Landslide image recognition and early warning method based on image recognition

Abstract

The present application relates to the technical field of geological disaster monitoring and early warning, and specifically discloses a landslide image recognition and early warning method based on image recognition. The method comprises: acquiring terrain and remote sensing image data of a target area; simulating slope deformation under different rainfall conditions through rock-soil mechanics finite element analysis to generate physically reasonable landslide visualization images; fusing real and virtual landslide images to construct a hybrid training set; training a conditional generative adversarial network to generate enhanced images that conform to mechanical laws; deploying the model and embedding a lightweight mechanics constraint module to visually and mechanically double-discriminate new images; triggering an early warning when visual feature matching and mechanical conditions are simultaneously met, and supporting dynamic tracking and model incremental updating. The present application combines mechanism driving and data driving to improve recognition robustness and early warning accuracy, reduce false positive rate, and achieve high-precision and strongly-explained landslide intelligent early warning.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster monitoring and early warning technology, specifically relating to a landslide image recognition and early warning method based on image recognition. Background Technology

[0002] With the deep integration of remote sensing monitoring and computer vision technologies, automated identification of geological hazards has become a key means to safeguard national security and major engineering projects. As one of the most destructive natural disasters, the accurate identification and real-time early warning of landslides are of great strategic significance for disaster prevention and mitigation. Traditional monitoring systems are gradually shifting from reliance on manual inspections to automated image analysis based on deep learning. Through real-time processing of high-definition UAV aerial survey images and high-resolution satellite remote sensing data, risk screening and dynamic monitoring of a wide-area geological environment can be achieved.

[0003] Landslide monitoring technology based on image recognition aims to use deep neural networks to extract morphological, textural, and color features from topography to pinpoint potential landslide hazard points on a macro scale. This type of technology attempts to establish a correlation between image representation and disaster categories through end-to-end feature mapping, thereby improving the coverage and response speed of disaster investigation.

[0004] The evolution of landslide disasters is essentially governed by the laws of geotechnical mechanics, involving the dynamic coupling of multiple physical fields such as slope strain, shear failure, and seepage pressure. How to deeply integrate the feature extraction capabilities of the visual domain with the mechanical mechanisms of geotechnical engineering to construct an intelligent recognition framework with physical and logical self-consistency has become a core requirement for improving the reliability of landslide early warning systems.

[0005] Traditional computer vision models rely heavily on large-scale labeled samples, while real extreme landslide cases are scarce and exhibit long-tailed distribution characteristics, resulting in weak generalization performance of trained models in complex environments.

[0006] Existing algorithms primarily focus on the visual characteristics of landslide cover, lacking an understanding of the intrinsic stress state and deformation mechanisms of landslides. This leads to high rates of visual misjudgment in non-landslide areas such as artificially excavated slopes or exposed construction sites, increasing the system's false alarm frequency. Furthermore, image recognition processes often operate without the support of physical simulations, making it difficult to capture the intrinsic connections between the evolution of micro-cracks and nonlinear displacements in landslide precursors. This results in early warning results lacking in-depth interpretation and mechanical support. Therefore, a landslide image recognition and early warning method based on image recognition is desired. Summary of the Invention

[0007] The purpose of this invention is to provide a landslide image recognition and early warning method based on image recognition, which can effectively solve the problems in the background art mentioned above.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is: a landslide image recognition and early warning method based on image recognition, comprising the following specific steps: Step 1: Acquire topographic data and remote sensing image data of the target area. The topographic data includes digital elevation model and geological parameter information, and the remote sensing image data includes UAV aerial images or satellite remote sensing images. Step 2: Based on the terrain data, call the finite element analysis kernel of geotechnical mechanics to simulate the evolution process of stress field and deformation field of slope under different rainfall conditions, and generate corresponding physical reasonable landslide deformation visualization images; Step 3: Fuse the physically reasonable landslide deformation visualization image with the real landslide remote sensing image to construct a hybrid training dataset containing real samples and virtual samples; Step 4: Based on the hybrid training dataset, train a conditional generative adversarial network model, which can generate enhanced images that conform to the mechanical evolution law of landslides based on the input images; Step 5: Deploy the conditional generative adversarial network model into the landslide identification system. After receiving new remote sensing images, the system uses the embedded lightweight mechanical constraint module to reverse-engineer the surface deformation features in the images to determine whether they meet the mechanical conditions for landslide occurrence. Step 6: If the image simultaneously meets the visual feature matching condition and the mechanical condition, it is determined to be a landslide hazard area and an early warning signal is triggered.

[0009] Preferably, the terrain data in step 1 is acquired through lidar scanning, stereo photogrammetry, or downloading from a public geographic information system platform. The geological parameter information includes soil and rock density, internal friction angle, cohesion, and permeability coefficient. The spatial resolution of the remote sensing image data is not lower than a predetermined threshold and covers the entire slope area of ​​the target region.

[0010] Preferably, in step 2, the finite element analysis kernel of geotechnical mechanics adopts an implicit solution algorithm, which determines whether the slope element has undergone shear failure based on the Mohr-Coulomb strength criterion, and combines the saturated-unsaturated seepage theory to simulate the influence of rainfall infiltration on pore water pressure. Finally, the crack distribution map, displacement vector field and plastic strain cloud map are output as the basic data source for visualizing physically reasonable landslide deformation images.

[0011] Preferably, in step 3, the real landslide remote sensing images need to be manually annotated to mark the landslide boundaries, the location of the main cracks and the sliding direction, while the virtual landslide images are automatically generated based on the finite element simulation results to form the corresponding visual representation. The two are mixed in a specific ratio to form a training set, with the virtual samples accounting for no less than half of the total sample size, in order to alleviate the long-tail distribution problem caused by the scarcity of real samples.

[0012] Preferably, the conditional generative adversarial network model in step 4 includes a generator and a discriminator. The generator receives the original remote sensing image and the corresponding mechanical field label as input and outputs an enhanced landslide feature image. The discriminator simultaneously evaluates the realism and mechanical consistency of the image. During the training process, perceptual loss and structural similarity loss are introduced to improve the detail fidelity of the generated image.

[0013] Preferably, the lightweight mechanical constraint module in step 5 is an embedded neural network substructure. Its input is the deformation feature map extracted from the image to be identified, and its output is the probability value of whether the deformation conforms to the typical mechanical evolution path of a landslide. During the training phase, the lightweight mechanical constraint module learns synchronously with the main recognition network through joint optimization, ensuring that only a small amount of computational overhead is added during the inference phase.

[0014] Preferably, in step 6, the visual feature matching condition is determined by a convolutional neural network classifier. This convolutional neural network classifier has learned the differences in texture and shape between landslide and non-landslide areas during the training phase. The mechanical condition is determined by whether the probability value output by the lightweight mechanical constraint module exceeds a preset threshold. The warning is triggered only when both conditions are met simultaneously, thus reducing the false alarm rate.

[0015] Preferably, the present invention also includes dynamic tracking and monitoring of the warning area, calculating the deformation rate change trend by inputting multiple consecutive remote sensing images, and if the deformation rate shows an accelerating growth characteristic, the warning level is raised and automatically pushed to the disaster emergency command platform.

[0016] Preferably, the hybrid training dataset is incrementally updated after each new real landslide case is added, and the finite element simulation is rerun to generate virtual samples that match the geological conditions of the new case, ensuring that the model continuously adapts to the landslide identification needs of different regions and different lithologies.

[0017] Preferably, the geotechnical mechanics finite element analysis kernel and the conditional generative adversarial network model used in this invention are both deployed on edge computing devices or cloud servers, supporting batch processing of large-scale mountainous areas. The area of ​​a single processing session can reach hundreds of square kilometers, and the processing time meets the hourly early warning response requirements.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a landslide recognition framework with physical and logical self-consistency by integrating geotechnical mechanics mechanisms and generative visual models, which solves the problem of insufficient model generalization ability caused by the scarcity of real landslide samples; it expands the training set by using virtual but physically real landslide images generated by mechanical simulation, thereby improving the model's recognition robustness in extreme or rare landslide scenarios.

[0019] 2. This invention introduces a lightweight mechanical constraint module in the identification stage, enabling the algorithm not only to identify visually similar landslide areas, but also to distinguish whether their internal mechanical state conforms to the landslide evolution law, reducing the false alarm rate of misjudging artificially excavated slopes or exposed construction sites as landslides; making the landslide early warning results have high accuracy, low false alarm and strong interpretability, providing reliable technical support for intelligent prevention and control of geological disasters. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework for enhancing landslide features in this invention; Figure 3 This is a flowchart illustrating the logical process of constructing the hybrid training dataset in this invention. Figure 4 This is a flowchart illustrating the logical process of reverse inference of surface deformation features from remote sensing images in this invention. Figure 5 This is a flowchart of the logic process for landslide identification and early warning triggering in this invention. Detailed Implementation

[0021] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0022] In the landslide image recognition and early warning method based on image recognition provided in this invention, its core execution logic is built upon the deep coupling of geotechnical mechanics mechanisms and a generative visual deep learning model. This method solves the core technical problems of sample scarcity and inaccurate identification in geological disaster recognition by introducing a physical information enhancement mechanism.

[0023] In the above method, step 1 involves acquiring terrain data and remote sensing image data of the target area. Specifically, the acquisition of terrain data involves high-precision geospatial sampling. The data acquisition methods include, but are not limited to, using a lidar scanning system mounted on an unmanned aerial vehicle (UAV) for active 3D measurement. By acquiring massive amounts of lidar point cloud data, and after point cloud denoising, ground point classification, and resampling processing, a digital elevation model is constructed.

[0024] Using stereo photogrammetry, aerial triangulation and density matching are performed on UAV optical images with a specific overlap rate to generate a high-resolution digital surface model. Based on this, vegetation and man-made structures are removed to obtain a digital elevation model reflecting the true topographic relief. Geological parameters include the physical and mechanical properties of the soil and rock mass, covering parameters such as density, effective internal friction angle, cohesion, permeability coefficient, and compression modulus.

[0025] These parameters were obtained through in-situ drilling and sampling in the target area, followed by laboratory geotechnical tests, or by referring to a detailed geological survey report of the target area. Remote sensing imagery data included UAV aerial images or satellite remote sensing images. To ensure identification accuracy, the spatial resolution of the remote sensing imagery data was set to a predetermined threshold of no less than 0.2 meters to 0.5 meters, ensuring that the subtle topological features of the slope cracks could be clearly captured. The acquired images needed to cover the entire slope area of ​​the target region, including the tension crack zone at the rear edge of the landslide, the deformation monitoring zone in the middle of the slope, and the accumulation and compression zone at the toe of the slope.

[0026] In the above method, step 2 involves using a geotechnical finite element analysis kernel based on the terrain data to simulate the evolution of the stress and deformation fields of the slope under different rainfall conditions, and generating a corresponding physically reasonable landslide deformation visualization image. The geotechnical finite element analysis kernel employs an implicit solution algorithm to ensure convergence when dealing with large deformation and nonlinear contact problems. The solution process first converts the digital elevation model obtained in step 1 into a three-dimensional finite element mesh model, with the mesh element type selected as tetrahedral or hexahedral elements according to accuracy requirements.

[0027] In terms of boundary condition settings, the bottom of the model is set as a fixed constraint, the lateral side as a normal constraint, and the slope surface as a free boundary and seepage boundary. When simulating rainfall conditions, the rainfall intensity is converted into the infiltration flow boundary condition of the slope surface, based on the theory of saturated to unsaturated seepage, to simulate the increase in pore water pressure in the slope caused by rainfall infiltration in real time. The Mohr-Coulomb strength criterion is used to determine whether shear failure occurs in the internal units of the slope. The specific logic is as follows: when the shear stress inside the unit reaches the critical strength value determined by cohesion, normal stress, and internal friction angle, the internal unit of the slope is determined to have entered the plastic state.

[0028] The finite element kernel outputs the displacement vector field, the first principal stress field, the third principal stress field, and the equivalent plastic strain contour map for each step. To generate physically plausible landslide deformation visualization images, the system establishes a texture mapping function to map the three-dimensional displacement field calculated by the finite element method onto a two-dimensional image space. A spatial offset algorithm is then used to simulate the visual effects of tensile cracks, uplift folds, and vegetation tilting and peeling caused by slope sliding onto the original terrain texture. These visualization images not only possess visual realism but also strictly adhere to the mechanical logic of landslide occurrence in their spatial distribution.

[0029] In the above method, step 3 involves fusing the physically reasonable landslide deformation visualization image with real landslide remote sensing images to construct a hybrid training dataset containing both real and virtual samples. Specifically, the real landslide remote sensing images are manually annotated by professional geological engineers before being included in the dataset. The annotations include high-precision pixel-level landslide boundary masks, main crack orientation, secondary crack distribution, and estimated sliding direction arrows. The virtual landslide images are automatically generated with corresponding visual representations and labels based on finite element simulation results. This generation method eliminates the subjective bias of manual annotation.

[0030] To address the long-tailed distribution of landslide samples during dataset construction, a specific mixing ratio strategy was adopted, with virtual samples accounting for no less than half of the total sample size. Through physical simulation, numerous landslide scenarios under extreme rainfall conditions or different slope angles, which are difficult to collect in reality, were generated, enriching the diversity of the samples. During the fusion process, radiometric calibration, atmospheric correction, and tone consistency processing were performed on the images to ensure that the virtually generated feature textures and the underlying background of the real images have good overlap in visual feature space.

[0031] In the above method, step 4 involves training a conditional generative adversarial network (GAN) model based on the hybrid training dataset. This GAN model can generate enhanced images that conform to the mechanical evolution laws of landslides from the input images. The GAN model comprises two core components: a generator and a discriminator. The generator receives the original remote sensing image and mechanical field labels provided by finite element simulation as input. The mechanical field labels are represented in the form of multi-channel images, each representing mechanical characteristics such as deformation rate and stress concentration in different regions.

[0032] The generator's internal architecture employs an encoder-decoder structure with skip connections. It extracts semantic information from the image through deep convolutional operations and fuses it with mechanical field constraints to output an enhanced landslide feature image. The discriminator is designed as a multi-scale classification network that not only evaluates the realism of the generated image at the pixel level but also assesses whether the deformation texture presented by the image has mechanical consistency with the input mechanical field label.

[0033] During training, the loss function was constructed using a comprehensive weighted approach, including adversarial loss, perceptual loss, and structural similarity loss. Perceptual loss extracts high-level feature differences from the pre-trained convolutional neural network to ensure that the generated image semantically matches the features of the real landslide; structural similarity loss maintains the sharpness and continuity of the deformed region's edges. Through adversarial interaction between the generator and discriminator, the model ultimately gains the ability to extrapolate the inherent mechanical instability from a single remote sensing image.

[0034] In the above method, step 5 involves deploying the conditional generative adversarial network model into the landslide identification system. Upon receiving new remote sensing images, the system uses an embedded lightweight mechanical constraint module to reverse-engineer the surface deformation features in the images to determine whether they meet the mechanical conditions for landslide occurrence.

[0035] The lightweight mechanical constraint module, as an embedded neural network substructure, mainly consists of lightweight convolutional kernels and global average pooling layers. Its input is a deformation feature map generated after the image to be identified passes through the feature extraction layer. This deformation feature map contains information on surface texture anomalies, tonal variations, and geometric distortions.

[0036] The output of the lightweight mechanical constraint module is a continuously distributed probability value, which represents the likelihood that the surface deformation reflected by the current image features conforms to the typical mechanical evolution path of a landslide. During the training phase, this lightweight mechanical constraint module learns synchronously with the main recognition part of the generative adversarial network through a joint optimization strategy, establishing a mapping model from visual deformation patterns to mechanical instability states.

[0037] During the reasoning phase, the lightweight mechanical constraint module analyzes the intersection of cracks, the centroid of the deformation zone distribution, and the directionality of surface displacement to inversely deduce whether these features are caused by the internal shear stress of the slope exceeding the shear strength, rather than false positive features caused by artificial excavation or surface shadow interference.

[0038] In the above method, in step 6, if the image simultaneously meets both visual feature matching conditions and mechanical conditions, it is determined to be a landslide hazard area, and an early warning signal is triggered. The visual feature matching conditions are determined by a trained deep convolutional neural network classifier, which is specifically designed to distinguish landslides from various non-landslide areas, such as artificially excavated building slopes, exposed industrial sites, agricultural terraces, and cloud shadows. The classifier outputs a visual landslide confidence score.

[0039] The mechanical conditions are determined by the probability values ​​output by the lightweight mechanical constraint module. The logic for early warning determination is as follows: the system only determines that there is a real landslide hazard in the area when the confidence score at the visual level exceeds the first preset threshold and the probability value at the mechanical level exceeds the second preset threshold. This dual constraint mechanism reduces the false alarm rate. Once the determination is made, the system automatically extracts the geographical coordinates, deformation area, and estimated impact range of the hazard area, and triggers different levels of early warning signals based on the severity of the identified deformation, which are then pushed to the disaster management department in real time via wireless network.

[0040] In further implementation of the above method, to address the complex and variable geological conditions, the hybrid training dataset is incrementally updated after each new real landslide case is added. Whenever a new landslide sample is captured in actual monitoring, the system automatically analyzes the topographic and geomechanical parameters of the case location, reruns the finite element simulation program, and generates a virtual enhanced sample that highly matches its geological background. This self-evolving mechanism ensures that the model can continuously adapt to the landslide identification needs of different regions and different lithologies.

[0041] In terms of processing performance, the geotechnical mechanics finite element analysis kernel and conditional generative adversarial network model used in this method are both deployed on high-performance edge computing devices or high-concurrency cloud server clusters, supporting streaming batch processing of remote sensing data in wide-area mountainous areas. The area processed in a single session can reach hundreds of square kilometers, and the processing time meets the requirements for rapid early warning response at the hourly or even minute level.

[0042] Example 2: Based on Example 1, this example further refines and expands the dynamic early warning of landslides under complex rainfall duration conditions, aiming to improve the system's ability to capture the evolution trend of landslides during rainfall.

[0043] In Example 2, the geotechnical finite element simulation process in step 2 is refined into a non-steady-state analysis with multiple time steps. Specifically, the rainfall conditions are set as a time-varying function sequence, including a long-duration, low-intensity rainfall process in the early stage and a short-duration, high-intensity downpour process in the later stage. The finite element kernel dynamically adjusts the shear strength parameters of the soil and rock mass according to the real-time simulated changes in soil moisture content. For example, as the moisture content increases, the cohesion decreases by a certain reduction ratio, while the internal friction angle is corrected according to the principle of unsaturated effective stress. The finite element simulation not only outputs the final failure mode but also outputs evolution sequence images at different time points. These sequence images show how tensile cracks gradually develop downwards from the top of the slope and how the plastic zone connects to form a continuous sliding surface.

[0044] In step 3, a time-series dimension is introduced when constructing the hybrid training dataset. The virtual samples are no longer single static images, but rather evolutionary sequence frames composed of multiple temporal phases. These frame sequences are matched with multiple real remote sensing images acquired through continuous monitoring to train a conditional generative adversarial network capable of capturing spatiotemporal features.

[0045] In step 4, the generator of the conditional generative adversarial network model further integrates a long short-term memory (LSTM) network structure. This LSM network structure can receive multiple consecutive remote sensing images as input and utilizes a recurrent neural network mechanism to extract the spatiotemporal correlation between images. When generating enhanced images, the generator considers not only the features of the current image but also the mechanical evolution state of the previous phase. The discriminator adds a temporal consistency scoring branch to judge the rationality of the evolution of the enhanced image sequence in the time dimension, such as whether the increase in deformation displacement conforms to the mechanical acceleration law.

[0046] In step 5, a dynamic tracking and monitoring step for the warning area is added. The system calculates the displacement vector changes of feature points within the potential hazard area using multiple consecutive remote sensing image inputs, and further derives the deformation rate trend. The calculation logic is as follows: by comparing the image displacement amounts over three consecutive observation periods, the derivative of the deformation rate, i.e., the deformation acceleration, is calculated. If the deformation rate exhibits a continuous accelerating growth characteristic, and the acceleration value exceeds the set critical warning value, the system automatically upgrades the warning level from the ordinary attention level to the orange or red critical warning level.

[0047] In step 6, the triggering logic for the early warning signal incorporates measured rainfall data. The judgment threshold of the lightweight mechanical constraint module is dynamically adjusted based on the current real-time rainfall intensity. When the measured rainfall exceeds the historical landslide induction threshold for the region, the system automatically lowers the sensitivity threshold for mechanical condition judgment, enabling the model to have a stronger risk identification and perception capability under extreme weather conditions. The judgment results are automatically pushed to the disaster emergency command platform via BeiDou satellite short message or 5G network, and the precise three-dimensional location of the hazard point is highlighted in real time on the electronic map.

[0048] Example 3: This example further explores how to use feedback data from ground monitoring equipment to correct and enhance the accuracy of image recognition in the context of multi-source data fusion.

[0049] In Example 3, in addition to remote sensing imagery, the data acquired in step 1 also includes measured physical monitoring data obtained through flexible displacement gauges, deep inclinometers, and pore water pressure gauges deployed on the slope surface. This data is wirelessly transmitted to the backend database via an IoT terminal.

[0050] In the simulation process of step 2, an inverse analysis optimization algorithm was employed. The measured displacement data and pore water pressure data obtained in step 1 were used as the constraint boundaries or verification indicators for the finite element analysis. After the initial simulation, the finite element kernel compared the calculated deformation field with the measured displacement at the points. If the absolute value of the error between the two was greater than a preset deviation value, the geotechnical mechanical parameters in the finite element model, such as the value of cohesion or the order of magnitude of the permeability coefficient, were automatically adjusted.

[0051] Through multiple iterative calculations, the deformation field output by the finite element model is made to match the actual physical state on site to the highest degree. The resulting physically reasonable landslide deformation visualization image is more targeted and can accurately reproduce the instability mode of a specific slope under specific working conditions.

[0052] In step 4, during model training, a multi-dimensional vector interface is added to the input of the conditional generative adversarial network (GAN) to accept current measured ground physical quantities. When constructing enhanced images, the generator uses these measured physical quantities as prior conditions to enhance local features in specific regions of the image. For example, when a pore water pressure gauge reading suddenly increases in a certain region, the model assigns a higher deformation weight to that specific region when generating the enhanced image and strengthens the darkening features caused by water saturation in the image.

[0053] In step 5, the lightweight mechanical constraint module is designed as a multimodal fusion network. This multimodal fusion network not only extracts spatial convolutional features from the imagery but also processes the measured physical monitoring time series through parallel fully connected subnetworks. The features from the two branches are vector-concatenated at the fusion layer before the final mechanical rationality evaluation is performed. In this way, the image recognition system no longer observes the surface appearance in isolation but combines the real physical feedback from the slope interior, achieving comprehensive monitoring and analysis of the potential hazard area.

[0054] In step 6, multi-source cross-verification logic is incorporated into the triggering of the warning signal. The system performs a weighted summation of the warning results obtained from image recognition and the warning results predicted based on physical monitoring equipment. The weighting scheme is dynamically adjusted according to current meteorological conditions and data reliability: during the day when visibility is good, the weight of image recognition results is increased; during heavy rain or when the remote sensing imaging quality deteriorates at night, the weight of ground monitoring equipment results is increased. This complementary logical architecture ensures the high reliability of the warning system under all-weather conditions.

[0055] Example 4: This example focuses on describing a rapid landslide image processing and adaptive model optimization scheme based on a distributed computing architecture at a large regional scale.

[0056] For the task of landslide survey in large-scale mountainous areas, the amount of remote sensing imagery data acquired in step 1 is enormous, typically reaching the terabyte level. The system adopts a block storage mechanism based on a distributed file system, dividing the large-scale imagery into several fixed-size geographic grid blocks. Each block is associated with corresponding terrain data and geological zoning parameters through metadata indexing.

[0057] In step 2, to improve the efficiency of large-scale finite element simulation, the system employs a parallel computing cluster. For tens of thousands of potential slopes in different geological zones, the computing cluster calls the finite element analysis kernel in parallel, utilizing a massive number of computing nodes to synchronously run stability simulations under different working conditions. To reduce redundant calculations, a pre-calculated library of physically reasonable deformation models was established. For slopes with geometric features similar to geological parameters, models are directly retrieved from the library and fine-tuned to generate visualization images.

[0058] In step 3, when constructing the hybrid training dataset, an active learning mechanism is introduced. After the initial identification, the system filters out samples whose classification confidence levels fall within the fuzzy range (e.g., at the edge of the judgment threshold). These problematic samples are then fed back to geological experts for manual verification, or the finite element kernel is guided to perform higher-precision encrypted simulations on specific parameter ranges for these samples, generating more discriminative virtual training data. This targeted dataset expansion method enables the model to quickly overcome specific scenarios with high recognition difficulty.

[0059] In steps 4 and 5, the conditional generative adversarial network model employs model pruning and quantization techniques to adapt to the operating environment of edge computing devices. Specifically, by analyzing the contribution of neuron connections, redundant parameter branches are eliminated, and floating-point operations are converted into fixed-point operations. The optimized lightweight mechanical constraint module can be directly embedded into UAV terminals or embedded smart cameras deployed in forest areas to achieve real-time front-end preprocessing and hazard screening, uploading only suspicious image slices to the cloud for in-depth verification.

[0060] In the early warning distribution phase of step 6, the system establishes a multi-level distributed early warning push system. The first level is an on-site real-time voice alarm, which is directly triggered by cameras with recognition capabilities; the second level is a mobile SMS alarm for nearby residents; and the third level is a detailed technical report for government disaster prevention and mitigation departments, which includes detailed mechanical assessment results of landslide hazards, the distribution of affected people, and recommended evacuation routes.

[0061] The hybrid training dataset is incrementally updated after each new real landslide case is added. The system employs an online learning algorithm, enabling the model to fine-tune using newly acquired data streams without retraining the entire network. In this way, the model can continuously accumulate knowledge of landslide evolution patterns in specific geographical areas over time.

[0062] The operation flow of the method of this invention is as follows: First, the system automatically triggers a data acquisition task, obtains rainfall forecasts from the meteorological early warning center, and determines key monitoring grids based on rainfall intensity. It then calls upon the terrain database and finite element analysis module stored in the cloud to perform mechanical stability simulations on all high and steep slopes within the key monitoring grids, generating a virtual image set containing displacement cloud maps and crack distribution maps. These images are used in real-time for fine-tuning the training set, putting the recognition model into a state of readiness.

[0063] Once satellites or drones acquire the latest live imagery, the system immediately initiates enhancement processing based on conditional generative adversarial networks. The enhanced imagery is then inspected using a lightweight mechanical constraint module. If a region exhibits significant deformation texture, and this texture closely matches the tensile failure mode simulated by finite element analysis in its spatial arrangement and expansion direction, the system will perform final confirmation using a convolutional neural network classifier. Once both visual and mechanical indicators exceed safety thresholds, the system immediately calculates the deformation area and potential slide direction, generating a detailed early warning information package.

[0064] The early warning information package is transmitted to the disaster emergency command platform via a high-speed fiber optic network or satellite link. The command platform automatically retrieves surrounding emergency resource data based on the early warning level. For orange and higher-level warnings, the system automatically generates a disaster impact map and calculates the number of buildings and population density within the threatened area. Throughout the process, the introduction of physical constraints based on geotechnical mechanics effectively eliminates visual false alarms caused by surface construction, large machinery movement, or natural vegetation decay, thus improving the scientific rigor and authority of the early warning system.

[0065] This invention also supports the evaluation of the effectiveness of landslide mitigation projects. After slope reinforcement, simply updating the support parameters in the finite element model generates a new predicted image under mechanical equilibrium. By comparing the image recognition results before and after mitigation, the contribution of the mitigation project to improving slope stability can be intuitively analyzed, providing technical support for closed-loop management of geological disasters.

[0066] This invention is not limited to landslide identification. For geological disasters with clear mechanical evolution mechanisms, such as collapses and debris flows, similar physical information enhancement identification can also be achieved by adjusting the constitutive model of the finite element kernel and the feature extraction layer of the identification model.

[0067] In summary, this invention, by embedding the essence of geotechnical mechanics into a deep learning framework, endows image recognition algorithms with physical thinking, enabling them to see through the mechanical essence beneath the surface of images like an expert, and accurately capture subtle signs of geological disasters in extremely complex natural environments. This deeply integrated methodology provides a novel technical path to solve the persistent problems of high cost, low efficiency, and high false alarm rates in large-scale geological disaster monitoring, possessing high engineering application value and social disaster reduction benefits.

[0068] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A landslide image recognition and early warning method based on image recognition, characterized in that, Includes the following steps: Step 1: Acquire topographic data and remote sensing image data of the target area. The topographic data includes digital elevation model and geological parameter information, and the remote sensing image data includes UAV aerial images or satellite remote sensing images. Step 2: Based on the terrain data, call the finite element analysis kernel of geotechnical mechanics to simulate the evolution process of stress field and deformation field of slope under different rainfall conditions, and generate corresponding physical reasonable landslide deformation visualization images; Step 3: Fuse the physically reasonable landslide deformation visualization image with the real landslide remote sensing image to construct a hybrid training dataset containing real samples and virtual samples; Step 4: Based on the hybrid training dataset, train a conditional generative adversarial network model, which generates an enhanced image that conforms to the mechanical evolution law of landslides based on the input image; Step 5: Deploy the conditional generative adversarial network model into the landslide identification system. After receiving new remote sensing images, the system uses the embedded lightweight mechanical constraint module to reverse-engineer the surface deformation features in the images to determine whether they meet the mechanical conditions for landslide occurrence. Step 6: If the image simultaneously meets the visual feature matching condition and the mechanical condition, it is determined to be a landslide hazard area and an early warning signal is triggered.

2. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, Step 1 specifically includes: using the lidar scanning system carried by the UAV to perform active three-dimensional measurement to obtain lidar point cloud data, and performing noise reduction, ground point classification and resampling processing on the lidar point cloud data to construct the digital elevation model; Using stereo photogrammetry, aerial triangulation and dense matching are performed on UAV optical images with a predetermined overlap rate to generate a high-resolution digital surface model. Vegetation and artificial buildings are removed from the digital surface model to obtain a digital elevation model that reflects the real terrain undulations. The geological parameter information is obtained by in-situ drilling and sampling in the target area and conducting indoor geotechnical tests, or by referring to the detailed geological survey report of the target area. The geological parameter information includes the density of the soil and rock mass, the internal friction angle, the cohesion, the permeability coefficient, and the compression modulus. The spatial resolution of the remote sensing image data is higher than a predetermined resolution threshold, and the remote sensing image data covers the complete slope area of ​​the target area. The complete slope area includes the tension crack zone at the rear edge of the landslide, the deformation monitoring zone in the middle of the slope, and the accumulation and compression zone at the toe of the slope.

3. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, Step 2 specifically includes: the geotechnical mechanics finite element analysis kernel adopts an implicit solution algorithm, first converting the digital elevation model into a three-dimensional finite element mesh model, and selecting tetrahedral mesh elements or hexahedral mesh elements according to the accuracy requirements; In terms of boundary condition settings, the bottom of the three-dimensional finite element mesh model is set as a fixed constraint, the lateral side is set as a normal constraint, and the slope is set as a free boundary and a seepage boundary. Combining the theory of saturated to unsaturated seepage, rainfall intensity is transformed into the infiltration flow boundary condition of the slope to simulate the process of increased pore water pressure in the slope caused by rainfall infiltration. Based on the Mohr-Coulomb strength criterion, it is determined whether the internal unit of the slope has undergone shear failure. When the shear stress inside the unit reaches the critical strength value determined by the cohesion, normal stress and internal friction angle, the unit is determined to have entered the plastic state. The finite element analysis kernel of geotechnical mechanics outputs the displacement vector field, the first principal stress field, the third principal stress field, and the equivalent plastic strain cloud map for each calculation step, which serve as the basis for generating the physically reasonable landslide deformation visualization image.

4. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, The logic for generating the corresponding physically reasonable landslide deformation visualization image in step 2 is as follows: establish a texture mapping function to map the three-dimensional displacement field calculated by the finite element analysis kernel of the soil and rock mechanics to the two-dimensional image space; By using a spatial offset algorithm, the visual effects of tensile cracks, uplift folds, and vegetation tilting and peeling caused by slope sliding are simulated on the original terrain texture, so that the generated physically reasonable landslide deformation visualization image follows the mechanical logic of landslide occurrence in spatial distribution.

5. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, Step 3 specifically includes: manually annotating the real landslide remote sensing image, the annotation content including pixel-level landslide boundary mask, main crack direction, branch crack distribution and estimated sliding direction; The physically reasonable landslide deformation visualization image is used as a virtual sample, and the corresponding visual representation and label are automatically generated based on the simulation results of the finite element analysis kernel of the soil and rock mechanics. When constructing the hybrid training dataset, the proportion of the virtual sample in the hybrid training dataset is set to be greater than or equal to half of the total number of samples; During the fusion process, the images in the hybrid training dataset are subjected to radiometric calibration, atmospheric correction, and tone consistency processing to ensure that the feature textures of the virtual samples and the underlying background of the real landslide remote sensing images have a degree of overlap in visual feature space.

6. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, Step 4 specifically includes: The conditional generative adversarial network model includes a generator and a discriminator. The generator receives the original remote sensing image and the mechanical field labels provided by the finite element simulation as input. The mechanical field labels are represented in the form of multi-channel images, which respectively represent the deformation rate and stress concentration of different regions. The generator's internal architecture employs an encoder-decoder structure with skip connections. It extracts semantic information from images through convolution operations and fuses it with mechanical field constraints to output an enhanced landslide feature image. The discriminator is a multi-scale classification network that evaluates the realism of the generated image at the pixel level and assesses whether the deformation texture presented by the image has mechanical consistency with the input mechanical field label. During training, the loss function is constructed using a comprehensive weighted approach, including adversarial loss, perceptual loss, and structural similarity loss. The perceptual loss is used to ensure that the generated image is consistent with the real landslide features at the semantic level, and the structural similarity loss is used to maintain the sharpness and continuity of the edges of the deformed region.

7. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, Step 5 specifically includes: the lightweight mechanical constraint module is an embedded neural network substructure, consisting of a convolutional kernel and a global average pooling layer, and its input is a deformation feature map extracted from the image to be identified, which contains information on surface texture anomalies, tone changes, and geometric distortions. The output of the lightweight mechanical constraint module is a continuously distributed probability value, which represents the probability that the surface deformation reflected by the current image features conforms to the typical mechanical evolution path of a landslide. During the training phase, the lightweight mechanical constraint module learns synchronously with the main recognition part of the conditional generative adversarial network model through a joint optimization strategy, establishing a mapping model from visual deformation patterns to mechanical instability states. During the reasoning phase, the lightweight mechanical constraint module analyzes the intersection relationship of cracks, the distribution centroid of deformation zones, and the directionality of surface displacement to inversely deduce whether the surface deformation characteristics are caused by the internal shear stress of the slope exceeding the shear strength.

8. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, Step 6 specifically includes: the visual feature matching conditions are determined by a trained deep convolutional neural network classifier, which is used to distinguish between landslide areas and non-landslide areas and outputs a visual confidence score of landslides. The mechanical conditions are determined by the probability value output by the lightweight mechanical constraint module; when the landslide confidence score at the visual level exceeds the first preset threshold and the probability value exceeds the second preset threshold, the system determines that there is a landslide hazard in the area. Once the determination is established, the geographical coordinates, deformation area, and estimated impact range of the landslide hazard area are extracted, and different levels of early warning signals are triggered based on the identified deformation severity, which are then pushed to the disaster management department via wireless network.

9. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, It also includes a dynamic tracking and monitoring step for the evolution trend of landslides during rainfall: the finite element simulation process of soil mechanics in step 2 is refined into a non-steady-state analysis with multiple time steps, the rainfall conditions are set as a function sequence that changes with time, and the shear strength parameters of the soil and rock mass are dynamically adjusted according to the real-time simulated changes in soil moisture content. By inputting remote sensing images from multiple consecutive periods, comparing the image displacement over three consecutive observation periods, the displacement vector changes of feature points within the potential hazard area are calculated, and the deformation rate and its rate of change over time are derived. If the deformation rate exhibits a continuous accelerating growth characteristic, and the rate of change of the deformation rate over time exceeds a set critical warning value, then the warning level is upgraded. The determination threshold of the lightweight mechanical constraint module is dynamically adjusted based on the current real-time rainfall intensity. When the measured rainfall exceeds the time-based landslide induction threshold of the target area, the sensitivity threshold of the mechanical condition determination is lowered.

10. The landslide image recognition and early warning method based on image recognition according to claim 1, characterized in that, It also includes the steps of adaptive optimization and distributed processing of the model: the hybrid training dataset is incrementally updated after each new real landslide case is added, the terrain parameters and geomechanical parameters of the location of the new case are analyzed, the finite element simulation is rerun to generate virtual samples that match the geological conditions of the new case, and the model is fine-tuned using online learning algorithms; The conditional generative adversarial network model employs model pruning and quantization techniques. By analyzing the contribution of neuron connections, redundant parameter branches are eliminated, and floating-point operations are converted into fixed-point operations to adapt to edge computing devices. For remote sensing image data of large-scale mountainous areas, a parallel computing cluster is used to simultaneously call the finite element analysis kernel of soil and rock mechanics, and a pre-calculated physical reasonable deformation model library is established. For slopes with similar geometric features and geological parameters, the model library is searched and fine-tuned to generate the physical reasonable landslide deformation visualization image.

Images