects, further calculate the key indicators to obtain a set of key indicators; S3. Match the affected analysis method according to the set of key indicators, and obtain the affected situation of the set of key indicators by reversely calculating the model gradient; S4. According to the affected situation, combined with the category of the analysis object, obtain the influence of each input of the model on the key features of the output; Among them, the influence of the key features is used to analyze the relationship and action mechanism between factors in the field of earth science, especially as an effective visual evidence for the complex and variable relationship between factors under the background of global climate change.
2. The method for attributing key indicators of a deep learning model for geoscience according to claim 1, wherein In step S1, according to the data type of the deep learning model output and various elements in the field of earth science in actual research, the object categories are divided, and the objects to be analyzed for the affected situation are divided into three categories: time, space, and time-space.
3. The method for attributing key indicators of a deep learning model for geoscience according to claim 1, wherein In step S2, according to the data values in the concerned time period, rectangular concerned area image, and rectangular concerned area image sequence obtained in S1, basic indicators, spatial indicators, time indicators, and time-space indicators are calculated.
4. The method for attributing key indicators of a deep learning model for geoscience according to claim 1, characterized in that, In step S3, matching the affected analysis method according to the set of key indicators and obtaining the affected situation of the set of key indicators by reversely calculating the model gradient includes: The influence of the maximum and minimum value index m of the basic index on the target contribution layer The influence of the mean index mean of the basic index on the target contribution layer The influence of the coefficient of variation CV of the basic index on the target contribution layer The influence of the regional difference D1 index in the spatial index on the target contribution layer The influence of the time difference D2 index in the time index on the target contribution layer The influence of the time difference slope B2 in the time index on the target contribution layer And the influence of the spatio-temporal index slope B1 on the target contribution layer 5. The key index attribution analysis method for deep learning models in geoscience according to claim 4, characterized in that, The basic index extreme value index m, including Max1, Min1, Max2, Min2; the influence situation by the target contribution layer Obtained by the following process: The basic index mean index mean, including M1 and M2; the influence situation of the target contribution layer Obtained by the following process: Coefficient of variation CV of the basic indicator, including CV1 and CV2; The influence thereof corresponding to the target contribution layer Obtained by the following process: The regional difference D1 index in the spatial index, and its corresponding influence by the target contribution layer Obtained by the following process: For the time difference D2 index in the time indicators, it respectively corresponds to the influence of the target contribution layer Obtained by the following process: The time difference slope B2 in the time index, and its corresponding influence by the target contribution layer Obtained by the following process: The slope B1 of the space-time index, corresponding to the influence of the target contribution layer Obtained by the following process:
6. The method for attributing key indicators of a deep learning model for geoscience according to claim 1, characterized in that, In step S4, according to the influence situation obtained in step S3 of the distribution, respectively representing the trend attribution, variability attribution, abnormality attribution and extremity attribution for the research object, and obtain the influence situation of the research object under different characteristics by the target feature layer.
7. A system for the attribution analysis of key indicators of a deep learning model for earth sciences according to any one of claims 1-6, characterized in that, The system specifically includes: The first processing unit is configured to: obtain the model output, screen out the objects to be analyzed for the affected situation, and divide the objects to be analyzed for the affected situation into three categories: time, space, and time-space; The second processing unit is configured to: further calculate the key indicators according to the classified categories of the analysis objects to obtain a set of key indicators; The third processing unit is configured to: match the affected analysis method according to the set of key indicators, and obtain the affected situation of the set of key indicators by reversely calculating the model weights; The fourth processing unit is configured to: according to the affected situation, combined with the category of the analysis object, obtain the attribution of the influence of each feature of the model target layer on the key characteristics of the output.