Loess landslide detection method and system based on grade differentiation learning grey wolf optimization

Abstract

The invention belongs to the technical field of image processing, and discloses a loess landslide detection method and system optimized by a multi-population grey wolf optimization algorithm based on grade differentiation learning. Performing data enhancement on the landslide data set by using a data enhancement generator; designing a grey wolf optimization algorithm model based on multi-level differential learning through the segmentation network model, and optimizing hyper-parameters in the segmentation network model by using an improved algorithm to realize adaptive adjustment and cooperative calculation of three hyper-parameters, namely a learning rate, an iteration number and a batch processing number, in the segmentation network; and finally carrying out loess landslide detection. According to the method, algorithm improvement is carried out, so that the balance capability of global exploration and local development is enhanced, and later diversity loss and local optimum are avoided. By combining the excellent semantic segmentation capability of VIG and the improved grey wolf optimization algorithm, the method can effectively overcome the feature extraction difficulty caused by factors such as geological structure and vegetation coverage, and significantly improves the accuracy of loess landslide recognition.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to a loess landslide detection method and system based on multi-group gray wolf optimization of graded differentiation learning. Background Art

[0003] The prevention and control of loess landslides primarily involves identifying potential loess landslide hazards. Early loess landslide surveys were conducted through field investigations, which yielded relatively accurate results. However, these surveys were time-consuming, required significant human and material resources, and were significantly influenced by human factors and the geographical environment. Furthermore, some geological hazard source areas are located in elevated areas, making it difficult for investigators to access and identify hidden landslide hazard sites within the area. New image processing technologies and high-performance computer models offer new solutions and technical approaches to this problem. However, an analysis of existing literature reveals that approaches to loess landslide detection still have the following shortcomings: In practical applications, false or missed detections are highly likely when encountering features with similar image characteristics to loess landslides, such as bare land, dry land, and certain man-made structures. Therefore, effectively avoiding confusion between loess landslides and other features such as dry land and bare land during identification remains a thorny issue facing various landslide detection methods.

[0004] In recent years, deep learning-based semantic segmentation models have become increasingly widely used in landslide detection and have become a mainstream technology. Among them, models based on convolutional neural networks (CNNs), such as fully convolutional networks (FCNs), UNet, and DeepLabv3+, have been applied to landslide detection, significantly improving the efficiency and accuracy of large-scale landslide detection. However, because CNNs expand their receptive fields through stacked convolutions, they have certain limitations in extracting global information. Visual graph neural networks (VIGs) can more effectively utilize the global information and contextual relationships of an image, thereby improving the accuracy of segmentation results. Furthermore, VIGs offer greater flexibility and adaptability, capable of processing image inputs of various sizes and resolutions without requiring additional image adjustments or cropping. Therefore, this paper proposes to use VIGs for loess landslide detection.

[0005] Compared with existing segmentation network models such as FCN, UNet, and DeepLabv3+, it is difficult to utilize the spatial relationship and contextual information between pixels in the task of loess landslide identification. At the same time, there is a problem of insufficient processing of target boundaries, resulting in unclear boundaries of the segmentation results, which in turn affects the segmentation effect. VIG can capture the overall semantic information in the image and maintain a high accuracy. However, the accuracy of the final segmentation of the network is related to parameters such as the number of iterations, the number of batches, and the learning rate. The VIG structure hierarchy and hyperparameter settings require a lot of experiments and manual experience to find the optimal one. If the parameters are not set properly, it will result in: (1) poor segmentation effect; (2) complex network structure; (3) high computational cost. Swarm intelligence optimization algorithms are highly favored because of their powerful ability to solve complex optimization problems. By applying the Grey Wolf Optimizer (GWO) algorithm with strong search capabilities to adjust the hyperparameters of the VIG network model, it is expected to further improve the recognition performance of the model and reduce errors. However, the Grey Wolf Optimization Algorithm also faces challenges similar to those of other swarm intelligence algorithms, such as being prone to getting trapped in local optima, which leads to slower convergence and decreased accuracy. Its performance is particularly affected when dealing with deep, complex, nonlinear, and highly constrained non-convex multimodal optimization problems.

[0006] Technical problems existing in industrial applications of existing technologies:

[0007] 1. Insufficient segmentation accuracy and inadequate utilization of contextual information in loess landslide identification.

[0008] Problem description:

[0009] Traditional segmentation network models such as FCN, UNet, and DeepLabv3+ struggle to fully utilize the spatial relationships and contextual information between pixels in loess landslide identification tasks.

[0010] These models have shortcomings in handling target boundaries, resulting in unclear boundaries in the segmentation results and affecting the overall segmentation performance.

[0011] Influence:

[0012] The accuracy of loess landslide identification is insufficient, especially in areas with complex terrain and rich details, leading to missed or false detections.

[0013] The ambiguity of boundaries affects subsequent disaster analysis and emergency response decisions.

[0014] 2. Optimization difficulty of the VIG network model

[0015] Problem description:

[0016] VIG can capture the overall semantic information in the image and maintain a high accuracy, but its final segmentation accuracy is closely related to hyperparameters such as the number of iterations, batch size, and learning rate.

[0017] Setting hyperparameters requires extensive experimentation and human experience; improper settings can lead to the following problems:

[0018] 1. Poor segmentation results.

[0019] 2. Complex network structure.

[0020] 3. High computational cost.

[0021] Influence:

[0022] The hyperparameter optimization process is time-consuming and labor-intensive, and requires a large amount of computing resources, which increases the complexity and cost of practical applications.

[0023] Improper parameter settings will seriously affect the performance and applicability of the model.

[0024] 3. Limitations of the Gray Wolf Optimization Algorithm

[0025] Problem description:

[0026] Although the Gray Wolf Optimization Algorithm has strong search capabilities and is suitable for adjusting the hyperparameters of the VIG network model, it also faces challenges similar to other swarm intelligence algorithms:

[0027] It is easy to fall into local optimum, resulting in slower convergence and lower accuracy.

[0028] Performance is significantly affected when dealing with deep, complex, nonlinear, and highly constrained nonconvex multimodal optimization problems.

[0029] Influence:

[0030] The optimization effect is unstable, resulting in the segmentation model's performance not meeting expectations.

[0031] The algorithm's slow convergence speed increases the time cost of the optimization process.

[0032] The main technical challenges faced by existing technologies in industrial applications of landslide identification include:

[0033] 1. Insufficient segmentation accuracy and utilization of contextual information: Traditional segmentation networks struggle to fully utilize the spatial relationships and contextual information between pixels, resulting in inadequate boundary processing and impacting segmentation performance.

[0034] 2. High difficulty in hyperparameter optimization: The segmentation performance of the VIG model depends on the hyperparameter settings. The optimization process is complex and time-consuming, and the computational cost is high.

[0035] 3. Limitations of the Grey Wolf Optimization Algorithm: Although it has strong search capabilities, it is prone to getting trapped in local optima, which affects the convergence speed and optimization accuracy.

[0036] These problems limit the effectiveness and practicality of loess landslide identification models in practical applications, and proposing new optimization methods and strategies to improve model performance has become an urgent technical challenge. Summary of the Invention

[0037] To address the problems existing in the prior art, this invention provides a loess landslide detection method and system optimized by a multi-group gray wolf optimization algorithm based on hierarchical differential learning. This not only improves the accuracy of detection but also increases the efficiency of subsequent landslide hazard monitoring.

[0038] The present invention is implemented as follows: a loess landslide detection method optimized by a multi-group gray wolf optimization algorithm based on hierarchical differentiation learning, comprising:

[0039] Data augmentation generators were used to augment the loess landslide dataset; a gray wolf optimization algorithm model based on multi-level differential learning was designed using a segmentation network model, and the improved algorithm was used to optimize the hyperparameters in the segmentation network model, realizing the adaptive adjustment and collaborative calculation of the three hyperparameters of learning rate, number of iterations, and number of batches in the segmentation network; finally, loess landslide detection was performed.

[0040] Further, the implementation method is as follows:

[0041] Step 1: Preprocess the images of the public landslide dataset landslide4sense for semantic segmentation and use the data enhancer method to enhance the landslide image data.

[0042] Step 2: Establish a segmentation network model UNet-VIG and design a Multiple group gray wolf optimization algorithm based on hierarchical differentiated learning (HDLMGWO) model. Use the HDLMGWO model to optimize the hyperparameters of the segmentation network model UNet-VIG, and achieve collaborative calculation and adaptive adjustment of the three hyperparameters of the segmentation network: learning rate, batch size, and number of iterations.

[0043] Step 3: Initialize the positions of individual gray wolves in the HDLMGWO model;

[0044] In step 4, a nonlinear control parameter based on the hyperbolic tangent function is proposed to adjust the exploration and development ratio of the algorithm.

[0045] In step five, a hierarchical differentiation learning mechanism is proposed, which greedily selects the original position update mechanism and the learning mechanism to improve the position update method, so that the algorithm can learn towards the global optimum and the neighborhood optimum.

[0046] In step six, a multi-population recombination evolution strategy was proposed. Gray wolves in different subgroups will adopt evolutionary strategies that are consistent with the current subgroup fitness level, thereby enhancing the diversity of the population.

[0047] Step 7: In order to both improve the global exploration capability of the gray wolf and enhance the quality of the search solution during the iteration process, this paper proposes a learning operator based on hyperbolic sine function.

[0048] Step 8: Determine whether the HDLMGWO model has reached the maximum number of iterations. If not, return to step 4. If yes, proceed to the next step.

[0049] Step nine: Use the optimal hyperparameter values ​​searched by the improved grey wolf optimization algorithm to optimize the segmentation network model for loess landslide detection.

[0050] Furthermore, in step 1, a data augmentation method is used to enhance the loess landslide data. Normalization, image rotation, horizontal translation, vertical translation, shear transformation, and horizontal flipping are selected as processing methods of the data augmentation method. The enhanced loess landslide dataset is used as an input variable of the segmentation network. The segmentation network generates an output variable, thereby establishing a mapping relationship between the input variable and the output variable.

[0051] Furthermore, in step 2, a segmentation network model UNet-VIG is established, which consists of an encoder, a data exchange module, and a decoder. The HDLMGWO model is designed to optimize the hyperparameters in the segmentation network. The three hyperparameters of learning rate, batch number, and number of iterations in the segmentation network are optimized using the HDLMGWO algorithm. The population size in the HDLMGWO algorithm is set to 3 and the maximum number of iterations is set to 10. In the algorithm, the solution space is set for the three hyperparameters of learning rate, batch number, and number of iterations in the VIG network. ub = {0.1, 8, 450} represents the upper limit of the solution space, and lb = {0.001, 1, 350} represents the lower limit of the solution space. Each dimension of ub and lb corresponds to the upper and lower bounds of the learning rate, batch number, and number of iterations of the segmentation network model, respectively.

[0052] Furthermore, in step three, the HDLMGWO model initializes the gray wolf population. Then, the fitness value of each individual gray wolf is calculated, and they are sorted according to their fitness values. The location of each gray wolf is interpreted as a hyperparameter of the segmentation network model. The segmentation network model is optimized, trained using the training dataset, and the F1 score for loess landslide identification is predicted using the test dataset. The result of 1-F1 is used as the fitness value of the gray wolf.

[0053] Furthermore, in step four, the convergence factor is adjusted using nonlinear control parameters based on the hyperbolic tangent function. The formula is as follows:

[0054]

[0055] Furthermore, in step five, a hierarchical differential learning mechanism is proposed, which greedily selects the original position update mechanism and the learning mechanism to improve the position update method, enabling the algorithm to learn towards the global optimum and the neighborhood optimum.

[0056] In the learning mechanism, the gray wolf can observe the differences between the gray wolves in four different levels. The mathematical expressions of the differences between the four groups are as follows:

[0057]

[0058] In this context, Gap1 represents the gap between alpha wolves and elite wolves, Gap2 represents the gap between wolves and bottom wolves, Gap3 represents the gap between elite wolves and bottom wolves, and Gap4 represents the gap between two wolves randomly selected from the population. α represents the globally optimal wolf, elite wolves are the best wolves, and bottom represents the bottom wolves. L1 and L2 are two wolves randomly selected from the pack.

[0059] LF k Normalized ratio of the Euclidean distance of the k-th group difference, LF k The expression is as follows:

[0060]

[0061] SF i Assess the knowledge that individual gray wolves need to learn. LR i Let LR represent the learning resistance of the i-th gray wolf. max The greatest learning resistance among individual gray wolves. Generally, the objective function will accept the i-th individual and return the corresponding output, i.e., LR. i SF i The expression is as follows:

[0062]

[0063] For an individual gray wolf, in terms of its own learning ability (SF) i ) and learning factor (LF k ) under the influence of the gray wolf from the k-th group difference (Gap k ) is the knowledge learned in KA k , the formula is shown in (5).

[0064] KA k =SFi ·LF k Gap k (k=1,2,3,4) (5)

[0065] Learning was completed by learning from the knowledge gained from the four sets of differences. The learning process of the i-th gray wolf is shown below:

[0066]

[0067] is the position of the gray wolf in the previous iteration, This shows the position of the gray wolf after the learning mechanism has been updated.

[0068] The original gray wolf's position update mechanism is as follows:

[0069]

[0070] in, is the updated position after the original gray wolf leadership mechanism. The update formula of X1, X2 and X3 is as follows:

[0071] X1=|X α -A1·D α |,X2=|X β -A2·D β |,X3=|X δ -A3·D δ | (8)

[0072] Among them, X α 、X β and X δ Denote the positions of α, β and δ gray wolves respectively. α 、D β and D δ , the formulas for A1, A2, and A3 are as follows.

[0073]

[0074] A1=2·a·r1-a, A2=2·a·r2-a, A2=2·a·r3-a (10)

[0075] r1, r2 and r3 are random numbers between [0,1].

[0076] The formulas for C1, C2, and C3 are as follows:

[0077] C1=2r4, C2=2r5, C3=2r6 (11)

[0078] r4, r5 and r6 are random numbers between [0,1].

[0079] Furthermore, in step six, to avoid the loss of population diversity, we propose a multi-population recombination evolution strategy.

[0080] The position update formula of the upper wolf is as follows:

[0081]

[0082] Where rand is a random number between [0,1], and X Avg The formula for the average position of the best gray wolf at the bottom layer and the best gray wolf at the middle layer is as follows:

[0083]

[0084] Among them, MiddleBest is the best gray wolf in the middle layer, BottomBest is the best gray wolf in the bottom layer, and LR is the growth resistance of gray wolves.

[0085] The formula for updating the position of the mid-level gray wolves is as follows:

[0086]

[0087] Among them, X UpperRand It is a gray wolf randomly selected from the upper layer, wblrand is a Weiber distribution with shape parameter and scale parameter of 1, and rand is a random number between [0,1].

[0088] The update formula for the bottom-level gray wolves is as follows:

[0089]

[0090] Where rand is a random number between [0,1], and BottomBest is the best gray wolf in the bottom layer. This represents the position of the gray wolf after undergoing multiple population recombination and evolutionary strategies. The parameters are P1 = 3 and P2 = 0.1.

[0091] Furthermore, in step seven, in order to both enhance the gray wolf's global exploration capabilities and improve the quality of the search solution during the iteration process, this paper creates a learning operator based on the probability P of the hyperbolic sine function, the mathematical expression of which is defined as follows:

[0092]

[0093] Where rand is a random number between [0,1], and X rand A random individual is selected from the gray wolves, and wblrnd is a random number generated by a Weiber distribution with a scale parameter of 1 and a shape parameter of 1. It is the position of the gray wolf after learning the operator based on the hyperbolic sine function.

[0094] The mathematical expression for P is as follows:

[0095]

[0096] Where sinh is the hyperbolic sine function, t is the current iteration number, and MaxIter is the maximum iteration number.

[0097] Furthermore, in step eight, based on the maximum number of iterations set by the HDLMGWO model, it is determined whether the current individual meets the termination condition. If the termination condition is met, the algorithm result is output. If the final iteration condition is not met, the process returns to step four to continue looping into the next generation until the condition is met.

[0098] Furthermore, in step nine, the optimal position finally found by the gray wolf is interpreted as the hyperparameter of the segmentation network model. The hyperparameter is input into the segmentation network model to construct a complete segmentation network model. The network model is trained using the training set and test set in the landslide dataset, and the trained model is used to make predictions to improve the accuracy of landslide identification.

[0099] Another objective of this invention is to provide a loess landslide detection system optimized by a multi-group gray wolf optimization algorithm based on hierarchical differential learning, comprising:

[0100] The data preprocessing module is used to preprocess images from the publicly available Landslide4Sense dataset for semantic segmentation and to augment the landslide image data using the data augmenter method.

[0101] Segmentation Network Model Module: Establishes the segmentation network model UNet-VIG for semantic segmentation of loess landslide images;

[0102] HDLMGWO Model Optimization Module: Design the HDLMGWO model based on multi-level differential learning gray wolf optimization algorithm to optimize the hyperparameters in the segmentation network model UNet-VIG;

[0103] Iteration Decision Module: Determines whether the HDLMGWO model has reached the maximum number of iterations. If not, it returns to continue iterating; if it does, it proceeds to the next step.

[0104] Optimization Module: The segmentation network model UNet-VIG is optimized using the best hyperparameter values ​​searched by the improved Grey Wolf Optimization Algorithm for loess landslide detection.

[0105] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0106] First, the segmentation network model established by the present invention can effectively capture the overall semantic information in the image and be applied to the identification of loess landslides. In order to overcome the challenges faced by the segmentation network model in network structure design and the high demand for computing and storage resources, we introduced a gray wolf optimization model. By optimizing key hyperparameters such as learning rate, batch processing number and iteration number, this model not only improves the accuracy of loess landslide identification, but also significantly reduces the recognition error. However, similar to other swarm intelligence optimization algorithms, the gray wolf optimization algorithm is also prone to falling into local optimal solutions, especially when dealing with non-convex multimodal optimization problems with deep complexity, nonlinear features and strict constraints, its performance will be affected. To meet this challenge, we draw on the learning mechanism of human society and propose four improvement strategies to enhance the model's ability to balance global exploration and local development, avoid the lack of diversity and fall into local optimal problems. By integrating the excellent semantic segmentation capabilities of VIG with an improved version of the gray wolf optimization algorithm, the present invention effectively solves the feature extraction problems caused by factors such as geological structure and vegetation cover, thereby significantly improving the accuracy of loess landslide identification.

[0107] Second, the present invention mainly solves the following technical problems in the prior art:

[0108] Complex and changeable loess landslide detection environment: The occurrence of loess landslides is affected by multiple factors such as geology, climate, topography, vegetation cover and human activities. These factors are intertwined to form complex and changeable environmental conditions, making the loess landslide detection task extremely challenging.

[0109] Complex segmentation network model design: In loess landslide detection, the design of the segmentation network model requires the optimization of multiple hyperparameters, including the learning rate, batch size, and number of iterations. The selection of these hyperparameters significantly affects model performance, but existing methods struggle to optimize them efficiently and accurately.

[0110] Existing optimization algorithms are prone to falling into local optimal solutions: When dealing with non-convex multimodal optimization problems with depth, complexity, nonlinear characteristics and high constraints, the Gray Wolf Optimization Algorithm is prone to falling into local optimal solutions, resulting in unsatisfactory optimization results.

[0111] The present invention has achieved the following significant technical advancements:

[0112] Improving the accuracy of loess landslide detection: By designing a grey wolf optimization algorithm HDLMGWO model based on multi-level differentiated learning, the hyperparameters of the segmentation network model UNet-VIG are effectively optimized, which improves the precision and accuracy of loess landslide detection and reduces recognition errors.

[0113] Efficient hyperparameter optimization: The HDLMGWO model optimizes the three hyperparameters of the segmentation network model—learning rate, batch size, and iteration number—through collaborative computing and adaptive adjustment, significantly improving model performance.

[0114] Adjusting the ratio of global search and local development: A nonlinear control parameter based on the hyperbolic tangent function is introduced to adjust the ratio of global search and local development of the Grey Wolf Optimization Algorithm, thereby improving the optimization efficiency.

[0115] Level-differentiated learning mechanism: A level-differentiated learning mechanism is established, and the original position update mechanism and learning mechanism are greedily selected to improve the position update method, so that the algorithm learns towards the global optimum and the neighborhood optimum, thereby improving the global search ability of the algorithm.

[0116] Multi-population recombination evolution strategy: By adopting a multi-population recombination evolution strategy, gray wolves from different subpopulations can select appropriate learning modes based on their fitness levels, thereby improving the adaptability and robustness of the algorithm.

[0117] To prevent the algorithm from getting stuck in local optima in the later stages: A learning operator based on hyperbolic sine function is introduced to prevent the algorithm from getting stuck in local optima in the later stages of iteration, thereby improving the algorithm's global exploration capability.

[0118] Reduced computing and storage resource requirements: By optimizing the algorithm and designing the segmentation network model, the demand for computing and storage resources is reduced, enabling the loess landslide detection system to operate efficiently even in resource-constrained environments, thereby expanding its scope of application.

[0119] Significantly improve the efficiency of loess landslide detection: The present invention improves the overall efficiency of loess landslide detection by optimizing the segmentation network model, making it possible to complete high-precision loess landslide identification and early warning in a shorter time, which helps to take disaster prevention and mitigation measures in a timely manner.

[0120] Third, the technical solution of this invention fills a technological gap in the industry both domestically and internationally:

[0121] This paper presents an improved Grey Wolf optimization algorithm, the first of its kind to improve the algorithm from the perspective of simulating the learning patterns of human society. Four improvement strategies are proposed: nonlinear control parameters based on the hyperbolic tangent function, a hierarchical differential learning mechanism, a multi-population recombination evolution strategy, and a learning operator based on the hyperbolic sine function. Comparison with other enhanced Grey Wolf algorithms shows that the algorithm exhibits better performance. The algorithm is then applied to loess landslide image feature recognition and detection. Given its versatility and robustness, this model can be applied to various other types of landslide applications.

[0122] Fourth, the method provided by this invention first enhances the landslide dataset using a data augmentation generator, employing multiple operations including normalization, image rotation, translation, shearing, and flipping to enrich the dataset's diversity and robustness. The enhanced landslide dataset serves as input, and a carefully designed segmentation network model (such as UNet-VIG) generates an output variable. This establishes a precise mapping between input and output, providing reliable data support for subsequent landslide detection.

[0123] During the optimization process, we introduced the Grey Wolf Optimizer (HDLMGWO) algorithm based on multi-level differentiated learning. This algorithm, through the establishment of a hierarchical differentiated learning mechanism combined with a greedy selection strategy, enables the algorithm to learn towards both the global optimum and the neighborhood optimum, thereby improving its search efficiency and optimization capabilities. Furthermore, by introducing nonlinear control parameters based on the hyperbolic tangent function and a multi-population recombination evolutionary strategy, the algorithm is able to strike a balance between global search and local exploitation, further improving its performance.

[0124] During the optimization process, the HDLMGWO algorithm adaptively adjusts and collaboratively computes three key hyperparameters in the segmentation network model: the learning rate, batch size, and number of iterations. By setting a reasonable solution space, the algorithm can find the optimal hyperparameter combination within a limited range to optimize the performance of the segmentation network model. This adaptive adjustment and collaborative computation significantly improves the accuracy and efficiency of landslide detection.

[0125] This method addresses the inaccurate and inefficient landslide detection challenges of existing technologies in industrial applications. By enhancing data, constructing mapping relationships, and applying the Grey Wolf Optimization algorithm based on multi-level differentiated learning, this method significantly improves the accuracy and efficiency of loess landslide detection. In industrial applications, this method enables rapid and accurate identification of landslide areas, providing strong technical support for landslide prevention and control, and possesses significant practical value. Attached Figure Description

[0126] Figure 1 This is a framework diagram of optimizing the segmentation network for landslide identification using the Grey Wolf Optimization Algorithm based on level-differentiated learning provided by an embodiment of the present invention.

[0127] Figure 2 This is a module diagram of landslide identification using a gray wolf optimization segmentation network based on level-differentiated learning provided by an embodiment of the present invention.

[0128] Figure 3 This is a diagram of the HDLMGWO-UNet-VIG recognition model for a landslide dataset provided by an embodiment of the present invention.

[0129] Figure 4This is a schematic diagram of a landslide dataset example provided in an embodiment of the present invention.

[0130] Figure 5 This is a convergence curve of the HDLMGWO algorithm provided in this embodiment of the invention compared with eight other comparison algorithms on the test function.

[0131] Figure 6 This is a box plot comparing the HDLMGWO algorithm provided in this embodiment of the invention with eight other comparison algorithms.

[0132] Figure 7 This is a framework diagram of the UNet-VIG network provided in an embodiment of the present invention.

[0133] Figure 8 This is a structural diagram of the encoder and decoder of the UNet-VIG network provided in an embodiment of the present invention.

[0134] Figure 9 This is the 1-F1 convergence curve of the HDLMGWO-UNet-VIG network model provided in this embodiment of the invention.

[0135] Figure 10 This is a comparison chart of the F1 score of the HDLMGWO-UNet-VIG model with the GWO-UNet-VIG and UNet-VIG models provided in this embodiment of the invention.

[0136] Figure 11 These are landslide identification images of the UNet-VIG, GWO-UNet-VIG, and HDLMGWO-UNet-VIG models provided in the embodiments of the present invention. DETAILED DESCRIPTION

[0137] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0138] Industrial Application Example 1: Geological Disaster Early Warning System

[0139] Geological disasters such as landslides and debris flows pose a serious threat to human life and property. Timely and accurate landslide detection and early warning are important means of reducing disaster losses. This example describes a loess landslide detection method optimized by the Hierarchical Differentiated Learning Multi-Population Grey Wolf Optimization (HDLMGWO) algorithm and its application in a geological disaster early warning system.

[0140] 1. Data preparation and preprocessing

[0141] The landslide dataset landslide4sense was used to perform data enhancement processing on the landslide images in the dataset, including normalization, image rotation, horizontal translation, vertical translation, shear transformation and horizontal flipping.

[0142] 2. Construction of the segmentation network model

[0143] A UNet-VIG segmentation network model is established, consisting of an encoder, a data exchange module, and a decoder.

[0144] 3. Optimization of the HDLMGWO model

[0145] Design an HDLMGWO model based on multi-level differential learning to initialize the position of individual gray wolves.

[0146] Adjust the ratio of global search to local exploitation in the algorithm, establish a hierarchical differentiated learning mechanism, and optimize the hyperparameters (learning rate, batch size, and number of iterations) in the segmentation network model.

[0147] 4. Landslide Detection

[0148] The hyperparameters optimized using the HDLMGWO model are used to train and optimize the segmentation network model, detect landslide areas in landslide images, and generate landslide early warning information.

[0149] 5. Warning Release

[0150] The detection results and early warning information are integrated into the geological disaster early warning system, and landslide early warning information is released through the early warning platform to promptly notify relevant departments and residents.

[0151] This system can significantly improve the accuracy and efficiency of landslide detection, provide timely warnings of geological disasters, and effectively reduce casualties and property losses caused by landslides.

[0152] Industrial Application Example 2: Mine Safety Monitoring System

[0153] Landslides are a common geological hazard during mining operations, and timely monitoring and detection are crucial for ensuring mine safety. This embodiment introduces the application of a landslide detection method optimized based on HDLMGWO in a mine safety monitoring system.

[0154] 1. Data collection and enhancement

[0155] High-resolution landslide images of the mining area were obtained using drones and ground monitoring equipment.

[0156] Data augmentation methods were used to preprocess the acquired landslide images, including normalization, image rotation, horizontal translation, vertical translation, shearing transformation, and horizontal flipping.

[0157] 2. Construct a segmentation network model

[0158] A UNet-VIG segmentation network model is established, which includes an encoder, a data exchange module and a decoder.

[0159] 3. Optimization of the HDLMGWO model

[0160] The HDLMGWO model is used to optimize the hyperparameters in the segmentation network, initialize the position of individual gray wolves, and introduce nonlinear control parameters based on the hyperbolic tangent function to adjust the ratio of global search to local exploitation in the algorithm.

[0161] A level-differentiated learning mechanism is established to optimize the learning rate, batch size, and iteration times in the segmentation network model.

[0162] 4. Landslide monitoring

[0163] The enhanced landslide image is input into the optimized segmentation network model to detect the landslide area.

[0164] Monitor landslides in the mining area in real time and feed back the detection results to mine safety managers.

[0165] 5. Safety Management

[0166] According to the landslide monitoring results, formulate and adjust the mining plan, take safety protection measures in time, and avoid the occurrence of landslide disasters.

[0167] The system can significantly improve the accuracy and real-time performance of mine landslide monitoring, provide a scientific basis for mine safety management, and effectively ensure safe production in mining areas.

[0168] The two examples above demonstrate the application of a landslide detection method optimized by the Hierarchical Differential Learning-based Multi-Group Gray Wolf Optimization Algorithm (HDLMGWO) in geological disaster early warning and mine safety monitoring. Through data augmentation, segmentation network model construction, and HDLMGWO model optimization, this method achieves high efficiency and high accuracy in landslide detection, and has broad prospects for industrial applications.

[0169] The present invention will now be described in detail with reference to the accompanying drawings:

[0170] Step 1: Use the data enhancement method to perform data enhancement on the training set and test set respectively;

[0171] Step 2: Build the VIG model, design the HDLMGWO model, and use the HDLMGWO model to optimize the hyperparameters of the segmentation network model UNet-VIG.

[0172] Step 3: Set the parameter values ​​of the gray wolf optimization algorithm, initialize the gray wolf population, calculate the fitness value of the gray wolf population, and sort the gray wolf fitness values.

[0173] Step 4: Modify the convergence factor a to a nonlinear control parameter based on the hyperbolic tangent function;

[0174] Step 5: Establish a grade-differentiated learning mechanism, and greedily select the original position update mechanism and the learning mechanism to improve the position update method.

[0175] Step 6: Sort the fitness values ​​of the gray wolf population and divide them into different subpopulations based on their fitness values. Different position update formulas are applied to gray wolves in different subpopulations;

[0176] Step 7: Use a learning operator based on hyperbolic sine function to update the formula for different positions;

[0177] Step 8: Determine if the termination condition is met. If not, return to step 4. If met, proceed to the next step.

[0178] Step 9: Optimize the segmentation network model based on the optimal hyperparameter values ​​searched by the HDLMGWO model to perform landslide identification.

[0179] The landslide HDLMGWO-UNet-VIG recognition model provided by the embodiment of the present invention is shown in FIG. Figure 1 shown.

[0180] The technical solution of the present invention is further described below in conjunction with specific embodiments, which mainly includes the following steps:

[0181] Step 1: Use data enhancement method to enhance the landslide data to expand the data scale and diversity. Figure 4 This study constructed a segmentation network model called UNet-VIG, as shown in Figure 7 As shown, the network consists of three parts: an encoder, a data exchange module, and a decoder. Specifically, both the encoder and decoder adopt the structure of the Unet network and are divided into two stages, each containing two blocks. Each block includes a convolutional block, a batch normalization layer, and a GELU activation function. In each stage of the decoder, the first block includes a transposed convolution with an upsampling rate of 2x. In the data exchange module, a total of 16 VIG blocks are used. In each block, the k edges that are semantically most relevant to the current node are dynamically determined using the k-nearest neighbor algorithm, where k is set to 9.

[0182] Step 2: Design the HDLMGWO model to optimize the hyperparameters of the UNet-VIG segmentation network model, achieving coordinated calculation and adaptive adjustment of the three hyperparameters: learning rate, batch size, and number of iterations. The HDLMGWO algorithm sets the population size to 3 and the maximum number of iterations to 10. Within the algorithm, a solution space is defined for the three hyperparameters of the UNet-VIG network: ub = {0.1, 8, 450} represents the upper bound of the solution space, and lb = {0.001, 1, 350} represents the lower bound. Each dimension of ub and lb corresponds to the upper and lower bounds of the learning rate, batch size, and number of iterations in the UNet-VIG model, respectively.

[0183] Step 3, further, in step three, the HDLMGWO model initializes the gray wolf population. Then, the fitness value of each individual gray wolf is calculated, and they are sorted according to their fitness values. The location of the individual gray wolf is interpreted as a hyperparameter of the segmentation network model UNet-VIG. The UNet-VIG network model is optimized, trained using the training dataset, and its landslide identification performance is evaluated by the F1 score on the test dataset. 1-F1 is used as the fitness value for each gray wolf.

[0184] Step 4: Adjust the convergence factor using nonlinear control parameters based on the hyperbolic tangent function. The formula is as follows:

[0185]

[0186] Step 5: Establish a hierarchical differentiation learning mechanism. A greedy selection of the original position update mechanism and the learning mechanism is used to improve the position update method, enabling the algorithm to learn towards the global optimum and neighborhood optimum. To integrate with the original hierarchy of the gray wolf, the value of P1 is set to 3. It is worth noting that, when set in the original literature, the optimization performance remained stable on most test functions.

[0187] In the learning mechanism, the gray wolf can observe the differences between gray wolves in four different levels, as shown in the following expression:

[0188]

[0189] In this context, Gap1 represents the gap between alpha wolves and elite wolves, Gap2 represents the gap between wolves and bottom wolves, Gap3 represents the gap between elite wolves and bottom wolves, and Gap4 represents the gap between two wolves randomly selected from the population. α represents the globally optimal wolf, elite wolves are the best wolves, and bottom represents the bottom wolves. L1 and L2 are two wolves randomly selected from the pack.

[0190] LF kThe normalized ratio of the Euclidean distance of the k-th difference will affect the learning ability of the i-th individual for the k-th difference. k The larger the value, the more the i-th individual learns from the k-th group's gap. LF k The expression is as follows:

[0191]

[0192] SF i Assess the knowledge that individual gray wolves need to learn, SF i The older you get, the more you need to learn. LR i Let LR represent the learning resistance of the i-th gray wolf. max The maximum learning resistance among the gray wolf individuals. In general, the objective function will accept the i-th individual and return the corresponding output, i.e. LR i The larger the LR i This means that if an individual isn't good enough, they need to learn more. SF i The expression is as follows:

[0193]

[0194] For an individual gray wolf, in terms of its own learning ability (SF) i ) and learning factor (LF k ) under the influence of the gray wolf from the k-th group difference (Gap k ) is the knowledge learned in KA k The formula is shown in (12).

[0195] KA k =SF i ·LF k Gap k (k=1,2,3,4) (5)

[0196] The update formula for the learning mechanism is shown below:

[0197]

[0198] This is the position of the gray wolf in the previous iteration. is the position of the gray wolf after being updated by the learning mechanism.

[0199] The original gray wolf leadership mechanism is:

[0200]

[0201] in, is the updated position after the original gray wolf leadership mechanism. The update formulas of X1, X2 and X3 are as follows:

[0202] X1=|x α -A1·D α |,X2=|X β -A2·D β |,X3=|X δ -A3·D δ | (8)

[0203] Among them, X α 、X β and X δ Let D represent the positions of the gray wolves α, β, and δ, respectively. α 、D β and D δ , the formulas for A1, A2, and A3 are as follows.

[0204]

[0205] A1=2·a·r1-a, A2=2·a·r2-a, A2=2·a·r3-a (10)

[0206] r1, r2, and r3 are random numbers between [0, 1].

[0207] The formulas for C1, C2, and C3 are as follows:

[0208] C1=2r4, C2=2r5, C3=2r6 (11)

[0209] r4, r5 and r6 are random numbers between [0,1].

[0210] Step 6: To avoid the loss of population diversity, we propose a multi-population recombination evolution strategy. The position update formula for the upper-level gray wolves is shown below:

[0211]

[0212] Where rand is a random number between [0,1], and X Avg The formula for the average position of the best gray wolf at the bottom layer and the best gray wolf at the middle layer is as follows:

[0213]

[0214] Among them, MiddleBest is the best gray wolf in the middle layer, BottomBest is the best gray wolf in the bottom layer, and LR is the growth resistance of gray wolves.

[0215] The formula for updating the position of the mid-level gray wolves is as follows:

[0216]

[0217] Among them, X UppeRand is a gray wolf randomly selected from the upper layer, wblrand is a Weibull distribution with shape parameter and scale parameter 1, and rand is a random number between [0,1].

[0218] The formula for updating the position of the bottom gray wolves is as follows:

[0219]

[0220] Among them, rand is a random number between [0,1], BottomBest is the best gray wolf in the bottom layer, is the position of the gray wolf after the multi-population recombination evolution strategy, where the parameters P1 = 3 and P2 = 0.1.

[0221] Step 7: To both enhance the gray wolf's global exploration capabilities and improve the quality of the search solution during the iteration process, this paper creates a learning operator based on the probability P of the hyperbolic sine function. The mathematical expression of this operator is defined as follows:

[0222]

[0223] Among them, rand is a random number between [0,1], Xrand is a random individual selected from the gray wolf, and wblrnd is a random number generated by a Weibull distribution with a scale parameter of 1 and a shape parameter of 1. It is the position of the gray wolf after learning the operator based on the hyperbolic sine function.

[0224] The mathematical expression for P is as follows:

[0225]

[0226] Where sinh is the hyperbolic sine function, t is the current iteration number, and MaxIter is the maximum iteration number.

[0227] Step 8: Determine if the HDLMGWO model has reached the maximum number of iterations. If it has, interpret the optimal position obtained by the algorithm as a hyperparameter of the UNet-VIG model. In this invention, the optimal position of the tunicate is x = {0.03234427}. Here, the learning rate is 0.0323, the batch size is 4, and the number of iterations is 427. If the maximum number of iterations has not been reached, return to step 3 to continue the iterative process.

[0228] Figure 5The following figure shows the convergence curves of the HDLMGWO algorithm provided by an embodiment of the present invention and eight other comparison algorithms on a test function. The X-axis represents the number of iterations, and the Y-axis represents the fitness value. It can be seen that the convergence curve of the HDLMGWO algorithm drops rapidly at the beginning of the iteration, quickly finding the optimal value. Compared with other algorithms, the HDLMGWO algorithm outperforms them in both optimization capability and convergence speed.

[0229] Figure 6 This is a boxplot of the HDLMGWO algorithm provided by an embodiment of the present invention and eight other comparison algorithms. A boxplot is a standardized method for summarizing data distribution based on five digits: minimum, median, maximum, upper quartile, and lower quartile. It allows for the rapid identification of outliers. Therefore, a boxplot can provide a detailed description of the overall stability of an algorithm and present it more intuitively in graphical form. As can be seen from the figure, the HDLMGWO algorithm data is highly concentrated and optimally positioned overall, indicating that its stability and mean are significantly superior to those of the other algorithms.

[0230] In step 8, the obtained hyperparameters are substituted into the UNet-VIG model to build a complete HDLMGWO-UNet-VIG model. The model is then trained using the training set and test set in the dataset, and the trained model is used to predict the test set. Figure 9 The 1-F1 convergence curve for training the HDLMGWO-UNet-VIG model. Figure 10 The F1 of the UNet-VIG model, the GWO-UNet-VIG model established by the standard GWO optimized UNet-VIG network, and the HDLMGWO-UNet-VIG model of the present invention are 68.7%, 69.94%, and 71.89%, respectively, among which the HDLMGWO-UNet-VIG model has the highest F1. Figure 11 The landslide recognition images for the UNet-VIG, UNet-VIG, and HDLMGWO-UNet-VIG models are shown. The results show that the HDLMGWO-UNet-VIG model provided by the present invention has better recognition performance than the UNet-VIG and UNet-VIG models. This indicates that the HDLMGWO-UNet-VIG model proposed in the present invention combines the excellent hyperparameter optimization capabilities of the HDLMGWO algorithm with the advantages of the UNet-VIG network structure, further improving landslide recognition performance and bringing significant significance to landslide prediction.

[0231] The schematic diagram of the landslide dataset sample provided by the embodiment of the present invention is as follows: Figure 4 This study constructed a segmentation network model called UNet-VIG, as shown in Figure 7As shown, the network consists of three parts: an encoder, a data exchange module, and a decoder. The encoder is mainly responsible for downsampling the landslide image and capturing rich semantic information in the process, ensuring that each downsampled pixel represents the semantic content of a specific region in the original image. The data exchange module consists of multiple VIG blocks, which treat each downsampled pixel as a graph node and exchange semantic information. Finally, through the decoder, the semantically exchanged image is restored to the original image size, and the texture and edge information of the image are enhanced.

[0232] Specifically, both the encoder and decoder of this network adopt the structure of the UNet network, which is divided into two stages, each containing two blocks. For example... Figure 8 As shown, each block includes a convolutional block, a batch normalization layer, and a GELU activation function. In each stage of the decoder, the first block includes a transposed convolution with a 2x upsampling rate. In the data exchange module, a total of 16 VIG blocks are used. In each block, the k edges that are semantically most relevant to the current node are dynamically determined using the k-nearest neighbor algorithm, where k is set to 9.

[0233] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes them in conjunction with the data, charts and other information of the experimental process.

[0234] Figure 10 The F1 scores of the UNet-VIG model, the GWO-UNet-VIG model established by the standard GWO optimized UNet-VIG network, and the HDLMGWO-UNet-VIG model in this invention are 68.7%, 69.94%, and 71.89%, respectively, with the HDLMGWO-UNet-VIG model having the highest F1 score. Figure 11 These are landslide recognition images of the UNet-VIG, UNet-VIG, and HDLMGWO-UNet-VIG models. The results show that the HDLMGWO-UNet-VIG model provided by this invention has better recognition performance than the UNet-VIG and UNet-VIG models. It can be seen that the HDLMGWO-UNet-VIG model proposed in this invention combines the excellent hyperparameter optimization capability of the HDLMGWO algorithm with the advantages of the UNet-VIG network structure, further improving the performance of landslide recognition and bringing significant benefits to landslide prediction. Evidence of the relevant effects of the embodiments is provided below. The embodiments of this invention have achieved some positive effects during research and development or use, and indeed have significant advantages compared with existing technologies. The following content describes these advantages using data, charts, etc., from the experimental process.

[0235] In this technology, we applied data augmentation techniques to the landslide dataset. By applying various image processing techniques, we expanded the data size and diversity, improved the model's generalization and robustness, and enhanced its ability to extract key features. This significantly improved the model's performance and practical application, ultimately boosting the accuracy of the recognition task.

[0236] Although segmentation networks demonstrate excellent performance in image recognition tasks, their ultimate recognition results are still affected by hyperparameters such as the number of iterations, batch size, and learning rate. The traditional gray wolf optimization algorithm, a swarm intelligence optimization method, also suffers from the problems of being trapped in local optima and inefficient convergence, especially when dealing with large-scale and multi-peak function optimization problems. Therefore, improvements to this algorithm are crucial.

[0237] The contribution of this technique lies not only in the introduction of the VIG network model but also in the effective data augmentation and innovative improvements to the traditional Grey Wolf Optimization algorithm. This improvement enables adaptive adjustment and optimization of key hyperparameters in the UNet-VIG model—the learning rate, batch size, and number of iterations. This improved algorithm, applied to the landslide identification task, significantly improved recognition accuracy. This technique not only provides important guidance for landslide identification but also offers new ideas and methods for other image recognition tasks, possessing profound theoretical significance and significant practical value.

[0238] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0239] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A loess landslide detection method optimized by multi-population gray wolf optimization algorithm based on hierarchical differentiation learning, It is characterized in that include: Use the data augmentation generator to perform data augmentation on the landslide dataset; A Grey Wolf optimization algorithm model based on multi-level differentiated learning is designed, and the improved algorithm is used to optimize the hyperparameters in the segmentation network model, realizing the adaptive adjustment and collaborative calculation of the three hyperparameters of learning rate, number of iterations, and number of batches in the segmentation network, and finally performing landslide detection.

2. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 1, It is characterized in that The implementation steps are: Step 1: Preprocess the images of the public landslide dataset landslide4sense for semantic segmentation, and use the data enhancer method to enhance the loess landslide image data; Step 2: Establish the segmentation network model UNet-VIG, design the HDLMGWO model based on multi-level differentiated learning, use the HDLMGWO model to optimize the hyperparameters in UNet-VIG, and realize the collaborative calculation and adaptive adjustment of the three hyperparameters of learning rate, batch number and iteration number in the segmentation network; Step 3, initialize the position of the gray wolf individual in the HDLMGWO model; Step 4: Introduce nonlinear control parameters based on hyperbolic tangent function to adjust the ratio of global search and local development of the algorithm; Step 5: Establish a hierarchical differentiation learning mechanism, greedily select the original position update mechanism and the learning mechanism to improve the position update method, so that the algorithm can learn towards the global optimum and the neighborhood optimum; Step 6: Using the multi-population recombination evolution strategy, the gray wolves in different subgroups will adopt the evolution strategy that matches the fitness level of the current subgroup; Step 7, using a learning operator based on a hyperbolic sine function; Step 8: Determine whether the HDLMGWO model has reached the maximum number of iterations. If not, return to step 4. If satisfied, proceed to the next step. Step nine, use the best hyperparameter value searched by the improved grey wolf optimization algorithm to optimize the segmentation network model for loess landslide detection.

3. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that In the step 1, a data enhancement method is used to enhance the loess landslide data; Select normalization, image rotation, horizontal translation, vertical translation, shear transformation, and horizontal flipping as the processing methods of the data enhancer; The enhanced loess landslide dataset is used as the input variable of the segmentation network, and an output variable is generated through the segmentation network, so that a mapping relationship is formed between the input variable and the output variable.

4. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that In the step 2, a segmentation network model UNet-VIG is established, which is composed of an encoder, a data exchange module and a decoder; The HDLMGWO model is designed to optimize the hyperparameters in the segmentation network. The HDLMGWO algorithm is used to optimize the three hyperparameters of learning rate, batch number and iteration number in the segmentation network. The population size in the HDLMGWO algorithm is set to 3 and the maximum number of iterations is set to 10. In the algorithm, the solution space is set for the three hyperparameters of learning rate, batch number and iteration number in the VIG network. ub={0.1,8,450} represents the upper limit of the solution space, and lb={0.001,1,350} represents the lower limit of the solution space. Each dimension of ub and lb corresponds to the upper and lower bounds of the learning rate, batch number and iteration number of the segmentation network model.

5. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that In the step 3, the HDLMGWO model initializes the gray wolf population, then calculates the fitness value of each gray wolf individual, and sorts them according to the fitness value; interprets the position of the gray wolf individual as a hyperparameter of the segmentation network model, optimizes the segmentation network model, trains the network model with the training data set, and predicts the F of loess landslide identification through the test data set. 1 , 1-F 1 As the fitness value of the gray wolf; In step 4, the convergence factor is adjusted using a nonlinear control parameter based on a hyperbolic tangent function. Convergence Factor In the early stages, the Large, larger than the original in later stages, adjusts the ratio of algorithm development and exploration:

6. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that In the step 5, a hierarchical differentiation learning mechanism is proposed, and the original position update mechanism and the learning mechanism are greedily selected to improve the position update method, so that the algorithm learns towards the global optimum and the neighborhood optimum; In the learning mechanism, the gray wolf can observe the differences between the gray wolves in four groups of different levels. The mathematical expression of the difference is: Among them, Gap 1 The gap between alpha wolves and elite wolves. 2 Gap is the gap between alpha wolves and bottom wolves. 3 The gap between the elite wolves and the bottom wolves. 4 is the gap between two randomly selected wolves; α is the global optimal wolf, elites is the elite wolf, bottom represents the bottom wolf, L 1 and L 2 For two wolves selected at random; LF k Normalized ratio of the Euclidean distances of the k-th group differences, LF k The expression is as follows: SF i Assessing the knowledge that gray wolves need to learn, LR i represents the learning resistance of the i-th gray wolf, LR max The greatest learning resistance among gray wolves; SF i The expression is as follows: For a gray wolf individual, in its own learning ability SF i and learning factor LF k Under the influence of Gap, the gray wolf from the kth group k The knowledge learned in KA k : YOU k =SF i ·LF k ·Gap k (k=1, 2, 3, 4) The update formula of the learning mechanism is as follows: Where t is the current iteration number, is the knowledge learned by the gray wolf at the tth iteration, KA k (k=1, 2, 3, 4) is the knowledge learned from the kth group of differences; The original gray wolf's leadership mechanism is: Among them, X 1 , X 2 and X 3 The update formula is as follows: X 1 =|X α -A 1 ·D α |,X 2 =|X β -A 2 ·D β |,X 3 |X δ -A 3 ·D δ | Among them, X α , X β and X δ Denote the positions of α, β and δ gray wolves respectively, D α , D β and D δ , A 1 , A 2 and A 3 The formula is: A 1 =2·a·r 1 -a,A 2 =2·a·r 2 -a,A 2 =2·a·r 3 -a r 1 、r 2 and r 3 are random numbers between [0,1]; among them, C 1 , C 2 and C 3 The formula is as follows: C 1 =2r 4 ,C 2 =2r 5 ,C 3 =2r 6 r 4 、r 5 and r 6 are random numbers between [0,1].

7. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that The step six proposes a multi-population recombination evolution strategy; The position update formula of the upper gray wolf is as follows: Among them, rand is a random number between [0,1], X Avg is the average position of the best gray wolf at the bottom and the best gray wolf at the middle. The formula is as follows: Among them, MiddleBest is the best gray wolf in the middle layer, BottomBest is the best gray wolf in the bottom layer, and LR is the growth resistance of the gray wolf; The position update formula of the middle-level gray wolf is as follows: Among them, X UpperRand is a gray wolf randomly selected from the upper layer, wblrand is a Weibull distribution with shape parameter and scale parameter 1, and rand is a random number between [0,1]; The update formula for the underlying gray wolf is as follows: Among them, rand is a random number between [0,1], BottomBest is the best gray wolf in the bottom layer, and the parameters P1=3, P2=0.

1.

8. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that In step 7, a learning operator based on the probability P of a hyperbolic sine function is created: Among them, rand is a random number between [0,1], X rand An individual randomly selected from the gray wolf, wblrnd is a random number generated by a Weibull distribution with a scale parameter of 1 and a shape parameter of 1. The mathematical expression of P is as follows:

9. The loess landslide detection method optimized by the multi-population gray wolf optimization algorithm based on level-differentiated learning as claimed in claim 2, It is characterized in that In the step 8, according to the maximum number of iterations set by the HDLMGWO model, it is determined whether the current individual meets the termination condition. If the termination condition is met, the algorithm result is output. If the final iteration condition is not met, it returns to step 4 and continues to cycle into the next generation until the condition is met. In the step nine, the optimal position finally found by the gray wolf is interpreted as the hyperparameter of the segmentation network model, the hyperparameter is input into the segmentation network model, a complete segmentation network model is constructed, the network model is trained using the training set and the test set in the landslide data set, and the trained model is predicted to improve the accuracy of landslide identification.

10. A loess landslide detection system optimized by a multi-population grey wolf optimization algorithm based on level-differentiated learning according to the method of claim 1, It is characterized in that include: Data preprocessing module: used to preprocess the images of the public loess landslide dataset Landslide4Sense for semantic segmentation, and use the data enhancer method to enhance the landslide image data; Segmentation network model module: establish the segmentation network model UNet-VIG for semantic segmentation of landslide images; HDLMGWO model optimization module: Design the HDLMGWO model based on multi-level differentiated learning, which is used to optimize the hyperparameters in the segmentation network model UNet-VIG; Iteration judgment module: judges whether the HDLMGWO model has reached the maximum number of iterations. If not, it returns to continue iteration. If it is satisfied, it proceeds to the next step. Optimization execution module: Optimize the segmentation network model UNet-VIG for loess landslide detection using the best hyperparameter values ​​searched by the improved Grey Wolf Optimization Algorithm.

Images