A geological disaster risk assessment method, device, medium and product

Abstract

The application discloses a geological disaster risk evaluation method, device, medium and product, and relates to the field of data processing. The method comprises the following steps: dividing an evaluated region into a source domain and a target domain according to the density of geological environment data; determining cross-regional shared features and performing dimension reduction and denoising optimization to obtain an optimized feature set; constructing a domain adversarial neural network containing a gradient inversion layer, and pre-training a source domain pre-training model based on the source domain optimized feature set; in the adversarial training, the gradient direction is reversed through the gradient inversion layer, the feature distribution difference between the source domain and the target domain is reduced, and a cross-regional adaptive model is obtained; the target domain expanded data is obtained by performing data expansion processing on the target domain optimized feature set by using the model; and finally, the geological disaster risk evaluation result of the target domain is obtained by performing geological disaster risk evaluation according to the target domain expanded data. The application can alleviate the problem that the evaluation precision is insufficient due to insufficient data in the data scarce region.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular relates to a geological disaster risk assessment method, device, medium and product. BACKGROUND

[0002] Under the background of global climate change and intensified human engineering activities, sudden geological disasters occur frequently, causing a large number of life and property losses. Therefore, geological disaster risk assessment can effectively reduce personnel casualties and property losses. Geological disaster risk assessment can quantitatively evaluate the spatial probability and potential damage intensity of disasters such as landslides and debris flows, and provide scientific basis for land space planning, infrastructure site selection and emergency management.

[0003] In related technologies, geological disaster risk assessment mainly relies on comprehensive collection and analysis of geological environment data. By integrating elevation topographic data, lithology distribution map, historical disaster point spatial coordinates and real-time meteorological monitoring data and other multi-source information, spatial analysis is performed using a geographic information system (GIS), and statistical models or machine learning algorithms are combined to assess the possibility and intensity of geological disasters. These schemes can provide relatively accurate and reliable evaluation results in data-rich areas, and provide scientific basis for the prevention and control of geological disasters.

[0004] However, the density of geological environment data in an area is often not uniform, with some areas having abundant data and some areas having scarce data. For data-scarce areas, there are situations such as incomplete geological environment data, missing meteorological data, and lack of historical disaster records. Therefore, using traditional evaluation methods to evaluate the geological disaster risk of data-scarce areas will not be able to accurately assess disaster risks due to lack of sufficient data support. SUMMARY

[0005] In view of the above technical problems and defects, the purpose of the present application is to provide a geological disaster risk assessment method, device, medium and product, which can alleviate the problem of insufficient evaluation accuracy caused by insufficient data in data-scarce areas.

[0006] To achieve the above object, in a first aspect, the present application provides a geological disaster risk evaluation method, comprising: dividing an evaluated region into a source domain and a target domain according to the density of geological environment data of the evaluated region, wherein the data density of the source domain is greater than that of the target domain; determining cross-regional shared features according to the geological environment data of the source domain and the target domain, the cross-regional shared features including at least one of slope, lithology combination and historical disaster density; performing dimensionality reduction denoising optimization on the cross-regional shared features by principal component analysis to obtain an optimized feature set, the optimized feature set including a source domain optimized feature set and a target domain optimized feature set; in the case of constructing a domain adversarial neural network containing a gradient reversal layer as a preset deep learning model, pre-training the deep learning model based on the source domain optimized feature set to obtain a source domain pre-training model; in the process of adversarial training of the source domain pre-training model, reversing the gradient direction when back propagation through the gradient reversal layer to reduce the feature distribution difference between the source domain and the target domain to obtain a cross-regional adaptation model; performing data expansion processing on the target domain optimized feature set by the cross-regional adaptation model to obtain target domain expansion data; and performing geological disaster risk evaluation according to the target domain expansion data to obtain a geological disaster risk evaluation result of the target domain.

[0007] The present application divides the evaluated region into a source domain and a target domain, determines cross-regional shared features based on the geological environment data of the data-intensive source domain and the data-sparse target domain, and performs dimensionality reduction denoising optimization to obtain a more refined feature set. Further, the domain adversarial neural network containing the gradient reversal layer is used to pre-train on the source domain data, and then the adversarial training is used to reduce the feature distribution difference between the source domain and the target domain to finally obtain a cross-regional adaptation model. The cross-regional adaptation model can perform data expansion processing on the optimized feature set of the target domain to generate more abundant data, and further realize more accurate geological disaster risk evaluation of the target domain. The whole process fully exploits the data advantage of the source domain, while adapting to the data characteristics of the target domain, effectively making up for the data deficiency of the target domain, improving the evaluation accuracy, and alleviating the problem of insufficient evaluation accuracy caused by insufficient data in the data-scarce region.

[0008] Optionally, in some embodiments, during the process of adversarial training on the source domain pre-training model, the following steps are included: generating a geological disaster spatial correlation matrix of the target domain according to the spatial distribution characteristics of the geological disasters in the target domain; constructing a spatial continuity constraint function based on the geological disaster spatial correlation matrix, the spatial continuity constraint function being used to quantify the spatial correlation strength of the geological disasters; generating spatial constraint features that satisfy domain distribution consistency and geological spatial continuity according to the spatial continuity constraint function; evaluating the model generalization performance of the target domain after feature alignment based on the spatial constraint features to obtain a performance evaluation index; determining the initial weight of the source domain and the target domain according to the performance evaluation index; and dynamically adjusting the weight distribution of the source domain and the target domain according to the real-time data increment of the target domain and the initial weight, so as to gradually reduce the source domain dependency.

[0009] By introducing the geological disaster spatial correlation matrix and the spatial continuity constraint function, the adaptability of the model to the spatial distribution characteristics of the geological disasters in the target domain is significantly improved. During the adversarial training process, this method not only reduces the feature distribution difference between the source domain and the target domain, but also ensures that the generated spatial constraint features can reflect the spatial continuity and correlation of the geological disasters. By determining the initial weight through the performance evaluation index and dynamically adjusting the weight distribution according to the real-time data increment, the dependency on the source domain is gradually reduced, and the generalization ability and prediction accuracy of the model in the target domain are enhanced. This method is particularly suitable for regions where the spatial distribution of geological disasters has obvious correlation.

[0010] Optionally, in some embodiments, dynamically adjusting the weight distribution of the source domain and the target domain according to the real-time data increment and the initial weight of the target domain includes: determining the size of the newly added data and the time series information of the target domain according to the real-time data increment of the target domain; determining the dynamic adjustment coefficient of the source domain and the target domain in transfer learning based on the size of the newly added data and the time series information; and updating the weight distribution ratio of the source domain and the target domain in model training according to the dynamic adjustment coefficient and the initial weight, so as to gradually reduce the proportion of the source domain weight.

[0011] By dynamically adjusting the weight distribution of the source domain and the target domain, the above-mentioned embodiment solves the problem of insufficient model adaptability caused by fixed weights in traditional transfer learning. In practical applications, this method can flexibly adjust the weight distribution ratio in model training according to the real-time data increment of the target domain, so that the model can better adapt to the actual situation of the target domain. By determining the size of the newly added data and the time series information, calculating the dynamic adjustment coefficient, and updating the weight distribution ratio combined with the initial weight, the flexibility and adaptability of the model are improved, and the stability of the model training and the reliability of the evaluation results are ensured. This method is particularly suitable for geological disaster risk assessment in target domains with dynamic data changes.

[0012] Optionally, in some embodiments, based on the scale of the new data and the time series information, the dynamic adjustment coefficient of the source domain and the target domain in the transfer learning is determined, including: determining the target domain new data amount at multiple moments according to the scale of the new data and the time series information; and calculating the dynamic adjustment coefficient according to the target domain new data amount at multiple moments and a preset coefficient adjustment formula, the coefficient adjustment formula including:

[0013] ;

[0014] wherein, λ ( t ) represents the dynamic adjustment coefficient, λ 0 represents the initial adjustment coefficient, β represents the source domain weight decay rate, γ represents the target domain data increment influence coefficient, ΔD( τ ) represents the target domain new data amount at moment τ, D total represents the target domain expected total data amount, k represents the time decay factor of the increment data, t represents the current time, τ is used to traverse all time points from the initial moment to the current time t, e is a natural constant.

[0015] By introducing the calculation formula of the dynamic adjustment coefficient, the above-mentioned technical scheme provides a clear mathematical basis and operability for weight adjustment. The formula comprehensively considers the scale of the new data of the target domain, the time series information and multiple preset parameters, and can accurately reflect the influence of the data increment of the target domain on the weight adjustment. In actual application, this method not only improves the scientificity and accuracy of the weight adjustment, but also makes the entire adjustment process smoother and more stable, avoiding the mutation and instability phenomenon in the weight adjustment process. By updating the weight distribution proportion with the dynamic adjustment coefficient and the initial weight, the proportion of the source domain weight is gradually reduced, and the influence of the target domain in the model training is enhanced, thereby improving the generalization performance of the model and the accuracy of the evaluation result.

[0016] Optionally, in some embodiments, the weight distribution proportion of the source domain and the target domain in the model training is updated according to the dynamic adjustment coefficient and the initial weight, including: calculating the weight distribution proportion according to a preset weight distribution formula, the weight distribution formula including:

[0017] ;

[0018] w g ( t )=1- w s ( t ).

[0019] wherein, w s t represents the weight proportion of the source domain, w g t represents the weight proportion of the target domain, λ t represents a dynamic adjustment coefficient, w i represents an initial weight, α represents a dynamic coefficient sensitivity, η represents a weight transfer rate, t represents a current time, τ for traversing all time points from an initial time to the current time t, e is a natural constant.

[0020] By using the technical solutions of the above embodiments, the weight distribution proportion of the source domain and the target domain in model training is calculated by a preset weight distribution formula. The formula is based on a dynamic adjustment coefficient and an initial weight, and by introducing parameters such as a natural constant, a dynamic coefficient sensitivity, and a weight transfer rate, the weight proportions of the source domain and the target domain are accurately calculated. In practical applications, this method not only dynamically reflects the influence of data changes on the weight, but also makes the weight adjustment process smoother and more stable. As the target domain data continues to increase, the weight proportion of the source domain gradually decreases, and the weight proportion of the target domain correspondingly increases, so that the model can better reflect the actual situation of the target domain, and improve the accuracy and timeliness of the evaluation results. The sudden and unstable phenomena in the weight adjustment process are effectively avoided, and the stability of the model training and the reliability of the evaluation results are improved.

[0021] Optionally, in some embodiments, the geological disaster risk evaluation is performed according to the target domain expansion data to obtain a geological disaster risk evaluation result of the target domain, including: dividing the target domain into a plurality of evaluation units according to the target domain expansion data; determining a susceptibility comprehensive index and a instability probability of each evaluation unit based on the expanded terrain data, human activity index, and meteorological data of each evaluation unit, the susceptibility comprehensive index being used to represent the possibility of geological disaster occurrence of the evaluation unit, and the instability probability being used to represent the probability of the evaluation unit losing stability and thus causing geological disaster occurrence; determining a geological disaster risk index according to the susceptibility comprehensive index and the instability probability; and generating the geological disaster risk evaluation result based on the geological disaster risk index.

[0022] ​​​According to the technical scheme of the above embodiment, how to evaluate the geological disaster risk according to the target domain expansion data is described in detail. First, the target domain is divided into multiple evaluation units, and then the susceptibility comprehensive index and the instability probability are determined based on the expanded terrain data, human activity index and meteorological data of each evaluation unit. By integrating these multi-source data, the geological disaster risk index is calculated by using a preset geological disaster risk index formula, and the evaluation result is generated. This method can not only more accurately reflect the geological disaster risk condition of the target domain, but also provide more detailed and scientific basis for the formulation of geological disaster prevention measures, and is especially suitable for the fine evaluation of the geological disaster risk of the target domain after data expansion.

[0023] Optionally, in some embodiments, determining the geological disaster risk index based on the susceptibility comprehensive index and the instability probability comprises: calculating the geological disaster risk index according to a preset risk index formula, and the risk index formula comprises: H i =Y i / Y max ×P i ;

[0024] wherein H i represents the geological disaster risk index of the i th evaluation unit under a certain rainfall condition, Y i represents the susceptibility comprehensive index of the i th evaluation unit, Y max represents the maximum value of the susceptibility comprehensive index, and P i represents the instability probability.

[0025] According to the technical scheme of the above embodiment, the calculation method of the geological disaster risk index is provided by using the preset risk index formula. The formula combines the susceptibility comprehensive index and the instability probability organically, and makes them consistent in dimension through normalization processing, so as to realize the double consideration of the possibility and intensity of geological disasters. In practical application, this method can not only more comprehensively reflect the risk degree of geological disasters, but also provide a standardized quantitative evaluation means for geological disaster risk evaluation. By calculating the geological disaster risk index, the high-risk areas can be more accurately identified, the prevention and control resources can be reasonably allocated, the overall efficiency of geological disaster prevention and control can be improved, and this method is especially suitable for geological disaster evaluation projects that need to be updated and optimized for a long time.

[0026] In a second aspect, the embodiments of the present application provide an electronic device, comprising: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code comprises computer instructions, and the one or more processors invoke the computer instructions to make the electronic device execute the method described in the first aspect or the second aspect and any possible implementation manner of the first aspect or the second aspect.

[0027] In a third aspect, the present application provides a computer readable storage medium comprising instructions which, when executed on the electronic device, cause the electronic device to perform the method according to the first aspect or the second aspect, and any possible implementation of the first aspect or the second aspect.

[0028] In a fourth aspect, the present application provides a computer program product comprising instructions which, when executed on the electronic device, cause the electronic device to perform the method according to the first aspect or the second aspect, and any possible implementation of the first aspect or the second aspect.

[0029] It can be understood that the electronic device provided by the second aspect, the storage medium provided by the third aspect and the computer program product provided by the fourth aspect are all used to execute the method provided by the present application. Therefore, the beneficial effects that can be achieved are referred to the beneficial effects in the corresponding method, which will not be described here.

[0030] The one or more technical solutions provided by the present application have at least the following technical effects or advantages:

[0031] 1. By dividing the evaluation area into a source domain and a target domain, pre-training the model using the rich data of the source domain, and then through adversarial training and gradient reversal layer, the model can adapt to the data characteristics of the target domain, thereby effectively solving the problem of insufficient evaluation accuracy caused by the scarcity of target domain data. At the same time, the dynamic weight adjustment is made for the real-time data increment of the target domain, which further improves the adaptability of the model to the target domain and the accuracy of the evaluation results.

[0032] 2. The domain adversarial neural network comprising the gradient reversal layer is constructed, the gradient direction is reversed during back propagation to reduce the feature distribution difference between the source domain and the target domain, so that the model can achieve good performance on the source domain and the target domain. At the same time, a spatial continuity constraint function is introduced in the adversarial training process to generate spatial constraint features that meet the domain distribution consistency and geological spatial continuity, further improving the generalization ability of the model and the adaptability to the target domain.

[0033] 3. According to the real-time data increment of the target domain, the size of the new data and the time series information are determined, and then the dynamic adjustment coefficient is determined. Combined with the initial weight, the weight distribution ratio of the source domain and the target domain in the model training is updated, and the proportion of the source domain weight is gradually reduced. This dynamic weight adjustment mechanism enables the model to adapt to changes in target domain data in a timely manner, improving the flexibility of the model and the reliability of the evaluation results. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a flowchart of a geological disaster risk evaluation method according to an embodiment of the present application.

[0035] Figure 2 Figure 1 is a schematic diagram of an architecture of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] The terminology used in the present application merely for the purpose of describing particular embodiments and is not intended to be limiting of the present application. As used in the description of the application, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0037] As shown in Figure 1 Figure 1, the present application provides a geological disaster risk assessment method, which comprises the following steps, specifically:

[0038] Step 201, dividing the evaluation area into source domain and target domain according to the density of geological environment data in the evaluation area.

[0039] Wherein, the data density of the source domain is greater than that of the target domain. The source domain is the data-intensive area, and the target domain is the data-scarce area.

[0040] Specifically, first, collect various types of geological environment data of the evaluation area, including elevation topographic data, lithology distribution map, historical disaster point spatial coordinates, real-time meteorological monitoring data, etc. Then, evaluate the integrity and density of these data, quantify the data density by indicators such as the number of data points per unit area, data missing rate and noise level, etc. According to the preset data density threshold or through clustering analysis, etc., divide the evaluation area into source domain and target domain.

[0041] The source domain is usually located in areas with convenient transportation and perfect monitoring facilities, data acquisition is relatively easy, data quantity is abundant and quality is high; the target domain is mostly remote and difficult to monitor, data is scarce and may have more missing and errors.

[0042] Through such division, the data characteristics of different regions are clarified, laying a foundation for subsequent cross-regional feature analysis and data processing, so that data processing strategies suitable for each region can be targetedly adopted, improving the scientificity and accuracy of the overall evaluation.

[0043] Step 202, determining the cross-regional shared features according to the geological environment data of the source domain and the target domain.

[0044] The cross-region shared features include at least one of slope, lithology combination and historical disaster density. The slope refers to the inclination degree of a slope, is usually expressed by an angle or a percentage, and reflects the degree of topographic relief. The lithology combination refers to the collocation and distribution of different rock types in a certain region, and affects the geological structure and stability. The historical disaster density refers to the number or frequency of past geological disasters per unit area, and reflects the frequency of occurrence of regional geological disasters.

[0045] Specifically, first, based on the geological prior knowledge of the geological disaster formation mechanism, candidate features with physical meaning and strong spatial generalization (such as slope, lithology combination, and historical disaster density) are screened out. Secondly, through the dual-domain (source domain and target domain) data distribution alignment test, the distribution difference of the source domain and the target domain on the candidate features is quantified to screen the shared features with cross-region consistency, and the cross-region shared features are obtained.

[0046] The dual-domain data distribution alignment test in the embodiment includes: using the non-parametric Kolmogorov-Smirnov (KS) test to calculate the maximum difference value (KS statistic) of the cumulative distribution function (CDF) of the feature values of the two domains. For example, when the slope feature is tested, if the KS statistic of the source domain and the target domain is less than the significance level threshold (such as the critical value 0.15 corresponding to α=0.05), the hypothesis that the slope distribution of the two domains is consistent is accepted; otherwise, if the KS value is too high (such as 0.3), it indicates that the feature has significant regional specificity (such as the target domain due to tectonic uplift, the slope is generally larger), and needs to be excluded.

[0047] Then, the mutual information method is used to calculate the global correlation between each candidate feature and the probability of occurrence of geological disasters (such as the mutual information of the lithology combination is 0.32, which is significantly higher than the 0.08 of the vegetation coverage), and the features with significant cross-region correlation are retained.

[0048] Finally, the contribution degree of the shared features to the explanation of the geological disasters is verified by combining the principal component analysis load matrix (such as the slope and lithology load in the first principal component are 0.76 and 0.68, respectively), to ensure that more than 80% of the original information quantity can be retained after dimension reduction, so as to construct the cross-region shared feature set with domain invariance and disaster sensitivity.

[0049] In this embodiment, the principal component analysis load matrix is used to quantify the contribution weight of the original features to the principal components, and to reveal the linear correlation strength between each principal component in the reduced feature set and the original features. Specifically, the elements in the load matrix represent the correlation coefficient of the original features (such as slope, lithological combination) and the principal components (such as the first principal component PC1), and the larger the absolute value (such as the slope in PC1 has a load of 0.76) indicates that the feature has a more significant impact on the construction of the principal component. By analyzing the load matrix (for example, the slope and lithological combination loads in PC1 are 0.76 and 0.68 respectively, while the vegetation coverage is only 0.12), core features (slope, lithology) with strong explanatory power for geological disasters can be screened out, and it can be verified whether the principal components after dimensionality reduction effectively inherit their physical meaning (such as PC1 reflecting the geological stability), and combined with the principal component variance contribution rate (such as the cumulative contribution rate of the first three principal components is 85%), it can be ensured that the optimized feature set retains more than 80% of the information of the original shared features, thereby supporting the disaster sensitivity modeling of the cross-regional model.

[0050] In step 203, the principal component analysis method is used to reduce dimensionality, denoise and optimize the cross-regional shared features to obtain an optimized feature set.

[0051] The optimized feature set includes a source domain optimized feature set and a target domain optimized feature set.

[0052] Specifically, first, the data set of the cross-regional shared features (such as slope, lithological combination, historical disaster density) is standardized. The Z-score method can be used to eliminate the dimensional differences of the feature values.

[0053] Then, the covariance matrix of the standardized data set is calculated, the eigenvalues and eigenvectors of the covariance matrix are solved, the principal component direction is determined according to the order of the eigenvalues from large to small, for example, the first principal component variance contribution rate is 62%; by analyzing the correlation coefficient of the original features and the principal components in the principal component load matrix, features with significant contribution to the construction of the principal components are selected, for example, the load of the slope in the first principal component is 0.76, the load of the lithological combination is 0.68, and the load of the vegetation coverage is only 0.12; according to the cumulative variance contribution rate threshold (such as 85%), the first k principal components are intercepted, and the noise features with a load absolute value less than 0.5 are removed.

[0054] Finally, the original high-dimensional features are projected into a low-dimensional principal component space to generate a source domain optimized feature set and a target domain optimized feature set, for example, a three-dimensional principal component vector. While retaining more than 80% of the original information, it eliminates the multicollinearity between features (such as the correlation coefficient between slope and terrain relief is reduced from 0.82 to 0.05) and sensor acquisition noise, providing low-dimensional feature expression with distribution alignment for subsequent cross-domain adversarial training.

[0055] Through the above processing mode, the source domain optimization feature set reduces data redundancy and noise interference on the basis of retaining the main information of the source domain data, improves the purity and representativeness of the data; the data of the target domain can better reflect its geological characteristics at a lower dimension, and has certain comparability and consistency with the source domain optimization feature set in the feature space, providing high-quality feature input for subsequent deep learning model training and cross-regional data adaptation, which helps to improve the performance and generalization ability of the model, and further enhances the accuracy and reliability of the geological disaster risk assessment.

[0056] Step 204, in the case of constructing a domain adversarial neural network containing a gradient reversal layer as a preset deep learning model, pre-training the deep learning model based on the source domain optimization feature set to obtain a source domain pre-training model.

[0057] The gradient reversal layer is a functional layer used for adversarial training in a deep learning model, and its core function is to reverse the direction of the gradient in the back propagation process. In the forward propagation, the gradient reversal layer does not change the input data, and directly passes it to the next layer of network; while in the back propagation, the gradient reversal layer will take the opposite sign of the gradient, that is, reverse the gradient direction. This mechanism makes the model not only consider reducing the loss of the current task in the optimization process, but also consider increasing the loss of another task, so as to realize the simultaneous optimization of two mutually antagonistic tasks. In the domain adaptation scene, the gradient reversal layer reverses the gradient, so that the feature extractor learns the features of the source domain data while making the feature distribution of the target domain data as close to the source domain as possible, thereby reducing the feature distribution difference between the two domains and improving the generalization ability of the model in the target domain.

[0058] In this step, the source domain optimization feature set is taken as input and fed into the constructed deep learning model for pre-training. In the pre-training process, the feature extractor learns efficient feature representation from the source domain optimization feature set, and the classifier predicts the geological disaster risk according to these features. A suitable loss function (such as mean square error loss or cross-entropy loss) is defined to measure the difference between the model prediction result and the actual label, and the back propagation algorithm is used to optimize the model parameters, and the training is iterated until the model reaches satisfactory performance on the source domain data, thereby obtaining the source domain pre-training model, laying a foundation for subsequent adversarial training and cross-regional adaptation.

[0059] In this embodiment, the domain adversarial neural network containing the gradient reversal layer is constructed as the preset deep learning model. First, the overall architecture of the model needs to be defined, which is usually composed of three main parts: a feature extractor, a classifier, and a domain discriminator. The feature extractor is responsible for extracting high-level feature representations from the input data, which can be implemented using a convolutional neural network (CNN) or a fully connected layer structure. The classifier performs specific classification tasks on the input data based on the features output by the feature extractor, such as different category prediction in geological disaster risk assessment. The task of the domain discriminator is to distinguish whether the input features come from the source domain or the target domain, which can also use network structures such as multi-layer perceptron (MLP).

[0060] Embedding the gradient reversal layer (GRL) in the model is the key to realizing domain adversarial training. GRL is located between the feature extractor and the domain discriminator. During the forward propagation process, GRL does not modify the input features and directly passes them to the domain discriminator. During the backward propagation process, GRL inverts the gradient sign from the domain discriminator, i.e., reverses the gradient direction. This mechanism makes the feature extractor consider not only reducing the classification task loss but also making the features of the source domain and the target domain difficult to distinguish on the domain discriminator during optimization, thereby forcing the feature extractor to learn feature representations with domain invariance.

[0061] The training process of the model is divided into two stages: pre-training and adversarial training. In the pre-training stage, the entire model is trained using the optimized feature set of the source domain, mainly optimizing the performance of the classifier, so that the model can accurately perform geological disaster risk assessment on the source domain data. In the adversarial training stage, the optimized feature sets of the source domain and the target domain are input simultaneously, and the gradient reversal layer is used to make the features learned by the feature extractor applicable to both the source domain and the target domain, reducing the feature distribution difference between the two domains. By alternately optimizing the classification loss and the domain discrimination loss, a cross-region adaptive model that can achieve good performance on both the source domain and the target domain is finally obtained.

[0062] In step 205, during the adversarial training of the source domain pre-trained model, the gradient reversal layer reverses the gradient direction during backpropagation to reduce the feature distribution difference between the source domain and the target domain, obtaining a cross-region adaptive model.

[0063] Specifically, during the training process, the feature extractor receives the optimized feature sets of both the source domain and the target domain as input. The gradient inversion layer inverts the gradient sign during backpropagation, causing the feature extractor to consider not only reducing the loss of the source domain prediction task but also making the feature representations of the source domain and the target domain as similar as possible when updating parameters, thereby reducing the feature distribution difference between the two domains.

[0064] During the training process, the model receives data from both the source domain and the target domain. The feature representations learned by the feature extractor need to be applicable to both domains. The classifier is responsible for predicting the geological disaster risk of the source domain data and calculating the corresponding classification loss. The domain discriminator attempts to distinguish whether the input features come from the source domain or the target domain and calculates the domain discrimination loss. The gradient inversion layer is located between the feature extractor and the domain discriminator. During backpropagation, the gradient inversion layer inverts the gradient sign from the domain discriminator, causing the feature extractor to consider not only reducing the classification loss but also making the features of the source domain and the target domain indistinguishable on the domain discriminator, thereby forcing the feature extractor to learn feature representations with domain invariance and reducing the feature distribution difference between the source domain and the target domain.

[0065] Through continuous iterative optimization, the final cross-region adaptation model can achieve good performance on both the source domain and the target domain, effectively improving the generalization ability of the model on the target domain and the accuracy of geological disaster risk evaluation.

[0066] Step 206, the target domain optimized feature set is expanded by the cross-region adaptation model, and the target domain expansion data is obtained.

[0067] Specifically, first, the optimized feature sets of the source domain and the target domain are simultaneously input into the pre-trained cross-region adaptation model. The cross-region adaptation model has learned the mapping relationship between the features of the source domain and the target domain during the previous adversarial training phase, effectively reducing the feature distribution difference between the two domains.

[0068] During the data expansion phase, the cross-region adaptation model maps the optimized feature set of the target domain to a space closer to the feature distribution of the source domain through the forward propagation mechanism, thereby generating expanded target domain data. Specifically, the feature extractor in the model will transform the features of the target domain to align with the source domain data in the feature space, and then generate new data samples through the decoder or generator. These newly generated samples not only retain the original feature information of the target domain but also integrate the distribution characteristics of the source domain data, effectively expanding the data volume of the target domain while improving the quality and representativeness of the data.

[0069] To ensure the accuracy and reliability of the augmented data, a series of verification and evaluation methods are used, such as comparing the statistical properties of the original data and the augmented data, using part of the known labels to test the model performance, etc., to obtain high-quality target domain augmented data, and provide more sufficient data support for subsequent geological disaster risk assessment.

[0070] In some embodiments, in order to make the target domain augmented data inherit the geological law knowledge of the source domain and conform to the local statistical properties of the target domain, both physical rationality and distribution authenticity are achieved. The following designs are also made in this embodiment to break through technical bottlenecks and overcome technical obstacles:

[0071] 1. Adversarial training architecture of cross-area joint distribution modeling:

[0072] In order to align the distribution offset problem between the source domain and the target domain, this embodiment constructs a domain adversarial network with an embedded gradient reversal layer (GRL). In the training process, the feature extractor and the generator are optimized at the same time: the feature extractor is responsible for extracting the common features of the source domain and the target domain, and the generator generates new samples to enhance the data of the target domain. Through GRL, the gradient direction is reversed during back propagation, forcing the generated samples to approximate both the disaster law of the source domain (such as the nonlinear relationship between slope and disaster occurrence) and the local statistical properties of the target domain (such as the frequency distribution of lithology combination). This embodiment also designs a dynamic weight adjustment strategy. In the early stage of training (such as the first 30% of training rounds), more emphasis is placed on the transfer of source domain knowledge, with the classification loss weight α set to 0.8 and the generator weight β set to 0.2. As the training progresses, the fitting of the target domain distribution is gradually enhanced, with α reduced to 0.3 and β increased to 0.7, so that more attention is paid to the characteristics of the target domain in the later stage of training, and finally the generated samples can inherit the geological law knowledge of the source domain and adapt to the local statistical properties of the target domain.

[0073] 2. Geological prior constraint generation mechanism:

[0074] In view of the technical obstacle that pure data-driven generation models may output samples that violate geological laws, a geological prior constraint generation mechanism is introduced. A geological constraint layer is added to the generator of GAN, which can map the slope value to the reasonable interval of [0°, 45°] after using the Sigmoid activation function in the output layer, avoiding the generation of unrealistic slope values; for lithology combination, the Softmax layer is used to force mutual exclusion coding, ensuring that combinations such as "granite + salt rock" that cannot coexist in geology are not generated.

[0075] In addition, the embodiment also introduces an explainability feedback mechanism, which quantifies the feature contribution of the generated samples using SHAP values. If the contribution of the slope feature is less than 30% (which indicates that the impact of slope on disaster occurrence is underestimated, deviating from geological cognition), the regeneration mechanism is triggered to ensure that the generated samples meet the basic principles of geology. SHAP value is an explainability index based on the Shapley fair allocation principle in game theory, which quantifies the contribution of each feature to the prediction result of a single sample (positive / negative impact) to explain the decision basis of the model output in an additive manner.

[0076] The generation mechanism of geological prior constraints in this embodiment ensures that the sample data output by the model not only conforms to geological laws but also has statistical authenticity.

[0077] 3. Adaptive enhancement of local statistical properties of target domain

[0078] To address the problem of overfitting or deviation from the true distribution of the generated distribution caused by the scarcity of target domain data, this embodiment uses a hybrid density estimation generation method. Specifically, it combines a parametric Gaussian mixture model and a non-parametric kernel density estimation method to model the latent space distribution of the target domain. When generating samples, sampling is performed according to the direction of the density gradient, preferentially filling areas with low density in the target domain, such as generating more "slope 28° + sand-shale interbedding" samples, which are originally lacking in the target domain. At the same time, this embodiment also implements a pseudo-label confidence screening strategy, only retaining high-confidence generated samples with a cross-region adaptation model prediction probability p(y) greater than 0.8, avoiding the pollution of low-quality data to the training set, and ensuring that the generated samples can enhance the local statistical properties of the target domain without introducing excessive noise or false information.

[0079] 4. Closed-loop optimization of generation-evaluation-iteration

[0080] To address the technical obstacle that static generated data cannot adapt to the distribution drift in the model training process, a closed-loop optimization mechanism of generation-evaluation-iteration is designed. After adding generated data to the target domain training set, the cross-region adaptation model is fine-tuned, and then the data is regenerated based on the updated cross-region adaptation model, and the iteration is optimized until the distribution matching degree (measured by maximum mean difference MMD, less than 0.05) and classification performance (evaluated by AUC, less than 1%) of the generated samples converge.

[0081] Meanwhile, an online physical verification module is introduced to monitor the physical parameters (e.g., slope value, lithology code) of the generated data in real time. If the violation rate of multiple batches of generated samples is greater than a set threshold (e.g., 5%), the online physical verification module automatically adjusts the generator parameters or feature weights, thereby ensuring that the generated data maintains high quality and high credibility throughout the training process and meets the strict requirements of geological disaster evaluation.

[0082] In designing the online physical verification module, the key physical parameters and constraint conditions of the geological disaster data are first determined, such as the reasonable range of slope and the logical rules of lithology combination. Then, a real-time monitoring mechanism is constructed based on these key physical parameters and constraint conditions to check the physical parameters of each generated sample data one by one during the generation of data. For example, check whether the slope value is within the interval [0°, 45°] and whether the lithology combination conforms to geological common sense. If it is found that the generated data has physical unreasonable situations, such as slope out of range or contradictory lithology combination, the online physical verification module will record the violation samples and calculate the violation rate. When the violation rate of 10 consecutive batches of generated samples exceeds 5%, the module automatically triggers an alarm and adjusts the generator's parameters or feature weights to guide the generator to regenerate data that conforms to physical laws.

[0083] This process ensures that the generated augmented data not only matches the target domain statistically but also has physical reasonableness and conforms to geological laws, thereby improving the stability of model training, the accuracy of prediction, and the reasonableness and accuracy of model output.

[0084] Step 207, perform geological disaster risk assessment based on the target domain augmented data to obtain the target domain geological disaster risk assessment result.

[0085] Specifically, when performing geological disaster risk assessment based on the target domain augmented data, first, the target domain augmented data is used as the basis to divide the target domain into multiple evaluation units, ensuring that each unit has certain homogeneity in geological features, providing a spatial framework for subsequent refined evaluation.

[0086] Then, based on the augmented terrain data of each evaluation unit, the influence of slope, aspect, and other topographic factors on geological disaster susceptibility is analyzed, combined with human activity indicators such as engineering construction and vegetation destruction to assess the contribution of human activities to the possibility of geological disaster occurrence. Meanwhile, meteorological data, especially rainfall data, are used to determine the instability probability under different rainfall conditions. By integrating these multi-source data, the pre-set geological disaster susceptibility comprehensive index formula is used to calculate the susceptibility comprehensive index of each evaluation unit, which comprehensively reflects the combined effect of topography, lithology, human activities, and other factors on the possibility of geological disaster occurrence.

[0087] At the same time, the influence of water pressure gradient on slope stability under different rainfall conditions is simulated by combining the rainfall-seepage-stress coupling model, and the spatial and temporal distribution of instability probability and the critical safety factor matrix are output, further refining the instability risk of each evaluation unit under different conditions.

[0088] Finally, according to the susceptibility comprehensive index and the instability probability, the geological disaster risk index is calculated, and the detailed geological disaster risk evaluation results are generated based on the index, which clearly identifies the geological disaster risk levels in different regions of the target domain, providing scientific and accurate decision support for the prevention, emergency response and regional planning of geological disasters.

[0089] Among them, the susceptibility comprehensive index is a quantitative index that represents the possibility of a region occurring geological disasters by comprehensively considering multiple geological environmental factors (such as slope, lithology, human activity, etc.) and assigning them appropriate weights. Instability probability is the probability of a slope unit losing stability and leading to geological disasters under certain conditions (such as light rain, heavy rain, heavy rain, etc.).

[0090] When calculating the geological disaster risk index, the susceptibility comprehensive index needs to be normalized, that is, the susceptibility comprehensive index of each evaluation unit is divided by the maximum susceptibility comprehensive index in all evaluation units to obtain the normalized susceptibility index. The purpose of this is to eliminate dimensional differences and make the susceptibility indices of different evaluation units comparable. Then, the normalized susceptibility index is multiplied by the instability probability to obtain the geological disaster risk index. This process considers the possibility of geological disasters and the intensity or impact degree of actual occurrence through weighting, and the final risk index can more comprehensively and accurately reflect the risk degree of each evaluation unit, providing a scientific basis for the prevention and emergency decision-making of geological disasters.

[0091] This embodiment divides the evaluated region into source and target domains, determines the cross-regional shared features based on the geological environment data of the data-intensive source domain and the data-sparse target domain, and optimizes the dimensionality reduction and denoising to obtain a more refined feature set. Further, the domain-adversarial neural network containing the gradient reversal layer is used to pre-train on the source domain data, and then the feature distribution difference between the source domain and the target domain is reduced through adversarial training, and finally a cross-regional adaptation model is obtained. The cross-regional adaptation model can perform data augmentation processing on the optimized feature set of the target domain, generating more abundant data, and thus realizing more accurate geological disaster risk evaluation of the target domain. The whole process fully taps the data advantages of the source domain, while adapting to the data characteristics of the target domain, effectively making up for the data shortage of the target domain, improving the accuracy of geological disaster risk evaluation, and alleviating the problem of insufficient evaluation accuracy due to insufficient data in data-scarce regions.

[0092] In some embodiments, with respect to step 205, in the process of adversarial training of the source domain pre-trained model, the following steps are included:

[0093] S51, generating a geological disaster spatial correlation matrix of the target domain according to the spatial distribution characteristics of geological disasters in the target domain.

[0094] The geological disaster spatial correlation matrix is a tool for describing and quantifying the degree of spatial correlation of geological disasters. It analyzes the spatial position, distance, direction, and other relationships between geological disaster points to construct a matrix representing the correlation strength or similarity between these disaster points. Each element in the matrix represents the correlation degree between two geological disaster points, usually calculated based on their spatial distance, similarity of geological characteristics, or other related factors. The geological disaster spatial correlation matrix can help researchers identify concentrated areas of geological disasters, analyze spatial distribution patterns of disasters, assess potential risks of disasters, and provide scientific basis for prediction and prevention of geological disasters.

[0095] In this step, first, collect the spatial coordinate data of all known geological disaster points in the target domain, which usually comes from historical geological disaster investigation records, remote sensing interpretation results, and field monitoring data, etc. Then, perform spatial analysis on these disaster points, calculate the spatial distance and direction relationship between each pair of disaster points, and construct a disaster point spatial relationship dataset. On this basis, use spatial statistical methods such as spatial autocorrelation analysis, kernel density estimation, etc. to quantify the spatial correlation pattern between disaster points. For example, calculate the Moran's I to measure the degree of aggregation or dispersion of disaster points in space, or use kernel density estimation to generate a spatial density distribution map of geological disasters, thereby identifying high-risk and low-risk areas of disasters.

[0096] According to these analysis results, a matrix is constructed, where the rows and columns of the matrix represent the respective geological disaster points in the target domain, and the element values in the matrix represent the spatial correlation strength between the corresponding disaster points, which is the geological disaster spatial correlation matrix of the target domain. This matrix can intuitively reflect the spatial distribution characteristics and mutual relationship of geological disasters in the target domain, providing a basis for the construction of the spatial continuity constraint function in the subsequent steps.

[0097] S52, constructing a spatial continuity constraint function based on the geological disaster spatial correlation matrix.

[0098] The spatial continuity constraint function is used to quantify the spatial correlation strength of geological disasters. The spatial continuity constraint function can quantify the spatial correlation strength of geological disasters to ensure the spatial continuity and correlation of geological disasters in the subsequent model training.

[0099] Specifically, first, the spatial correlation matrix of geological disasters is analyzed in depth, and key information such as the average correlation strength between disaster points, the distribution pattern of correlation strength, and the maximum and minimum correlation strength is extracted.

[0100] Then, according to these statistical characteristics, a function form is designed that can quantify the spatial correlation strength of geological disasters. Common function forms include Gaussian function, exponential function or polynomial function, etc. These functions can convert the spatial distance between disaster points and the correlation strength into a continuous numerical value for measuring spatial continuity. For example, a Gaussian function can be used to take the distance between disaster points as an input variable, and by adjusting the bandwidth parameter of the function, the decay rate of spatial correlation can be controlled, so that disaster points have higher correlation strength at closer distances, and the correlation strength gradually decreases as the distance increases.

[0101] After determining the function form, the function parameters are estimated and optimized using actual geological disaster data in the target domain to ensure that the spatial continuity constraint function accurately reflects the spatial distribution and continuity characteristics of geological disasters in the target domain.

[0102] Finally, the rationality and effectiveness of the constructed spatial continuity constraint function are verified and evaluated through methods such as cross-validation, model goodness-of-fit test, etc. to ensure that it can provide reliable quantitative basis for subsequent spatial constraint feature generation.

[0103] S53, according to the spatial continuity constraint function, generate spatial constraint features that meet the domain distribution consistency and geological spatial continuity.

[0104] Specifically, first, the feature data of the source domain and the target domain are taken as input, and combined with the spatial continuity constraint function, a feature generation model is constructed. The goal of the feature generation model is to generate new features that not only maintain consistency with the source domain feature distribution, but also reflect the spatial continuity of geological disasters in the target domain.

[0105] The following steps can be used in this embodiment: first, the feature data of the source domain and the target domain are jointly analyzed to calculate their differences in statistical distribution, such as mean, variance, etc.; then, the spatial continuity constraint function is used to quantify the spatial correlation strength of geological disasters in the target domain, and the spatial relationship between disaster points is converted into a numerical weight; then, through feature interpolation or feature enhancement technology, the weight of the spatial continuity constraint function is used to adjust the feature data of the target domain, so that the generated new features have better spatial continuity, and at the same time, through domain distribution alignment technology such as adversarial training or distribution matching loss function, ensure that the generated features are consistent with the source domain features in statistical distribution; finally, through model evaluation and verification, such as calculating the prediction performance of the generated features in the target domain or the similarity of the feature distribution, the feature generation process is continuously optimized until the generated spatial constraint features meet the requirements of domain distribution consistency and geological spatial continuity.

[0106] S54, based on the spatial constraint features, evaluate the model generalization performance of the target domain after feature alignment, and obtain performance evaluation indicators.

[0107] First, the generated spatial constraint features are input into the deep learning model that has been pre-trained and preliminarily adversarially trained, and forward propagation calculation is performed to obtain the prediction output of the model on the target domain data.

[0108] Then, these prediction outputs are compared with the limited known geological disaster label data in the target domain, and the corresponding loss function values such as mean square error loss or cross-entropy loss are calculated to measure the difference between the model prediction results and the true labels.

[0109] Some unsupervised evaluation indicators such as clustering stability indicators and feature distribution similarity measures can also be used to evaluate the generalization performance of the model on the target domain. For example, by calculating the distribution similarity of the source domain and the target domain features in the model hidden layer space, such as using Kullback-Leibler divergence or Wasserstein distance indicators, to determine whether the model successfully transfers the source domain knowledge to the target domain.

[0110] Finally, these evaluation indicators are comprehensively analyzed to obtain a performance evaluation indicator system that comprehensively reflects the generalization performance of the model, which provides a basis for subsequent weight adjustment and model optimization.

[0111] If the evaluation results show that the generalization performance of the model on the target domain still needs to be improved, the model parameters need to be further adjusted or the feature generation process needs to be optimized to ensure that the model can accurately evaluate the geological disaster risk of the target domain.

[0112] S55, according to the performance evaluation indicators, determine the initial weights of the source domain and the target domain.

[0113] In determining the initial weights of the source domain and the target domain, first, a deep analysis is conducted on multiple key indicators in the performance evaluation indicators, such as prediction accuracy, recall rate, F1 score, and feature distribution similarity measure, which comprehensively reflect the generalization ability and adaptability of the model in the target domain.

[0114] Then, according to the specific values of the performance evaluation indicators, a weighted voting method or a Bayesian optimization algorithm is used to calculate the contribution of the source domain and the target domain data in model training. The data with high contribution will obtain relatively high initial weights. In addition, considering the reliability and representativeness of the source domain data, the integrity and quality score of the source domain data are calculated to further adjust the weight distribution, ensuring that the initial weight of the source domain is neither too high nor too low to fully play its guiding role.

[0115] Finally, the calculated weights are normalized to ensure that the sum of the weights is 1, thereby obtaining the initial weights of the source domain and the target domain, providing a reasonable starting point for subsequent dynamic weight adjustment, and improving the applicability and accuracy of the model in the target domain.

[0116] S56, according to the real-time data increment of the target domain and the initial weight, dynamically adjusting the weight distribution of the source domain and the target domain.

[0117] In dynamically adjusting the weight distribution of the source domain and the target domain, first, the real-time data increment of the target domain needs to be analyzed in detail to determine its new data size and time series information. The new data size can be quantified by calculating the proportion of the number of newly obtained data points to the total amount of existing data in the target domain, and the time series information involves the distribution of these data at different time points.

[0118] According to this information, a dynamic adjustment coefficient is calculated using a preset coefficient adjustment formula. This formula considers multiple parameters such as the initial adjustment coefficient, the source domain weight decay rate, the target domain data increment influence coefficient, as well as the new data amount and the expected total data amount of the target domain at different time points.

[0119] Through the synergistic effect of these parameters, the dynamic adjustment coefficient can accurately reflect the influence of the target domain data increment on weight adjustment. Combined with the initial weight and the dynamic adjustment coefficient, the weight distribution ratio of the source domain and the target domain in model training is updated.

[0120] As the target domain data continues to accumulate, the weight proportion of the source domain gradually decreases, and the weight proportion of the target domain correspondingly increases. This process dynamically adjusts the weights so that the model can better adapt to the actual situation of the target domain, improving the accuracy and timeliness of the evaluation results.

[0121] In some embodiments, step S56 can further include the following steps:

[0122] S561, determine the size of the newly added data of the target domain and time series information according to the real-time data increment of the target domain.

[0123] First, deploy a real-time data stream monitoring module in the target domain, which continuously receives incremental data generated from sensors, business systems or user interaction behaviors. The calculation of data increment size is based on a sliding time window mechanism, and the window length is set to be hourly or daily according to business needs. By counting the number of new data records, the change amount of feature dimensions (such as image resolution improvement or text field expansion) and the growth trend of storage capacity occupation in each window, the size of the target domain data increment is dynamically quantified.

[0124] The extraction of time series information needs to combine the generation frequency and periodicity of incremental data. The Autoregressive Integrated Moving Average Model (ARIMA) is used to analyze the stationarity of time series, and Fourier transform is used to identify the seasonal and trend components of data increment.

[0125] For unstructured data, the semantic feature evolution rate associated with the timestamp needs to be calculated additionally, such as the time decay coefficient of word distribution in natural language processing tasks. When the target domain data increment breaks through the preset threshold (such as the daily incremental data exceeding 2 standard deviations of the historical average), the data size re-estimation process is triggered, and the target domain training set storage space is dynamically expanded through elastic computing resources, while generating a data increment report containing timestamp, increment size, and feature type distribution, providing quantitative basis for subsequent transfer learning parameter adjustment.

[0126] S562, based on the size of the newly added data and the time series information, determine the dynamic adjustment coefficient of the source domain and the target domain in transfer learning.

[0127] The determination of the dynamic adjustment coefficient of the source domain and the target domain in transfer learning requires the construction of a dual-domain feature space similarity evaluation system. After obtaining the size of the newly added data of the target domain and the time series information, first calculate the marginal distribution difference between the source domain and the target domain, and use the Maximum Mean Discrepancy (MMD) to measure the distribution distance of the feature space of the two domains, while introducing the Dynamic Time Warping (DTW) algorithm to align the time series fluctuation patterns of the two domains.

[0128] Based on the time continuity strength of the newly added data of the target domain (such as the incremental data exceeding 150% of the baseline value for three consecutive days), set a time decay factor to weight the historical source domain data, and configure a higher decay rate for sudden data growth.

[0129] The calculation of the dynamic adjustment coefficient adopts a double-layer optimization structure: the outer layer optimization determines the basic adjustment coefficient by the ratio of the target domain incremental scale to the original data volume of the source domain, and the inner layer optimization corrects the coefficient according to the time series correlation coefficient (the proportion of the time window with a Pearson correlation coefficient greater than 0.7).

[0130] When the target domain continuously and stably generates high-quality data, the migration focus shift mechanism is started, the weight coefficient of the domain adaptation item in the transfer learning loss function is dynamically adjusted by calculating the KL divergence (Kullback-Leibler Divergence) between the class confidence distribution of the newly added samples of the target domain and the prior knowledge of the source domain.

[0131] The finally output dynamic adjustment coefficient matrix contains parameters of three dimensions of feature level, sample level and task level, ensuring fine control of knowledge transfer in the transfer process.

[0132] In some embodiments, step S562 can specifically include:

[0133] S5621, determining the target domain new data volume at multiple moments according to the new data scale and time series information.

[0134] Specifically, after deploying a real-time data collection system in the target domain, first, the time stamps of all data sources are calibrated to the millisecond level based on a high-precision clock synchronization protocol (such as NTP / PTP), ensuring the consistency of the time base of cross-device data; for structured data streams (such as database transaction logs), a sliding time window statistical mechanism is adopted to divide the continuous time axis into fixed-interval statistical units (for example, every 5 minutes as a window), and in each window, the number of newly added records, the number of field extensions and the storage capacity increment are accumulated in real time to generate structured incremental statistical results with the window start time as the key value; for unstructured data streams (such as video streams or image collection systems), the creation time stamp in the file metadata is parsed, and a preset equal-interval time slice division rule (such as every 30 minutes as a statistical period) is combined to count the number of newly added files, the total number of pixels or the number of text characters in each period.

[0135] In the time series processing stage, the dynamic time warping algorithm is used to align the time stamp offset of multi-source heterogeneous data, and the exponential weighted moving average method is used to smooth the bursty data stream, eliminating the data jitter at the boundary of the statistical window.

[0136] The final output form is a discrete moment sequence {τ1,τ2,...,τn} arranged in chronological order, each moment τi corresponds to a calibrated and verified target domain new data volume ΔD(τi), forming a discrete time function of the target domain incremental scale, providing accurate time sequence input parameters for subsequent dynamic adjustment coefficient calculation.

[0137] S5622, based on the amount of newly added data in the target domain at multiple time points and the preset coefficient adjustment formula, calculates the dynamic adjustment coefficient. The coefficient adjustment formula includes:

[0138] ;

[0139] in, λ ( t () represents the dynamic adjustment coefficient. λ 0 represents the initial adjustment factor. β Represents the decay rate of the source domain weights. γ The incremental impact coefficient of the target domain data, ΔD( τ ) represents a moment τ The target domain has a new amount of data, D total This represents the expected total data volume of the target domain. k The time decay factor representing incremental data. t Indicates the current time. τ Used to iterate through all time points from the initial time to the current time t. e It is a natural constant.

[0140] β The determination of the β value is based on the effectiveness decay law of source domain knowledge: the performance decay curve of the source domain pre-trained model is monitored by the validation set during the incremental data process of the target domain (such as the rate of decrease in test accuracy per cycle), and the decay rate is calculated by the piecewise exponential fitting method; or the initial value is set according to the task prior (such as β=0.05 corresponding to a half-life of about 14 training cycles), and then fine-tuned in the range of 0.01-0.1 by grid search, and the optimal β value that makes the transfer model converge and stabilize on the target domain validation set is selected.

[0141] γ The calibration is achieved through correlation analysis between the target domain data increment and the model performance gain: in the offline phase, historical incremental data is used for simulation to calculate different... γ The Pearson correlation coefficient between the rate of decrease in loss in the target domain and the incremental size is used to select the value that maximizes the correlation. γ Value (e.g.) γ =0.6); Bayesian optimization is introduced in the online phase, with the F1 score of the target domain validation set as the optimization objective, and a dynamic search is performed in the range of 0.2-1.0. γ Optimal value, while constrained γ and κ An upper limit for the product is set to avoid weight oscillations.

[0142] k The setting relies on the timeliness analysis of the target domain data: through time series cross-validation, the contribution decay curves of samples at different time intervals in historical data to the current model prediction are calculated, and the k value is determined by exponential fitting (e.g., for daily data decay).k =0.1); or preset the time decay strength based on business rules (e.g., requiring the influence of data from one week ago to decrease to 10%). k = ln (0.1) / 7≈-0.33), and then k is adjusted by verifying through a sliding window to ensure that the weight distribution of the integral term meets the timeliness requirements of the domain. The dynamic adjustment coefficient formula in this embodiment has a dual mechanism design that takes into account both the knowledge decay law of the source domain and the incremental data influence of the target domain:

[0143] The first term of the formula λ 0· e -βt The gradual forgetting process of source domain knowledge is simulated using an exponential decay function, with the decay rate... β Based on the experience of domain experts or the performance degradation curve of the validation set, ensure that necessary prior knowledge of the source domain is retained in the early stage of target domain data accumulation.

[0144] Formula 2 Using a convolutional integral structure, the amount of newly added data ΔD in the target domain at each historical time step is calculated. τ The relative contribution of ) (ΔD( τ ) / D total ) and time decay factor e -k(t-τ) This combination reflects both the driving effect of the scale of data increment on the direction of migration and the higher weight of recent data through exponential decay, which is in line with the cognitive pattern that "data timeliness affects model evolution".

[0145] coefficient γ and k The global strength and time decay rate of the influence of target domain data can be controlled separately through offline simulation or online Bayesian optimization, so that the formula can be both mathematically interpretable and engineering adjustable, and finally achieve a dynamic balance between source domain knowledge annealing and target domain knowledge accumulation.

[0146] S563 updates the weight distribution ratio between the source and target domains during model training based on dynamically adjusted coefficients and initial weights.

[0147] Initial weights are typically set based on the reliability and representativeness of the source domain data. As the target domain data increases and the dynamic adjustment coefficients change, the weight proportion of the source domain gradually decreases, while the weight proportion of the target domain increases accordingly. This is because, with the continuous accumulation of target domain data, the model needs to gradually adapt to the actual situation of the target domain. In the initial stage, due to the scarcity of target domain data, the rich data and knowledge of the source domain are crucial for model training. However, as the amount of target domain data increases, its guiding role in the model becomes increasingly important, enabling it to more accurately reflect the geological hazard characteristics and patterns of the target domain. To enable the model to better adapt to the target domain and improve its generalization ability and prediction accuracy in the target domain, it is necessary to dynamically adjust the weight allocation, reducing the weight proportion of the source domain and increasing the weight proportion of the target domain, allowing the model to be optimized more based on the target domain data.

[0148] Specifically, a weighted summation method can be used, multiplying the dynamic adjustment coefficient by the initial weights to obtain a new weight allocation ratio. Simultaneously, to ensure the rationality of the weight allocation, the new weight ratios are normalized so that their sum equals 1.

[0149] After each weight update, the model performance is evaluated using a validation set to ensure that the weight adjustments help improve the model's generalization ability and prediction accuracy in the target domain.

[0150] In some embodiments, step S563 may specifically include: calculating the weight allocation ratio according to a preset weight allocation formula, wherein the weight allocation formula includes:

[0151] ;

[0152] w g ( t )=1- w s ( t );

[0153] in, w s ( t ) represents the weight ratio of the source domain. w g ( t ) represents the weight ratio of the target domain. λ ( t () represents the dynamic adjustment coefficient. w i Indicates the initial weights. α Indicates the sensitivity of dynamic coefficients. η Indicates the weight transfer rate. t Indicates the current time. τ Used to iterate through all time points from the initial time to the current time t. eIt is a natural constant.

[0154] α The optimal α value can be determined by combining offline optimization with online adaptive optimization: First, based on a grid search on the validation set, the generalization performance (e.g., AUC, F1 score) of the model in the target domain under different α values ​​is evaluated within a preset range (0.1-5.0), and the optimal value that balances the decay rate of the source domain weights with the improvement of the target domain accuracy is selected; then, Bayesian optimization is used to construct the objective function (e.g., a weighted combination of validation loss and weight stability), and the model is iteratively converged to the optimal value through a Gaussian process surrogate model. α During online training, a dynamic adjustment strategy is designed. When the loss reduction rate of the target domain is insufficient, α is increased proportionally to accelerate the forgetting of the source domain. Conversely, if overfitting occurs, α is decreased to retain the regularization effect of the source domain, thereby achieving real-time adaptation of parameters and data evolution.

[0155] η can be determined through calibration dependency theory modeling and data-driven tuning: an initial value is preset based on the number of training epochs (e.g., ...). η (≈2 / total number of cycles), offline simulation is run using historical data, and a bisection method is used to search for the value that minimizes the error between the simulated weight curve and the theoretical decay. η Value; during the online phase, a sliding window is deployed to monitor the actual weight transfer rate and calculate... η The deviation from the measured value is adjusted by introducing a smoothing factor (e.g., 0.8) and asymptotic updates to the momentum term. η This ensures that weight transfer is synchronized with data increments; ultimately, the early stopping mechanism of the validation set is used to determine... η The optimal value balances migration speed and stability.

[0156] The weight allocation formula in this embodiment possesses an adaptive balancing strategy that integrates a gradual forgetting mechanism and a cumulative effect:

[0157] The formula is passed through the numerator. w i ·e -αλ(t) This achieves exponential decay control of the source domain weights, where the dynamic adjustment coefficient λ(t) is modulated by the sensitivity parameter α and applied to the initial weights. w i This makes it possible when the data increment in the target domain is significant ( λ ( t (Increase), source domain weights decay more rapidly to suppress interference from outdated knowledge.

[0158] in the denominator Partially constrained by the hyperbolic tangent function, the historical cumulative integral term will be used to dynamically adjust the coefficients throughout the transfer learning cycle. λ(τ) Perform time integration, combined with weight transfer rate ηSmooth the growth trend of the target domain weight, avoid the weight shock caused by the sudden change of single moment data, and limit the upper limit of the target domain weight through the nonlinear saturation characteristic (tanh value domain [-1, 1]) to prevent the overfitting of the target domain.

[0159] Finally, through the competitive weighting of the molecular denominator, the two-domain weight is normalized and allocated, forming a triple control mechanism of initial weight guidance, dynamic coefficient adjustment and historical trend strengthening, ensuring the stable transition and global optimal balance of the annealing of source domain knowledge and the accumulation of target domain knowledge in the process of transfer learning.

[0160] The embodiment effectively solves the problem of large distribution difference between source domain and target domain data and the scarcity of target domain data in geological disaster risk assessment. First, by constructing a geological disaster spatial correlation matrix and a spatial continuity constraint function, the spatial distribution law of geological disasters in the target domain is fully tapped and integrated into the model training, so that the generated spatial constraint features not only conform to the geological disaster law of the source domain, but also match the spatial characteristics of the target domain. Then, the performance evaluation index is used to determine the initial weight of the source domain and the target domain, and the weight distribution is dynamically adjusted according to the real-time data increment of the target domain, gradually reducing the dependence on the source domain. This process not only improves the adaptability of the model to the target domain, but also ensures that the model can be continuously optimized as the target domain data increases. Finally, by introducing the domain adversarial network with gradient reversal layer and the generation mechanism of geological prior constraints, the shortcomings of traditional generative models in cross-regional distribution alignment and physical rationality are overcome, and the generated expansion data not only inherits the geological knowledge of the source domain, but also conforms to the local statistical characteristics of the target domain. Overall, these steps work together to significantly improve the accuracy and reliability of the geological disaster risk assessment model in data-scarce areas, providing stronger technical support for the prevention and control of geological disasters.

[0161] In some embodiments, step 207 can specifically include the following steps:

[0162] S701, according to the target domain expansion data, the target domain is divided into multiple evaluation units.

[0163] Specifically, first, collect the target domain expansion data, which contains the geological environment information of the target domain, such as topography, lithology, weather and other multi-source data. Preprocess these data, including data cleaning, missing value filling and outlier processing, to ensure the integrity and accuracy of the data.

[0164] Then, according to the requirements of the geological disaster evaluation and the spatial distribution characteristics of the data, the target domain is divided into multiple evaluation units. In the division process, the geological features in each evaluation unit are ensured to be relatively consistent, so that the subsequent geological disaster risk evaluation can more accurately reflect the actual geological disaster risk of each unit.

[0165] In the division process, the geological features in each evaluation unit are ensured to be relatively consistent, so that the subsequent geological disaster risk evaluation can more accurately reflect the actual geological disaster risk of each unit.

[0166] S702, based on the expanded terrain data, human activity index and meteorological data of each evaluation unit, the susceptibility comprehensive index and the instability probability of each evaluation unit are determined.

[0167] Among them, the susceptibility comprehensive index is used to represent the possibility of geological disaster occurrence in the evaluation unit, and the instability probability is used to represent the probability of the evaluation unit losing stability and leading to geological disaster occurrence.

[0168] For each evaluation unit, its expanded terrain data, human activity index and meteorological data after the previous embodiment steps are collected. The terrain data includes slope, slope direction, elevation and other information, which can be extracted from the elevation model or topographic map through geographic information system (GIS) technology. The human activity index covers land use type, engineering construction activity, vegetation coverage and other aspects, which can come from field investigation, remote sensing interpretation or relevant statistical data. Meteorological data mainly includes rainfall, rainfall intensity, rainfall duration, etc., which can be obtained from meteorological station observation data or meteorological satellite data.

[0169] After collecting these data, the susceptibility comprehensive index of each evaluation unit is calculated by using the preset susceptibility comprehensive index calculation method combined with the terrain, human activity and meteorological data of each evaluation unit. For example, the data is standardized to eliminate dimensional differences. Then, according to the possibility of geological disaster occurrence, the weight of each factor is allocated, such as slope accounting for 30%, lithology accounting for 25%, human activity accounting for 20%, etc. The standardized values of each factor are multiplied by their corresponding weights and then added to obtain the susceptibility comprehensive index by using the weighted summation method. For example, if the standardized slope value of an evaluation unit is 0.8, the lithology value is 0.6, and the human activity value is 0.7, then the susceptibility comprehensive index is 0.8x0.3 + 0.6x0.25 + 0.7x0.2 + … (other factors). The higher the susceptibility comprehensive index, the greater the possibility of geological disaster occurrence in the evaluation unit.

[0170] Meanwhile, based on the rainfall conditions in meteorological data and geological characteristics, the stability of the slope unit under different rainfall conditions is analyzed to determine the instability probability of each evaluation unit. For example, first, according to historical rainfall data and geological disaster occurrence records, the instability threshold under different rainfall conditions is determined, such as the rainfall amount and duration corresponding to light rain, moderate rain, heavy rain, and rainstorm. For each evaluation unit, the stability under different rainfall conditions is analyzed. For example, under the rainstorm condition, the influence of rainfall infiltration on slope stability is considered, and the water pressure change is calculated using a hydrogeological model to determine the probability of slope instability. At the same time, combined with geological characteristics such as the shear strength of rock-soil mass and slope, the safety factor is calculated through stability analysis formulas (such as circular sliding method, limit equilibrium method). If the safety factor is less than 1, it is considered that the slope may be unstable, and the instability probability is determined according to the size of the safety factor and the statistical results of historical data, such as when the safety factor is 0.8, the instability probability may be 60%.

[0171] S703, determine the geological disaster risk index according to the susceptibility comprehensive index and the instability probability.

[0172] Specifically, first, the susceptibility comprehensive index is normalized to convert it to a unified dimension range, so as to effectively combine with the instability probability. The normalization method can use the maximum-minimum normalization to map the susceptibility comprehensive index value to the interval of 0 to 1.

[0173] Then, according to the preset geological disaster risk index calculation formula, the normalized susceptibility comprehensive index and the instability probability are weighted and summed or other mathematical operations to obtain the geological disaster risk index. The formula is usually designed according to the actual needs of geological disaster evaluation and expert experience to ensure that the influence of susceptibility and instability on geological disaster risk can be reasonably considered.

[0174] The geological disaster risk index calculated in this way can comprehensively and accurately reflect the geological disaster risk degree of each evaluation unit, providing a scientific basis for subsequent generation of geological disaster risk evaluation results.

[0175] In some embodiments, the geological disaster risk index can be calculated according to the preset risk index formula, and the risk index formula includes: H i =Y i / Y max ×P i ;

[0176] Wherein, H i represents the geological disaster risk index of the i-th evaluation unit under a certain rainfall condition, Y i represents the susceptibility comprehensive index of the i-th evaluation unit, Y maxP represents the maximum value of the susceptibility comprehensive index. i P represents the instability probability.

[0177] Specifically, in calculating the geological disaster risk index, it is first necessary to determine the susceptibility comprehensive index and the instability probability of each evaluation unit. The susceptibility comprehensive index is obtained by collecting the geological environmental data of the evaluation unit, including slope, lithology, land use type, vegetation coverage, human activity intensity, etc., standardizing these data to eliminate dimensional differences, then assigning weights to each factor according to the likelihood of geological disaster occurrence, such as slope accounting for 30%, lithology accounting for 25%, human activity accounting for 20%, etc., using the weighted summation method to multiply the standardized values of each factor by their corresponding weights, and then adding them up to obtain the susceptibility comprehensive index. The calculation of the instability probability is based on meteorological data and geological characteristics, and the instability threshold under different rainfall conditions is determined according to historical rainfall data and geological disaster occurrence records, such as the corresponding rainfall and duration of light rain, moderate rain, heavy rain, and heavy rain. For each evaluation unit, analyze its stability under different rainfall conditions, consider the influence of rainfall infiltration on slope stability, calculate the change of water pressure using hydrogeological models, and then determine the probability of slope instability. At the same time, combined with geological characteristics such as rock-soil shear strength and slope, the safety factor is calculated through the stability analysis formula. If the safety factor is less than 1, it is considered that the slope may be unstable, and the instability probability is determined according to the size of the safety factor and the statistical results of historical data.

[0178] After obtaining the susceptibility comprehensive index and the instability probability, the geological disaster risk index H i is calculated according to the preset risk index formula H i =Y max / Y i ×P i . Wherein H i represents the geological disaster risk index of the i-th evaluation unit under a certain rainfall condition, Y i / represents the susceptibility comprehensive index of the i-th evaluation unit, Y max represents the maximum value of the susceptibility comprehensive index among all evaluation units, used for normalization processing to ensure the comparability of the susceptibility index of different evaluation units, and P i represents the instability probability of the i-th evaluation unit under a specific rainfall condition.

[0179] By normalizing the susceptibility comprehensive index and multiplying it by the instability probability, the obtained geological disaster risk index can comprehensively reflect the risk degree of the evaluation unit under a specific rainfall condition, considering both the long-term influence of geological environmental factors on disaster occurrence and the short-term effect of rainfall and other triggering factors on disaster occurrence, providing a scientific basis for the early warning and prevention of geological disasters.

[0180] S704, generate a geological disaster risk assessment result based on the geological disaster risk index.

[0181] Specifically, first, the geological disaster risk index is classified, and the risk index is divided into different levels according to the preset classification standard, such as low risk, medium risk, high risk, and extremely high risk. The establishment of the classification standard is usually based on the comprehensive analysis of historical geological disaster data, geological environment characteristics and relevant specifications, so as to ensure that the classification result can scientifically and reasonably reflect the actual risk degree of geological disasters.

[0182] Then, the risk level results of each evaluation unit are integrated into a geographic information system (GIS) platform or other visualization tools to generate intuitive geological disaster risk assessment result maps. These maps clearly show the distribution of geological disaster risk levels in different regions in the form of color, symbols, etc., which facilitates decision-makers and relevant personnel to quickly understand the geological disaster risk situation in the target domain.

[0183] In addition, detailed reports can also be generated based on the evaluation results, including the risk level of each evaluation unit, analysis of main disaster-causing factors, prevention suggestions, etc., to provide comprehensive and scientific decision support for the prevention, emergency response and regional planning of geological disasters.

[0184] Based on the above series of steps, the evaluation region is first divided into a data-intensive source domain and a data-scarce target domain according to the density of geological environment data. Then the cross-regional shared features of the source domain and the target domain are determined, including slope, lithology combination and historical disaster density, etc., and the principal component analysis method is used to optimize the features to obtain the optimized feature set of the source domain and the target domain.

[0185] Then a domain adversarial neural network containing a gradient reversal layer is constructed, a source domain pre-training model is obtained by pre-training using the source domain optimized feature set, and the gradient direction is reversed through the gradient reversal layer in the adversarial training to reduce the feature distribution difference between the source domain and the target domain, and finally a cross-regional adaptation model is obtained. The model is used to expand the data of the target domain to obtain the target domain expansion data.

[0186] Then, according to the target domain expansion data, the target domain is divided into multiple evaluation units, the susceptibility comprehensive index and the instability probability are determined based on the topographic data, human activity indicators and meteorological data of each evaluation unit, and the geological disaster risk index is calculated according to the preset risk index formula, and finally the geological disaster risk assessment result is generated.

[0187] In the whole scheme, the weight distribution of the source domain and the target domain is dynamically adjusted according to the real-time data increment of the target domain, so as to gradually reduce the dependence on the source domain, improve the adaptability and generalization ability of the model to the target domain, and effectively solve the problem of insufficient precision of geological disaster risk evaluation in data scarce areas, and provide a more scientific and reliable decision basis for the prevention and control of geological disasters.

[0188] The method provided by the above embodiment can be executed by an electronic device. Next, the electronic device in the embodiment of the application is described from the perspective of hardware processing. Please refer to Figure 2 , which is a schematic diagram of an entity device structure of the electronic device in the embodiment of the application.

[0189] It should be noted that Figure 2 The structure of the electronic device shown is only an example, and should not limit the functions and use range of the embodiment of the application.

[0190] As Figure 2 shown, the electronic device includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 402 or programs loaded from a storage part 408 to a random access memory (RAM) 403, such as performing the method described in the above embodiment. In the random access memory (RAM) 403, various programs and data required for system operation are also stored. The central processing unit (CPU) 401, the read-only memory (ROM) 402, and the random access memory (RAM) 403 are connected to each other through a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0191] The following components are connected to the input / output (I / O) interface 405: an input section 406 including an audio input device, a button switch, and the like; an output section 407 including a display, an audio output device, an indicator, and the like; a storage section 408 including a hard disk and the like; and a communication section 409 including a network interface card such as a LAN (Local Area Network) card, a modem, and the like. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the input / output (I / O) interface 405 as necessary. A removable media 411 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is attached to the drive 410 as necessary, so that a computer program read out therefrom is installed in the storage section 408 as necessary.

[0192] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program according to embodiments of the present application. For example, an embodiment of the present application includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing a computer program for executing the method shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network by the communication section 409, and / or installed from the removable media 411. When the computer program is executed by the central processing unit (CPU) 401, various functions defined in the present application are executed.

[0193] Note that specific examples of the computer-readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, a fiber optic device, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present application, a computer-readable storage medium can be any tangible medium that contains or stores a program used by an instruction execution system, apparatus, or device to function or be combined therewith.

[0194] The computer program product of the present application can be a computer program embodied on a non-transitory computer readable medium. When the program is executed by a computer, it can achieve the objectives of the present application as discussed above. Accordingly, the scope of the present application should not be limited by the computer in which the computer program is executed. The computer program can be a stand-alone program, or it can be part of an operating system that controls an operation of a computer.

[0195] In particular, the electronic device of the embodiment includes a processor and a memory coupled with the one or more processors, the memory configured to store computer program code comprising computer instructions to be invoked by the one or more processors to cause the electronic device to perform the method provided by the above embodiment.

[0196] As another aspect, the present application also provides a computer readable storage medium, which can be included in the electronic device described in the above embodiments, or can exist separately and not be assembled into the electronic device. The storage medium carries one or more computer programs, which, when executed by a processor of the electronic device, cause the electronic device to implement the method provided in the above embodiments.

[0197] The above-described embodiments are merely intended for describing the technical solutions of the present application, but not to limit the present application; even though the present application has been described in detail with reference to the foregoing embodiments, those ordinarily skilled in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or equivalently replace some 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 application.

[0198] In the above embodiments, according to the context, the term "when" can be interpreted as meaning "if" or "after" or "in response to determining" or "in response to detecting". Similarly, according to the context, the phrase "upon determining" or "if detecting (the stated condition or event)" can be interpreted as meaning "if determining" or "in response to determining" or "upon detecting (the stated condition or event)" or "in response to detecting (the stated condition or event)".

[0199] Those skilled in the art can understand that all or part of the processes in the above-mentioned method embodiments can be implemented by a computer program instructing relevant hardware to complete, the program can be stored in a computer readable storage medium, and the program can include the processes of the above-mentioned method embodiments when executed. The aforementioned storage medium includes ROM or random storage memory RAM, magnetic disc or optical disc and various storage code medium.

Claims

1. A method for evaluating a geological disaster risk, characterized by, The method comprises the following steps: According to the density of the geological environment data of the evaluated region, the evaluated region is divided into a source domain and a target domain, wherein the data density of the source domain is greater than that of the target domain; According to the geological environment data of the source domain and the target domain, determine the cross-regional shared features, which include at least one of slope, lithology combination and historical disaster density; Optimize the cross-regional shared features by principal component analysis to obtain an optimized feature set, which includes a source domain optimized feature set and a target domain optimized feature set; In the case of constructing a domain adversarial neural network containing a gradient reversal layer as a preset deep learning model, pre-training the deep learning model based on the source domain optimized feature set to obtain a source domain pre-training model; In the process of adversarial training of the source domain pre-training model, the gradient direction is reversed by the gradient reversal layer during back propagation to reduce the feature distribution difference between the source domain and the target domain, and a cross-regional adaptation model is obtained; The target domain expansion data is obtained by expanding the data of the target domain optimized feature set through the cross-regional adaptation model; According to the target domain expansion data, the geological disaster risk assessment of the target domain is carried out to obtain the geological disaster risk assessment result of the target domain; In the process of adversarial training of the source domain pre-training model, it comprises: According to the spatial distribution characteristics of the geological disasters of the target domain, a geological disaster spatial correlation matrix of the target domain is generated; Based on the geological disaster spatial correlation matrix, a spatial continuity constraint function is constructed, which is used to quantify the spatial correlation strength of geological disasters; According to the spatial continuity constraint function, a spatial constraint feature meeting the domain distribution consistency and geological spatial continuity is generated; Based on the spatial constraint feature, the model generalization performance of the target domain after feature alignment is evaluated to obtain a performance evaluation index; According to the performance evaluation index, the initial weight of the source domain and the target domain is determined; According to the real-time data increment of the target domain and the initial weight, the weight distribution of the source domain and the target domain is dynamically adjusted.

2. The method of claim 1, wherein, According to the real-time data increment of the target domain and the initial weight, the weight distribution of the source domain and the target domain is dynamically adjusted, which comprises: According to the real-time data increment of the target domain, the size of the new data of the target domain and the time series information are determined; Based on the size of the new data and the time series information, the dynamic adjustment coefficient of the source domain and the target domain in transfer learning is determined; According to the dynamic adjustment coefficient and the initial weight, the weight distribution ratio of the source domain and the target domain in model training is updated to gradually reduce the proportion of the source domain weight.

3. The method of claim 2, wherein, According to the size of the new data and the time series information, the dynamic adjustment coefficient of the source domain and the target domain in transfer learning is determined, which comprises: According to the size of the new data and the time series information, the target domain new data amount at multiple time points is determined; According to the target domain new data amount of the plurality of moments and a preset coefficient adjustment formula, the dynamic adjustment coefficient is calculated, and the coefficient adjustment formula comprises: ； wherein, λ (t) represents the dynamic adjustment coefficient, λ 0 represents the initial adjustment coefficient, β represents the source domain weight decay rate, γ represents the target domain data increment influence coefficient, ΔD( τ ) represents the target domain new data amount at time τ, D total represents the target domain expected total data amount, k represents the time decay factor of the incremental data, and t represents the current time, τ for traversing all time points from the initial time to the current time t, e is a natural constant.

4. The method according to claim 2 or 3, characterized in that, According to the dynamic adjustment coefficient and the initial weight, the weight distribution ratio of the source domain and the target domain in model training is updated, which comprises: According to a preset weight distribution formula, the weight distribution ratio is calculated, and the weight distribution formula comprises: ； w g ( t )=1- w s ( t )； wherein, w s (t) represents the weight proportion of the source domain, w g t ) represents the weight proportion of the target domain, λ t ) represents the dynamic adjustment coefficient, w i represents the initial weight, α represents the dynamic coefficient sensitivity, η represents the weight transfer rate, t represents the current time, τ for traversing all time points from the initial time to the current time t, e is a natural constant.​​ 5. The method of claim 1, wherein, According to the target domain expansion data, the geological disaster risk evaluation result of the target domain is obtained, which comprises: According to the target domain expansion data, the target domain is divided into a plurality of evaluation units; Based on the expanded terrain data, human activity index and meteorological data of each evaluation unit, the susceptibility comprehensive index and instability probability of each evaluation unit are determined, the susceptibility comprehensive index is used to represent the geological disaster occurrence possibility of the evaluation unit, and the instability probability is used to represent the probability of the evaluation unit losing stability and causing geological disasters; According to the susceptibility comprehensive index and the instability probability, the geological disaster risk index is determined; Based on the geological disaster risk index, the geological disaster risk evaluation result is generated.

6. The method of claim 5, wherein, According to the susceptibility comprehensive index and the instability probability, the geological disaster risk index is determined, which comprises: According to a preset danger index formula, the geological disaster danger index is calculated, and the danger index formula comprises: H i =Y i / Y max ×P i ; wherein H i represents the geological disaster risk index of the i-th evaluation unit under a certain rainfall condition, Y i represents the i-th evaluation unit, Y max represents the maximum value of the susceptibility comprehensive index, P i represents the instability probability.

7. An electronic device, comprising: One or more processors and memories are included; The memory is coupled with the one or more processors, the memory is used to store computer program code, the computer program code comprises computer instructions, and the one or more processors call the computer instructions to make the electronic device execute the method in any one of claims 1-6.

8. A computer readable storage medium storing computer instructions, characterized in that, When the computer instructions run on the electronic device, the electronic device executes the method in any one of claims 1-6.

9. A computer program product, characterised in that, When the computer program product runs on the electronic device, the electronic device executes the method in any one of claims 1-6.

Images