Geological disaster three-dimensional identification method based on deep learning

Abstract

The invention discloses a geological disaster three-dimensional identification method based on deep learning, and relates to the technical field of automatic identification. Comprising the following steps: acquiring multi-source geological data, and generating a multi-modal data set; performing preprocessing and feature extraction on each data type in the multi-modal data set to generate an initial feature set; a dynamic weight adjustment mechanism is designed, an iterative step length feature set is generated according to the node time step length and the data features, and the saliency of feature iteration is calculated; calculating fluctuation stable values of the features, performing risk feature division on all the features in the monitoring period, and respectively generating corresponding risk sets; and in combination with historical data, according to the risk set, calculating a three-dimensional identification early warning value of the geological disaster in the monitoring period, and generating a final three-dimensional identification result. The geological disaster recognition accuracy and early warning timeliness are improved, and it is ensured that the potential risk of the disaster can be captured at the critical moment.

Description

Technical Field

[0001] The present invention relates to the field of automatic identification technology, and in particular to a three-dimensional identification method of geological disasters based on deep learning. Background Art

[0002] With the intensification of global climate change and the increase of human activities, the frequency and destructiveness of geological disasters have shown a significant upward trend. After years of development, the detection and identification technology of geological disasters has gradually shifted from traditional ground monitoring methods to modern remote sensing, sensor networks and data analysis technologies.

[0003] However, the existing geological disaster monitoring and identification technologies still have some shortcomings. For example, when processing multimodal data, existing technologies often lack effective fusion strategies, resulting in information redundancy or contradiction between different data sources, reducing the recognition accuracy of the model. Although deep learning models have strong feature extraction capabilities, their processing of time series data is still not flexible enough, and it is difficult to dynamically adjust weights for different seasons or climate conditions, resulting in a lack of timeliness and accuracy in monitoring data. Summary of the invention

[0004] In view of the problems existing in the existing geological disaster monitoring and identification technology, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to improve the accuracy of geological disaster identification and the timeliness of early warning.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides a method for three-dimensional identification of geological disasters based on deep learning, which includes acquiring multi-source geological data to generate a multimodal data set; preprocessing and feature extraction of each data type in the multimodal data set to generate an initial feature set; designing a dynamic weight adjustment mechanism to arrange and map the time node steps in the monitoring period; generating an iterative step feature set based on the node time step and data characteristics, and calculating the significance of feature iteration; calculating the fluctuation stability value of the feature based on the significance of the feature iteration; dividing all features within the monitoring period into risk features based on the fluctuation stability value of the feature, and generating corresponding risk sets respectively; combining historical data and calculating the warning value of geological disasters within the monitoring period based on the risk set to generate a final three-dimensional identification result.

[0008] As a preferred solution of the method for three-dimensional identification of geological disasters based on deep learning described in the present invention, the process of forming the initial feature set includes: extracting features using a deep learning model to generate a soil erosion feature set, wherein the features include vegetation coverage, slope, rainfall intensity and soil type; obtaining the time information of each feature and recording the feature value at each time point: T k →TZ k :(a,b,c,d,e), where T k represents the kth time point, TZ k Indicates the feature values ​​at different time points, a~e respectively represent the feature values ​​of key features; generate a performance feature set according to different time points: BS ij ={TZ1,TZ2,...,TZ m}, where BS ij It represents the set of features of feature j of data type i, and m represents the number of time points.

[0009] As a preferred solution of the method for three-dimensional identification of geological hazards based on deep learning described in the present invention, the dynamic weight adjustment mechanism is: the monitoring period T is divided into multiple time node steps, recorded as TN A , A is the step number, and the step setting changes dynamically according to seasonal characteristics; the cumulative sum of all time node steps is equal to the total duration of the entire monitoring period; let the initial weight set be W s2 ={w1,w2,...,w m0}, where m0 is the number of features; obtain the generated final stereo recognition result, analyze the contribution of each feature in the final stereo recognition result to risk assessment, calculate the actual impact weight of each feature in the final stereo recognition result, and form the actual impact weight set W s4 ; Calculate the actual impact weight set W s4 and the initial weight set W s2 The ratio R z , and according to the ratio R z To W s2 Adjust the weights in:

[0010] w new_h =w h ×(1+α×log(R z ))

[0011] Among them, α is the adjustment coefficient, w new_h is the weight of the hth feature after adjustment; use the adjusted weight to recalculate the feature significance of the past N cycles, compare the changes in the significance distribution before and after the adjustment, and accept the adjusted weight if the change in the significance distribution exceeds the preset threshold; otherwise, retain the initial weight set W s2The weight in .

[0012] As a preferred solution of the method for three-dimensional identification of geological hazards based on deep learning described in the present invention, wherein: the step feature set of the generation iteration and the calculation of the significance of the feature iteration include: for each time node step TN A , extract eigenvalues, and generate step feature set SS(TN A ); Step feature set SS(TN A )={f1(TN A ),f2(TN A ),...,f ne (TN A )}, where f ne is the neth feature at time step TN A The value within; the significance evaluation calculation is performed on the iterative process of each feature, and the significance evaluation calculation formula is as follows:

[0013] Sig(f h ,TN A )=w h ×|f h (TN A )-f h (TN A -1)| / f h (TN A -1)

[0014] Among them, Sig(f h ,TN A ) indicates that the hth feature is at time step TN A The significance of the feature on the current step size is the change between the current step size and the previous step size, w h is the weight of the h-th feature.

[0015] As a preferred solution of the method for three-dimensional identification of geological hazards based on deep learning described in the present invention, wherein: the risk feature classification of all features within the monitoring period and the generation of corresponding risk sets respectively include: setting a fluctuation stability value threshold SV related to soil and water loss risk z , if SV(f h ) <SV z , then the feature f h is marked as a high-risk feature; if SV(f h )≥SV z , then the feature f h is marked as a low-risk feature, where SV(f h ) is the fluctuation stable value of the feature; all high-risk features are assigned to the high-risk feature set HR s, all low-risk features are assigned to the low-risk feature set LR s .

[0016] As a preferred solution of the method for three-dimensional identification of geological disasters based on deep learning described in the present invention, wherein: the combination of historical data and the calculation of the three-dimensional identification warning value of geological disasters within the monitoring period according to the risk set includes the following steps: based on historical geological disaster data, a correlation model between geological disaster risks and various features is established through a machine learning method; the high-risk feature and low-risk feature data of the current monitoring period are input into the trained correlation model, the correlation model analyzes the input feature data, combines historical patterns, judges the importance of the feature in the monitoring period, and calculates the risk contribution to the occurrence of geological disasters; the correlation model outputs the impact weight value of each feature according to the performance of the feature; the warning value is calculated according to the fluctuation stability value and the impact weight value of each feature in the high-risk feature set and the low-risk feature set; according to the calculated warning value PW, combined with the preset warning value threshold PW y , generate the final recognition result: if PW ≥ PW y , it is judged that there is a high risk of geological disasters in the current monitoring period and a high-risk warning is generated; if PW <PW y , it is judged that the geological disaster risk in the current monitoring period is low and a low-risk warning is generated.

[0017] As a preferred solution of the method for three-dimensional identification of geological disasters based on deep learning described in the present invention, the calculation formula of the warning value is as follows:

[0018]

[0019] Among them, ω h is the feature f h The influence weight value of .

[0020] In a second aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the deep learning-based three-dimensional identification method of geological disasters as described in the first aspect of the present invention are implemented.

[0021] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the method for three-dimensional identification of geological disasters based on deep learning as described in the first aspect of the present invention are implemented.

[0022] The beneficial effects of the present invention are as follows: the present invention introduces a dynamic weight adjustment mechanism, realizes dynamic adjustment of the data time series according to the step lengths of different time nodes in the monitoring period, especially in combination with seasonal factors, optimizes the step length mapping by dynamically adjusting the time step length, so that the data at different time points can better reflect the specific geological feature changes; and the characteristic change trend between different time nodes can be reflected by the calculation formula of the characteristic significance, so that the model can automatically identify the characteristics of high-risk geological disasters, ensuring that the key risk factors within the monitoring period can be effectively identified. The present invention improves the accuracy of geological disaster identification and the timeliness of early warning, ensures that the potential risks of disasters can be captured at critical moments, and thus provides strong decision-making support for preventing and responding to geological disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 This is a flow chart of the three-dimensional identification method of geological hazards based on deep learning.

[0025] Figure 2 FIG. 1 is a schematic diagram of the structure of a terminal in the first embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0027] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0028] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0029] Example 1

[0030] Reference Figure 1-2, which is the first embodiment of the present invention, and provides a method for three-dimensional identification of geological disasters based on deep learning, comprising:

[0031] S1: Acquire multi-source geological data and generate a multimodal dataset; perform preprocessing and feature extraction on each data type in the multimodal dataset to generate an initial feature set.

[0032] Specifically, multi-source data related to soil erosion are obtained, including but not limited to: high-resolution satellite images and aerial images, LiDAR point cloud data (for terrain change detection), rainfall data, data required for soil erosion models (such as RUSLE model parameters), terrain data (such as DEM, slope, aspect, etc.), land use / land cover data, and hydrological data (such as rivers, reservoirs, etc.).

[0033] The acquired multi-source data are preprocessed, including denoising, registration, and fusion operations, to generate a multimodal dataset of soil and water loss.

[0034] The deep learning model is used to extract features and generate a set of soil and water loss features, which include vegetation coverage, slope, rainfall intensity, and soil type.

[0035] Furthermore, the time information of each feature is obtained, including the acquisition time and processing time, and the feature value at each time point is recorded: T k →TZ k :(a,b,c,d,e).

[0036] Among them, T k represents the kth time point, TZ k Represents the characteristic value of the time point, a~e represent the characteristic values ​​of key features such as vegetation coverage, slope, rainfall intensity, soil erosion factor and human activity intensity; generate performance feature sets according to different time points: BS ij ={TZ1,TZ2,...,TZ m}, where BS ij It represents the set of features of feature j of data type i, and m represents the number of time points.

[0037] S2: Design a dynamic weight adjustment mechanism to arrange and map the time node steps in the monitoring cycle; generate an iterative step feature set based on the node time step and data characteristics, and calculate the significance of the feature iteration.

[0038] S2.1: Design a dynamic weight adjustment mechanism, divide the monitoring period into multiple time node steps, consider seasonal factors (such as rainy season and dry season), evenly distribute the time nodes throughout the monitoring period, ensure that the feature data is evenly sampled throughout the period, map the data of each time node to the corresponding time point, construct a time feature mapping relationship, and ensure that the data at different time points have temporal coherence.

[0039] Among them, the dynamic weight adjustment mechanism is designed as follows:

[0040] The monitoring period T is divided into multiple time node steps, denoted as TN A , A is the step number, and the step setting changes dynamically according to the seasonal characteristics. For example, it can be set to a monthly or quarterly step, depending on the seasonal intensity of the soil erosion process.

[0041] Ensure that the cumulative sum of the steps of all time nodes is equal to the total duration of the entire monitoring period. The step mapping can be optimized based on historical data and its changing trends, so that the step weights in different time periods can be dynamically adjusted to ensure that the step can accurately reflect the changes in geological characteristics of the time nodes.

[0042] In the division of step length and weight adjustment, the influence of seasonal climate (such as severe soil erosion in the rainy season and less severe soil erosion in the dry season) on geological characteristics is considered as an important basis for the division and adjustment of step length. For example, during the rainy season, the step length should be refined to capture the frequent changes in soil erosion characteristics, while in the dry season, the step length can be appropriately relaxed to reduce unnecessary calculation burden. The initial weight set used to calculate the feature significance is W s2 ={w1,w2,...,w m0}, where m0 is the number of features.

[0043] S2.2: Iteratively update the features within each time node step and calculate the significance of the feature iteration.

[0044] Specifically, for each time node step TN A , extract eigenvalues, and generate step feature set SS(TN A ), the characteristic values ​​include vegetation coverage, slope, rainfall intensity and soil type, among which the step feature set SS(TN A )={f1(TN A ),f2(TN A ),...,f ne (TN A )}, where f ne is the neth feature at time step TN A The value within.

[0045] According to the step feature set of each time node, historical data is used for iterative update.

[0046] Furthermore, the significance evaluation is calculated for the iterative process of each feature to determine the variation range of the feature at different time nodes.

[0047] Furthermore, the calculation formula for significance evaluation is as follows:

[0048] Sig(f h ,TN A )=w h ×|f h (TN A )-f h (TN A -1)| / f h (TN A -1)

[0049] Among them, Sig(f h ,TN A ) indicates that the hth feature is at time step TN A The significance of the feature on the current step size is the change between the current step size and the previous step size, w h is the initial weight of the h-th feature.

[0050] S2.3: Based on the final stereo recognition results, adjust the dynamic weights in S2.1. For example, if a feature is more significant in a high-risk area, increase its weight; if a feature is less significant in a low-risk area, reduce its weight.

[0051] Specifically, the final stereo recognition result is obtained, the contribution of each feature in the S4 result to the final risk assessment is analyzed, and the actual impact weight of each feature in the S4 result is calculated to form the actual impact weight set W s4 .

[0052] Compare the current weight of S2 with the actual impact weight of the feature in S4, and compare the actual impact weight set W s4 and the initial weight set W s2 The ratio R z , and according to the ratio R z To W s2 The weights in are adjusted:

[0053] w new_h =w h ×(1+α×log(R z ))

[0054] Among them, α is the adjustment coefficient, which controls the adjustment range and is usually set between 0.1 and 0.5. new_h is the weight of the h-th feature after adjustment.

[0055] In addition, it is necessary to ensure that the sum of the adjusted S2 weights is 1, that is, weight normalization is required.

[0056] Use the adjusted S2 weights to recalculate the feature significance of the past N cycles, and compare the changes in the significance distribution before and after the adjustment. If the change in the significance distribution exceeds the preset threshold (such as 20%), accept the new weights; otherwise, retain the original weights; if retained, use the verified adjusted weights to update the feature significance calculation formula, and pass the adjusted weights and updated feature significance to subsequent steps.

[0057] It can be seen that by comparing the final stereo recognition result with the initial weight, the feature weight is dynamically adjusted. Compared with the traditional static feature weight allocation method, the present invention significantly improves the system's ability to flexibly respond to feature changes.

[0058] In the prior art, weights are usually preset and it is difficult to respond to changes in the environment or data in real time. The present invention dynamically adjusts weights so that the system can continuously optimize its judgment capabilities as features change, avoiding misjudgments or inefficient identification caused by outdated weight configurations. This not only effectively improves the accuracy of risk identification, but also reduces human intervention in an adaptive manner.

[0059] S3: Calculate the fluctuation stability value of the feature according to the significance of the feature iteration; divide all features within the monitoring period into risk features according to the fluctuation stability value of the feature, and generate corresponding risk sets respectively.

[0060] S3.1: Based on the significance of feature iteration, calculate the fluctuation stability value of the feature.

[0061] Preferably, the calculation process of the fluctuation stability value of the feature is as follows:

[0062]

[0063] Where W represents the total number of steps in the monitoring period.

[0064] It can be seen that the stability of the feature during the entire monitoring period is evaluated by calculating the average value of the significance, and the obtained fluctuation stability value can be used as a basis for measuring the stability of feature changes.

[0065] S3.2: Classify all characteristics within the monitoring period into risk characteristics.

[0066] Set the fluctuation stability threshold SV related to soil erosion risk z , if SV(f h ) <SV z , then the feature f his marked as a high-risk feature; if SV(f h )≥SV z , then the feature f h Labelled as a low risk feature.

[0067] Assign all high-risk features to the high-risk feature set HR s , where the high-risk feature set may include features such as high rainfall intensity, low vegetation cover, and steep slopes.

[0068] Assign all low-risk features to the low-risk feature set LR s , where the low-risk feature set usually includes relatively stable environmental conditions, such as flat terrain, stable rainfall, and high vegetation coverage.

[0069] S4: Combined with historical data, based on the high-risk feature set and the low-risk feature set, calculate the warning value of geological disasters within the monitoring period and generate the final three-dimensional recognition result.

[0070] S4.1: Based on historical geological disaster data, a correlation model between geological disaster risk and various characteristics is established through machine learning methods (such as random forests, decision trees, etc.). This correlation model is used to evaluate the impact weight of high-risk characteristics and low-risk characteristics in the current monitoring period on the occurrence of geological disasters.

[0071] Specifically, historical geological disaster data are collected, especially key features related to risk events (such as rainfall, vegetation coverage, soil erosion, etc.), and the performance of each feature before the geological disaster occurs is marked. The historical data is trained using machine learning methods to form an association model between features and geological disaster risks. This association model will assign corresponding weights to each feature, representing the degree of influence of the feature in the geological disaster risk assessment.

[0072] The high-risk and low-risk feature data of the current monitoring period are input into the trained association model. The association model analyzes the input feature data and combines historical patterns to determine the importance of these features in the monitoring period and calculate their contribution to the risk of geological disasters.

[0073] The association model outputs the impact weight value of each feature based on the performance of the feature. The higher the weight value, the greater the impact of the feature on geological disasters. For example, if the rainfall increases significantly during the current monitoring period, the impact weight of the rainfall feature may be higher, indicating that rainfall plays a key role in triggering geological disasters.

[0074] S4.2: Group HR by high-risk characteristics s and low risk feature set LR sThe early warning value is calculated based on the fluctuation stability value and influence weight of each feature:

[0075]

[0076] Among them, ω h is the feature f h High-risk characteristics are usually assigned higher impact weights to reflect their significant impact on geological hazard risks.

[0077] S4.3: Based on the calculated warning value PW, combined with the preset warning value threshold PW y , generate the final recognition result:

[0078] If PW ≥ PW y , it is judged that there is a high risk of geological disasters in the current monitoring period and a high-risk warning is generated;

[0079] If PW <PW y , it is judged that the geological disaster risk in the current monitoring period is low and a low-risk warning is generated.

[0080] Optionally, a three-dimensional disaster risk map can be generated by combining high / low risk areas and major risk characteristics using GIS platform and 3D modeling technology. The risk map display includes:

[0081] Different colors represent different risk levels, for example: red for high risk, yellow for medium risk, and green for low risk; disaster types are represented by different symbols; the impact range is represented by contour lines on topographic maps, and potential disaster-stricken areas are marked.

[0082] It can be seen that the use of the fluctuation stability value of the present invention enables the system to make real-time adjustments based on the current monitoring data, and the introduction of the weight value further enhances the accuracy of the assessment and prevents the misjudgment phenomenon that may exist in the traditional static model. By accurately calculating the warning value, this step can give personalized risk warnings for specific characteristics under different geological conditions, solving the one-size-fits-all risk assessment problem that existed in previous warning systems.

[0083] This embodiment also provides a computer device, which is suitable for the case of a three-dimensional identification method of geological disasters based on deep learning, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the three-dimensional identification method of geological disasters based on deep learning as proposed in the above embodiment.

[0084] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0085] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for three-dimensional identification of geological disasters based on deep learning as proposed in the above embodiment is implemented.

[0086] In summary, the present invention introduces a dynamic weight adjustment mechanism, and realizes the dynamic adjustment of the data time series according to the step lengths of different time nodes in the monitoring period, especially in combination with seasonal factors. The step length mapping is optimized by the dynamic adjustment of the time step length, so that the data at different time points can better reflect the changes in specific geological characteristics; and the characteristic change trend between different time nodes can be reflected by the calculation formula of the characteristic significance, so that the model can automatically identify the characteristics of high-risk geological disasters, ensuring that the key risk factors within the monitoring period can be effectively identified. The present invention improves the accuracy of geological disaster identification and the timeliness of early warning, ensuring that the potential risks of disasters can be captured at critical moments, and thus provides powerful decision-making support for preventing and responding to geological disasters.

[0087] Example 2

[0088] Referring to Table 1, which is the second embodiment of the present invention, this embodiment provides a three-dimensional identification method of geological disasters based on deep learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0089] In a certain mountainous basin, in order to better prevent and control soil erosion, this deep learning-based three-dimensional identification method of geological disasters was applied. The basin covers an area of ​​about 500 square kilometers, with complex terrain, uneven vegetation coverage, and uneven rainfall distribution. It is a typical area prone to soil erosion.

[0090] First, multi-source geological data from the past five years were collected in the region. These data include: monthly high-resolution satellite images (resolution of 0.5 meters), quarterly LiDAR point cloud data (point density of 10 points / square meter), daily rainfall data, annually updated land use / land cover data, and monthly updated DEM data (resolution of 5 meters). Records of soil erosion events in the region over the past 10 years were also collected as historical data for model training and verification, and the data was preprocessed.

[0091] Next, the improved U-Net deep learning model was used to extract features from the preprocessed multimodal data. The extracted features included vegetation coverage, slope, rainfall intensity, soil erosion factor, and human activity intensity. These features were recorded monthly to form an initial feature set.

[0092] In the design of the dynamic weight adjustment mechanism, considering the seasonal characteristics of the region, a year is divided into 12 monthly steps, but a finer half-month step is used in the rainy season (June-September) to capture the frequent changes in soil erosion characteristics. Within each step, the average value and change rate of the feature are calculated to generate a step feature set.

[0093] For the significance calculation of feature iteration, an improved significance evaluation formula is used, which not only considers the change of feature value, but also introduces the sensitivity coefficient of feature to soil erosion. For example, the sensitivity coefficients of rainfall intensity and vegetation coverage are high, while the sensitivity coefficient of slope is relatively low.

[0094] In the risk feature classification stage, the fluctuation stability value threshold of each feature is determined by analyzing historical data. Features above the threshold are marked as high-risk features, and those below the threshold are marked as low-risk features. Fuzzy set theory is also introduced to make the risk level of features smoothly transition between high and low risks, which is more in line with the actual situation.

[0095] Finally, in the calculation of stereoscopic recognition warning values, a soil and water loss risk prediction model was trained using the random forest algorithm. This model not only takes into account the current feature status, but also introduces time series analysis to consider the historical change trend of the features. The warning value is calculated using a weighted summation method, and the weights are automatically learned by the random forest model.

[0096] In order to verify the effectiveness of this method, six typical areas in the basin were selected for a 12-month monitoring experiment. These areas include different terrain features and land use types, which can represent the overall situation of the basin. Some of the experimental results are shown in the following table:

[0097] Table 1

[0098]

[0099] As shown in the table above, this method has achieved a high early warning accuracy in 6 different types of areas, with an average of 93.33%. In particular, in Area D (bare land), where the risk of soil erosion is the highest, the early warning accuracy is as high as 96%, which shows that this method is particularly effective in identifying high-risk areas. In contrast, the accuracy of traditional soil erosion early warning methods in areas with complex terrain and diverse land use types is usually only about 80-85%.

[0100] It can be seen from the table that this method shows good early warning effects in different types of areas (such as farmland, woodland, grassland, bare land, etc.), which shows that this method has strong adaptability and can automatically adjust the early warning strategy according to the characteristics of different regions. Especially in the composite area (F area), the early warning accuracy rate reaches 94%, indicating that this method can handle complex surface conditions well.

[0101] By comparing various indicators in different areas, it can be seen that the method has achieved a balance in the three key indicators of accuracy, false alarm rate and missed alarm rate. This deep learning-based three-dimensional identification method of geological disasters has shown significant advantages in soil and water loss early warning. It not only improves the accuracy of the early warning, but also enhances the adaptability and real-time performance of the system.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A three-dimensional identification method of geological hazards based on deep learning, characterized by: include, Acquire multi-source geological data and generate multimodal datasets; Preprocessing and feature extraction are performed on each data type in the multimodal data set to generate an initial feature set; Design a dynamic weight adjustment mechanism to arrange and map the time node steps in the monitoring cycle; Generate iterative step feature set according to node time step and data features, and calculate the significance of feature iteration; According to the significance of the feature iteration, the fluctuation stability value of the feature is calculated; according to the fluctuation stability value of the feature, all features within the monitoring period are divided into risk features, and corresponding risk sets are generated respectively; Combined with historical data and based on the risk set, the warning value of geological disasters within the monitoring period is calculated to generate the final three-dimensional recognition result.

2. The method for three-dimensional identification of geological hazards based on deep learning as claimed in claim 1, characterized in that: The process of forming the initial feature set includes: Using deep learning models to extract features and generate a set of soil erosion features, including vegetation coverage, slope, rainfall intensity, and soil type; Get the time information of each feature and record the feature value at each time point: T k →TZ k :(a,b,c,d,e), where T k represents the kth time point, TZ k represents the characteristic values ​​at different time points, and a~e represent the characteristic values ​​of key features respectively; Generate performance feature sets based on different time points: BS ij ={TZ1,TZ2,...,TZ m }, where BS ij It represents the set of features of feature j of data type i, and m represents the number of time points.

3. The method for three-dimensional identification of geological hazards based on deep learning as claimed in claim 2, characterized in that: The dynamic weight adjustment mechanism is: The monitoring period T is divided into multiple time node steps, denoted as TN A , A is the step number, and the step setting changes dynamically according to seasonal characteristics; The cumulative sum of all time node steps is equal to the total duration of the entire monitoring cycle; Let the initial weight set be W s2 ={w1,w2,...,w m0 }, where m0 is the number of features; Obtain the generated final stereo recognition result, analyze the contribution of each feature in the final stereo recognition result to risk assessment, calculate the actual impact weight of each feature in the final stereo recognition result, and form the actual impact weight set W s4 ; Calculate the actual influence weight set W s4 and the initial weight set W s2 The ratio R z , and according to the ratio R z To W s2 Adjust the weights in: w new_h =w h ×(1+α×log(R z )) Among them, α is the adjustment coefficient, w new_h is the weight of the h-th feature after adjustment; Use the adjusted weights to recalculate the feature significance of the past N cycles, and compare the changes in the significance distribution before and after the adjustment. If the change in the significance distribution exceeds the preset threshold, the adjusted weights are accepted; otherwise, the initial weight set W is retained. s2 The weight in .

4. The method for three-dimensional identification of geological hazards based on deep learning as claimed in claim 3, characterized in that: The step of generating an iterative step feature set and calculating the iterative significance of the feature comprises: For each time node step TN A , extract eigenvalues, and generate step feature set SS(TN A ); Step feature set SS(TN A )={f1(TN A ), f2(TN A ),...,f ne (TN A )}, where f ne is the neth feature at time step TN A The value within The significance evaluation calculation is performed on the iterative process of each feature. The significance evaluation calculation formula is as follows: Sig(f h ,TN A )=w h ×|f h (TN A )-f h (TN A -1)| / f h (TN A -1) Among them, Sig(f h ,TN A ) indicates that the hth feature is at time step TN A The significance of the feature on the current step size is the change between the current step size and the previous step size, w h is the weight of the h-th feature.

5. The method for three-dimensional identification of geological hazards based on deep learning as claimed in claim 4, characterized in that: The risk feature classification of all features within the monitoring period and the generation of corresponding risk sets respectively include: Set the fluctuation stability threshold SV related to soil erosion risk z , if SV(f h ) <SV z , then the feature f h is marked as a high-risk feature; if SV(f h )≥SV z , then the feature f h is marked as a low-risk feature, where SV(f h ) is the characteristic fluctuation stable value; Assign all high-risk features to the high-risk feature set HR s , all low-risk features are assigned to the low-risk feature set LR s .

6. The method for three-dimensional identification of geological hazards based on deep learning as claimed in claim 5, characterized in that: The calculation of the three-dimensional identification warning value of geological disasters within the monitoring period in combination with historical data and based on the risk set includes the following steps: Based on historical geological disaster data, a correlation model between geological disaster risk and various characteristics is established through machine learning methods; The high-risk and low-risk feature data of the current monitoring period are input into the trained association model. The association model analyzes the input feature data and combines historical patterns to determine the importance of the feature in the monitoring period and calculate the risk contribution to the occurrence of geological disasters. The association model outputs the impact weight value of each feature based on the performance of the feature; Calculate the warning value based on the fluctuation stability value and impact weight value of each feature in the high-risk feature set and the low-risk feature set; According to the calculated warning value PW, combined with the preset warning value threshold PW y , generate the final recognition result: If PW ≥ PW y , it is judged that there is a high risk of geological disasters in the current monitoring period and a high-risk warning is generated; If PW <PW y , it is judged that the geological disaster risk in the current monitoring period is low and a low-risk warning is generated.

7. The method for three-dimensional identification of geological hazards based on deep learning according to claim 6, characterized in that: The calculation formula of the warning value is as follows: Among them, ω h is the feature f h The influence weight value of .

Images