An intelligent control system for monitoring geological disaster hazards by using an airborne LiDAR

Abstract

The application discloses an intelligent control system for monitoring geological disaster hidden dangers by using an airborne LiDAR, and relates to the technical field of geological disaster monitoring.The system comprises a data acquisition module, a data processing and fusion module, an image intelligent interpretation module, a disaster risk intelligent prediction module, a dynamic monitoring control module, a self-adaptive learning optimization module and an early warning and decision support module.In the application, laser point clouds, positioning and orientation and surface image data are collected, original point clouds are adaptively denoised and strip adjustment is performed, a digital elevation model and a digital orthographic image are generated, vegetation cover is penetrated and terrain data noise interference is suppressed, geological disaster features are intelligently identified and boundaries are extracted, risk dynamic prediction and model parameter self-learning optimization are performed, monitoring strategies can be adaptively adjusted, hidden danger identification precision, prediction reliability and monitoring resource allocation efficiency are improved, and the risk of misjudgment and missed judgment caused by environmental complexity and data uncertainty is reduced.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and in particular to an intelligent control system that uses airborne LiDAR to monitor potential geological disaster risks. Background Technology

[0002] Airborne LiDAR is a technology that uses a lidar mounted on an aircraft or drone to quickly acquire three-dimensional surface data with centimeter-level precision. By penetrating vegetation, detecting subtle changes on the surface, and comparing data from multiple periods, it can identify potential geological hazards such as landslides, collapses, and debris flows, enabling full-process monitoring from early identification and risk assessment to post-disaster emergency response.

[0003] Currently, relying on a single data source or static analysis model makes it difficult to penetrate vegetation to obtain the true surface morphology. The automation and intelligence of processing massive point cloud and image data are insufficient, and reliance on manual interpretation leads to low efficiency and strong subjectivity. At the same time, the lack of deep integration of multi-source monitoring data and model self-optimization mechanism based on real-time feedback makes it impossible to dynamically adapt to different terrain and vegetation conditions, resulting in limited accuracy of hazard identification and delayed risk warning.

[0004] Therefore, an intelligent control system that utilizes airborne LiDAR to monitor potential geological hazards is proposed to address the aforementioned problems. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent control system for monitoring potential geological hazards using airborne LiDAR, in order to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: an intelligent control system for monitoring potential geological hazards using airborne LiDAR, the system comprising: Data acquisition module: Real-time acquisition of laser point cloud data, positioning attitude data and surface image data of geological hazard-prone areas through airborne LiDAR sensors, GNSS / IMU system and optical camera, and output of multi-source raw monitoring data; Data processing and fusion module: Receives raw monitoring data from multiple sources, performs noise reduction, smoothing and flight strip adjustment on point clouds, generates digital elevation models and digital orthophotos, and outputs standardized terrain and image datasets. Image intelligent interpretation module: Based on standardized terrain and image datasets, combined with deep learning algorithms, it identifies geological hazard features and extracts boundaries, outputting the location, type, and morphological parameters of potential hazard points; The intelligent disaster risk prediction module: Based on the location, type, and morphological parameters of potential hazards, combined with historical disaster data and real-time environmental factors, it uses random forest algorithm and time series analysis to predict the disaster risk level and occurrence time, and generates a risk prediction report; The dynamic monitoring and control module adaptively adjusts the monitoring frequency, sensor parameters, and flight path based on the risk prediction report, thereby achieving dynamic optimization of monitoring resources. The adaptive learning optimization module dynamically updates the disaster identification and prediction model parameters based on reinforcement learning algorithms by comparing the prediction results with the actual monitoring data, thereby improving the system's identification accuracy and prediction reliability. The early warning and decision support module automatically triggers multi-level early warnings and generates emergency plans when the disaster risk level exceeds a preset threshold.

[0007] Preferably, the data acquisition module includes a LiDAR acquisition unit, a positioning and attitude determination unit, and an image acquisition unit; The LiDAR acquisition unit acquires three-dimensional point cloud data of geological disaster-prone areas in real time through an airborne lidar sensor, and generates a high-density point cloud data stream based on multi-echo technology to penetrate vegetation cover. The positioning and attitude determination unit records the precise position and attitude information of the flight platform in real time through the GNSS / IMU integrated navigation system, and generates a positioning and attitude determination data stream; The image acquisition unit acquires true-color images of the ground surface using a synchronously mounted high-resolution optical camera, and generates an orthophoto data stream.

[0008] Preferably, the data processing and fusion module includes a point cloud preprocessing unit, a digital elevation model / digital orthophoto generation unit, and a multi-source fusion unit; The point cloud preprocessing unit performs noise reduction, smoothing and flight strip adjustment on the original point cloud based on wavelet transform and Kalman filter algorithms to generate clean point cloud data. The digital elevation model / digital orthophoto generation unit generates digital elevation models and digital orthophotos based on purified point cloud data and optical images, through triangulation construction and orthorectification technology. The multi-source fusion unit integrates digital elevation models, digital orthophotos, and positioning data into a standardized terrain and image dataset with a unified coordinate system through spatiotemporal registration and data fusion algorithms.

[0009] Preferably, the point cloud preprocessing unit is based on wavelet transform and Kalman filtering algorithms, and includes the following steps: Primary filtering stage: A moving average filter is used to perform preliminary smoothing on the original point cloud, filtering out high-frequency noise and isolated noise points; Intermediate filtering stage: A threshold denoising algorithm based on wavelet transform is used to adaptively filter point cloud signal components of different frequency bands to separate vegetation and surface reflection signals. Advanced filtering stage: The dynamic point cloud is predicted and corrected using a Kalman filter, and the flight strip adjustment and system error elimination are achieved by combining POS data; Consistency verification phase: The correlation coefficient analysis method is used to verify the consistency between the filtered point cloud and the original data to ensure that the terrain features are completely preserved.

[0010] Preferably, the image intelligent interpretation module includes a feature extraction unit, a disaster identification unit, and a boundary refinement unit; The feature extraction unit automatically extracts terrain slope, aspect, curvature, and texture features from digital elevation models and digital orthophotos based on a deep learning network; The disaster identification unit identifies geological disaster types such as landslides, collapses, and debris flows through a convolutional neural network classification model, and outputs the location and type labels of potential hazards. The boundary refinement unit uses edge detection and morphological algorithms to optimize the boundaries and calculate morphological parameters of the recognition results.

[0011] Preferably, the disaster risk intelligent prediction module includes a risk assessment unit, a time prediction unit, and an environmental correction unit; The risk assessment unit uses a random forest algorithm to assess the geological disaster risk level based on the morphological parameters of potential hazard points and a historical disaster database. The time prediction unit uses an LSTM time series model to predict the possible time points of a disaster and generates a disaster timeline. The environmental correction unit combines real-time rainfall, soil moisture, and seismic activity data to dynamically correct the prediction model.

[0012] Preferably, the dynamic monitoring and control module includes a route planning unit, a sensor control unit, and a monitoring and scheduling unit; The route planning unit dynamically optimizes the UAV's flight path and altitude using a genetic algorithm based on risk level and terrain complexity. The sensor control unit adaptively adjusts the LiDAR scanning frequency and camera exposure parameters according to vegetation cover and weather conditions. The monitoring and scheduling unit intelligently allocates monitoring resources and task priorities based on prediction results and real-time data.

[0013] Preferably, the adaptive learning optimization module includes a difference analysis unit, a model update unit, and a performance evaluation unit; The difference analysis unit calculates the identification error and prediction deviation by comparing the predicted disaster location with the actual monitoring data; The model update unit dynamically adjusts the weights of the deep learning network and the parameters of the prediction model based on a reinforcement learning mechanism. The performance evaluation unit periodically assesses the model accuracy and system stability, triggering the model reconstruction and optimization process.

[0014] Preferably, the model update unit is based on a reinforcement learning mechanism and includes the following steps: State perception phase: Real-time acquisition of the error between predicted disaster location and actual monitoring data, changes in environmental variables, and system identification accuracy indicators to construct a state vector; Action decision-making stage: Based on the deep Q-network model, the optimal model parameter update strategy is selected according to the current state, including neural network weight adjustment and prediction model parameter correction; Reward feedback phase: Design a reward function based on the improvement rate of recognition accuracy, the accuracy of early warning, and the system response time, and dynamically optimize the strategy through temporal difference learning; Iterative optimization phase: The network weights are gradually updated through multiple rounds of interactive training, and an experience replay mechanism is used to prevent overfitting, so as to achieve continuous self-learning and performance improvement of the model in complex environments.

[0015] Preferably, the early warning and decision support module includes an early warning issuing unit and a report generation unit; When the risk level exceeds a preset threshold, the early warning issuing unit automatically issues an audible and visual alarm, an SMS notification, and a platform push notification. The report generation unit generates geological disaster monitoring reports and emergency plans based on the interpretation and prediction results.

[0016] The present invention has the following beneficial effects: 1. In this invention, when performing intelligent identification of geological disaster hazards, the data acquisition module simultaneously acquires multi-source data such as laser point cloud, positioning and attitude determination, and surface imagery. Wavelet transform and Kalman filter multi-level algorithms are used to adaptively denoise, smooth, and adjust the flight strips of the original point cloud. This effectively penetrates dense vegetation and suppresses high-frequency noise and systematic errors in the terrain data, ensuring the accuracy of the generation of high-precision digital elevation models (DEM) and digital orthophotos (DOM), improving the reliability of subsequent feature extraction and hazard identification, and reducing misjudgments and omissions caused by poor data quality.

[0017] 2. In this invention, when conducting dynamic assessment and monitoring and control of geological disaster risks, the disaster risk intelligent prediction module combines historical data and real-time environmental factors to predict the risk level and occurrence time. Based on the prediction report, the dynamic monitoring and control module adaptively adjusts the flight path, sensor parameters, and monitoring task priority, enabling the system to achieve on-demand optimized allocation of monitoring resources and risk tracking. This significantly improves the monitoring efficiency and response timeliness in complex mountainous areas, overcoming the shortcomings of rigid resource utilization and delayed early warning in traditional static monitoring modes.

[0018] 3. In this invention, during continuous system optimization and reliability improvement, the adaptive learning optimization module compares the prediction results with the actual monitoring data in real time, and dynamically updates the identification and prediction model parameters based on the deep reinforcement learning mechanism. This enables the system to adapt to different terrain, vegetation and environmental changes, realize the model's self-learning and performance iteration improvement under complex conditions, thereby continuously improving the accuracy of hazard identification, prediction accuracy and overall system robustness, and solving the key limitations of existing methods that rely on fixed models and are difficult to adapt to dynamic scenarios. Attached Figure Description

[0019] Figure 1 This is a framework diagram of an intelligent control system for monitoring potential geological hazards using airborne LiDAR, as described in this invention. Figure 2 This is an implementation framework diagram of the point cloud preprocessing unit of an intelligent control system for monitoring potential geological hazards using airborne LiDAR according to the present invention; Figure 3 This is a model update unit framework diagram of an intelligent control system for monitoring potential geological hazards using airborne LiDAR, as described in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-3 This invention provides a technical solution: an intelligent control system for monitoring potential geological hazards using airborne LiDAR, the system comprising: Data acquisition module: Real-time acquisition of laser point cloud data, positioning attitude data and surface image data of geological hazard-prone areas through airborne LiDAR sensors, GNSS / IMU system and optical camera, and output of multi-source raw monitoring data; Data processing and fusion module: Receives raw monitoring data from multiple sources, performs noise reduction, smoothing and flight strip adjustment on point clouds, generates digital elevation models and digital orthophotos, and outputs standardized terrain and image datasets. Image intelligent interpretation module: Based on standardized terrain and image datasets, combined with deep learning algorithms, it identifies geological hazard features and extracts boundaries, outputting the location, type, and morphological parameters of potential hazard points; The intelligent disaster risk prediction module: Based on the location, type, and morphological parameters of potential hazards, combined with historical disaster data and real-time environmental factors, it uses random forest algorithm and time series analysis to predict the disaster risk level and occurrence time, and generates a risk prediction report; The dynamic monitoring and control module adaptively adjusts the monitoring frequency, sensor parameters, and flight path based on the risk prediction report, thereby achieving dynamic optimization of monitoring resources. The adaptive learning optimization module dynamically updates the disaster identification and prediction model parameters based on reinforcement learning algorithms by comparing the prediction results with the actual monitoring data, thereby improving the system's identification accuracy and prediction reliability. The early warning and decision support module automatically triggers multi-level early warnings and generates emergency plans when the disaster risk level exceeds a preset threshold.

[0022] The data acquisition module includes a LiDAR acquisition unit, a positioning and attitude determination unit, and an image acquisition unit; The LiDAR acquisition unit acquires real-time 3D point cloud data of geological hazard-prone areas using an airborne lidar sensor. Based on multi-echo technology, it penetrates vegetation cover to generate a high-density point cloud data stream, including the following steps: Laser emission and echo reception: The airborne lidar sensor emits laser pulses to the ground at a preset pulse frequency and simultaneously receives multiple echo signals reflected from the ground surface, vegetation and artificial objects. Each echo signal records its emission and return timestamps, reflection intensity and echo sequence number. Distance and coordinate calculation: based on the time difference of laser pulse propagation. and the speed of light Through formula Calculate the slant range from the sensor to the ground target; then combine this with the real-time acquired position of the flight platform. With attitude angle (roll) , up and down ,course ), and laser scanning angle The three-dimensional coordinates of each laser foot point are calculated using a spatial geometric transformation model. The conversion formula is: ; in The rotation matrix is ​​calculated from the attitude angles; Point cloud stream generation: The calculated coordinates, reflection intensity and echo sequence of all laser foot points are integrated in time sequence to form a high-density three-dimensional point cloud data stream with multiple echo attributes, and output to the data processing and fusion module in real time; The positioning and attitude determination unit records the precise position and attitude information of the flight platform in real time through the GNSS / IMU integrated navigation system, generating a positioning and attitude determination data stream, including the following steps: Multi-source positioning data synchronous acquisition: The system receives satellite signals from multiple constellations (GPS, BeiDou, GLONASS) in real time through a Global Navigation Satellite System (GNSS) receiver, and calculates the absolute position (longitude, latitude, and altitude) and time information of the flight platform; at the same time, the system measures the three-axis acceleration and three-axis angular velocity of the flight platform in real time through an inertial measurement unit (IMU). Integrated navigation calculation: The Kalman filter algorithm is used to fuse GNSS position data and IMU inertial data. Through state prediction and measurement updates, the precise position of the flight platform is estimated in real time. With attitude angle The state vector includes position, velocity, attitude, and sensor error. The state update equation is: ; in For state vectors, Here is the state transition matrix. for The state vector at any given time; Measurement updates utilize GNSS position and velocity observations to correct the state, outputting a high-frequency, high-precision positioning and attitude determination data stream.

[0023] Data time synchronization and output: The calculated positioning and attitude data is strictly synchronized with the LiDAR laser pulse emission time, and time tags are added to generate a positioning and attitude data stream with a unified time reference for subsequent point cloud calculation and data fusion. The image acquisition unit acquires true-color images of the Earth's surface using a synchronously mounted high-resolution optical camera, generating an orthophoto data stream, including the following steps: Image exposure control and acquisition: Based on the preset flight path and flight parameters, when the UAV reaches the designated exposure point, the flight control system triggers the high-resolution optical camera to perform exposure; the camera adaptively adjusts the aperture, shutter speed and ISO according to real-time lighting conditions to ensure correct and clear image exposure; Image geolocation: At the moment the camera is exposed, the precise location provided by the positioning and attitude determination unit is recorded simultaneously. With attitude angle This serves as the exterior orientation element of the image; simultaneously, the camera interior orientation element (focal length) is utilized. Like the main point Distortion coefficient Establish a geometric imaging model for the image; Image data generation and output: The acquired raw image data (usually in RAW or JPEG format) is packaged with its corresponding exterior orientation elements, timestamps and camera parameters to generate a true-color orthophoto data stream with georeferenced information, and then output to the data processing and fusion module in real time.

[0024] The data processing and fusion module includes a point cloud preprocessing unit, a digital elevation model / digital orthophoto generation unit, and a multi-source fusion unit; The point cloud preprocessing unit performs noise reduction, smoothing and flight strip adjustment on the original point cloud based on wavelet transform and Kalman filter algorithms to generate clean point cloud data. The digital elevation model / digital orthophoto generation unit generates digital elevation models and digital orthophotos based on cleaned point cloud data and optical images, using triangulation construction and orthorectification techniques, including the following steps: Point cloud classification and ground point extraction: Based on the echo count, reflection intensity and geometric features of the purified point cloud, a progressive triangulation densification filtering algorithm is used to separate ground points from non-ground points (vegetation, buildings); by iteratively constructing an initial sparse triangulation and gradually densifying it, points located within a certain distance threshold below the triangulation are retained as ground points, forming the point set required for the digital terrain model (DTM); Triangulation network construction and DEM interpolation: An irregular triangulation network (TIN) is constructed for the extracted ground points to ensure that there are no intersections or holes; linear interpolation is used to interpolate the TIN into a regular grid to generate a digital elevation model (DEM) with a specified resolution (e.g., 0.5 meters). The elevation value of each grid point is obtained by linearly weighting the elevations of the three vertices of the triangle in which it is located. Image orthorectification and DOM generation: Using the digital elevation model (DEM) and the exterior orientation elements of the image (obtained from the positioning and orientation unit), the original optical image is digitally differentially corrected based on the collinearity equation to eliminate the projection difference caused by terrain undulation; multiple corrected images are processed for uniform lighting and color, and then mosaicked and blended along the preset stitching lines to generate a digital orthophoto image (DOM) with consistent color and geometric accuracy. The multi-source fusion unit, through spatiotemporal registration and data fusion algorithms, integrates digital elevation models, digital orthophotos, and positioning data into a standardized terrain and image dataset with a unified coordinate system, including the following steps: Spatiotemporal benchmark unification and registration: The generated DEM and DOM data, as well as POS trajectory data, are all converted to a unified plane coordinate system (such as CGCS2000 Gaussian projection) and elevation benchmark (such as the 1985 National Elevation Benchmark); Fine registration is performed using the elevation information of prominent features on the DOM and the corresponding DEM to ensure that the image pixels and terrain grid points correspond one-to-one in space. Feature-level data fusion: Extracting spectral and texture features (such as Normalized Difference Vegetation Index NDVI and Gray-Level Co-occurrence Matrix Energy) from the DOM, and extracting terrain features (such as slope) from the DEM. Slope aspect (curvature), where slope is calculated as ,in The elevation is the derivative in the x and y directions; the above features are correlated and superimposed at the pixel / grid level to form a multi-dimensional feature layer; Standardized dataset generation and output: The registered DEM and DOM raw data layers are integrated with multi-dimensional feature layers and encapsulated into a standardized terrain and image dataset with unified geographic range, resolution, coordinate reference and attribute structure, providing standardized and complete input data for the subsequent image intelligent interpretation module.

[0025] The cloud preprocessing unit, based on wavelet transform and Kalman filtering algorithms, includes the following steps: Primary filtering stage: A moving average filter is used to perform preliminary smoothing on the original point cloud, filtering out high-frequency noise and isolated noise points, including the following steps: For each point in the raw point cloud data stream Establish a spatial neighborhood (such as a sphere with radius R or K nearest neighbors) centered on it, and calculate the elevation values ​​of all points within this neighborhood. arithmetic mean , to take as The smoothed elevation is used to filter out high-frequency random noise. Based on the statistical outlier detection method, the average elevation distance between each point and its K nearest neighbors is calculated. and standard deviation , will satisfy (in The global average distance. This is a preset coefficient, usually set to 2-3. For all points Points with a standard deviation of 0.5% are identified as isolated noise points and removed. Intermediate filtering stage: A threshold denoising algorithm based on wavelet transform is used to adaptively filter point cloud signal components of different frequency bands, separating vegetation and surface reflection signals. This includes the following steps: The point cloud data after primary filtering is used to construct a one-dimensional elevation signal sequence along the elevation direction (or along the scan line). A suitable wavelet basis function (such as Daubechies wavelet) is selected to perform multi-level wavelet decomposition on the signal to obtain approximation coefficients (low frequency) and detail coefficients (high frequency) at different scales. For signal components in different frequency bands, an adaptive thresholding function (such as a soft thresholding function) is used to process the detail coefficients. In the In layer decomposition, based on the noise variance of that layer Adaptive determination: ,in Given the signal length, this process effectively suppresses high-frequency noise representing non-surface reflections such as vegetation, while retaining low-frequency and some mid-frequency signals representing the actual terrain undulations. Wavelet reconstruction is performed on the coefficients after thresholding to obtain the denoised elevation signal, which is then mapped back to the 3D point cloud. Advanced filtering stage: Predictive correction of dynamic point clouds is performed using a Kalman filter, combined with POS data to achieve flight strip adjustment and systematic error elimination, including the following steps: The point cloud data of each flight strip is matched with its corresponding POS data (location data). ,attitude Strict time synchronization is used to establish the state-space model of the Kalman filter, and the state vector... Includes systematic error parameters that need to be estimated, such as the three-dimensional translational deviation between flight strips. and rotational deviation (under the assumption of small angles) ); The system state equations describe the propagation of error parameters: ,in Here is the state transition matrix. To mitigate process noise, pairs of corresponding points automatically matched within the overlapping area of ​​adjacent flight strips are used as observations, and their three-dimensional coordinate differences constitute the observation vector. Observation equation , For the observation matrix, To observe noise, the optimal state estimate is iteratively solved through the prediction and update steps of Kalman filtering, and then the point cloud coordinates of each flight strip are corrected to achieve flight strip adjustment and generate a seamlessly stitched clean point cloud. Consistency verification phase: The consistency between the filtered point cloud and the original data is verified using correlation coefficient analysis to ensure complete preservation of terrain features. This includes the following steps: To ensure that filtering and adjustment do not damage the real terrain features, a subset of the original point cloud and the corresponding purified point cloud from a typical terrain region (such as a stable, continuous natural slope) were selected. Their high-order sequences were extracted, and the Pearson correlation coefficient between them was calculated. : ; in, and The first The elevation of each point in the original and purified point clouds. and To determine the average elevation, set a threshold close to 1 (e.g., 0.98); if the calculated... If the value exceeds this threshold, the purified point cloud is considered to have effectively preserved the original terrain features and passes verification; otherwise, it is necessary to return to check the preceding filtering parameters and possibly reprocess.

[0026] The image intelligent interpretation module includes a feature extraction unit, a disaster identification unit, and a boundary refinement unit; The feature extraction unit automatically extracts terrain slope, aspect, curvature, and texture features from digital elevation models and digital orthophotos based on a deep learning network, including the following steps: Input data preparation and preprocessing: The digital elevation model (DEM) and digital orthophoto (DOM) in the standardized terrain and image dataset are used as input. The DEM is normalized to eliminate the influence of dimensions; the DOM is color normalized to reduce the interference of illumination changes; then, the two are accurately registered in space to ensure that the terrain information of each pixel / grid point corresponds one-to-one with the image information. Deep learning-based composite feature extraction: Construct a two-branch deep learning network (such as an encoder structure), with one branch taking the multispectral bands of the DOM as input and the other branch taking the DEM as input; the DOM branch automatically extracts multi-scale spectral texture features (such as edges, corners, and local patterns) through convolutional layers; the DEM branch automatically extracts terrain contours and spatial context features through convolutional layers; simultaneously, classical terrain factors are directly calculated from the DEM as a supplement. ; ; in For elevation exist Rate of change of direction For elevation exist Rate of change of direction; Multi-source feature fusion and output: The abstract feature map extracted by the deep learning network and the calculated classic terrain factor feature map are concatenated or weighted in the feature dimension to form a high-dimensional composite feature tensor containing spectral, texture, morphological and terrain physical attributes, which serves as the direct input of the disaster identification unit; The disaster identification unit identifies geological disaster types such as landslides, collapses, and debris flows using a convolutional neural network classification model, and outputs the location and type labels of potential hazards, including the following steps: Convolutional Neural Network Model Construction and Training: A multi-task convolutional neural network (CNN) model is constructed, typically consisting of multiple convolutional layers, pooling layers, and fully connected layers. The model's input is the composite feature tensor generated in the previous step. The output layer is designed with two branches: one branch is used for geological hazard type classification (e.g., landslide, collapse, debris flow, no hazard), employing the Softmax activation function; the other branch is used for hazard location regression (i.e., bounding box prediction). The model is trained under supervision using a historical dataset containing labeled hazard locations and types. The loss function is a weighted sum of classification cross-entropy loss and location regression loss (e.g., Smooth L1 Loss). (in This is the total loss function; These are the weighting coefficients for the classification loss; For classification loss, cross-entropy loss is typically used; These are the weighting coefficients for the localization loss; To determine the loss, Smooth L1 Loss is typically used. Forward inference and initial hazard assessment: The composite feature tensor of the region to be interpreted is input into a trained CNN model. After forward propagation, the probability distribution of hazard type for each potential hazard region is obtained from the classification branch. ; Obtain the corresponding bounding box coordinates from the regression branch; By setting a probability threshold (e.g. The non-maximum suppression (NMS) algorithm is used to filter out preliminary hazard identification results with high confidence, including their center coordinates, circumscribed rectangle range, and hazard type label; Structured information output: The information of each hidden danger point obtained in the initial judgment (including category label, confidence level, and bounding rectangle) is structured and organized to form a list of hidden danger points, which is then output to the boundary refinement unit for subsequent fine processing; The boundary refinement unit uses edge detection and morphological algorithms to optimize the boundaries and calculate morphological parameters of the recognition results, including the following steps: Edge detection-based boundary optimization: For each initially identified potential hazard point, within its bounding rectangular region, corresponding image and terrain patches are cropped from the original DOM and DEM. First, edge detection operators such as Canny are applied to the DOM patches to calculate the image gradient magnitude. The initial edge pixels are extracted by using double thresholding and hysteresis thresholding, and the slope change rate is calculated for small DEM blocks to help identify terrain change lines. Morphological operations and boundary fusion: Morphological operations are performed on the extracted initial edge map. First, the closing operation (dilation followed by erosion) is used to connect the broken edge segments; then, the opening operation (erosion followed by dilation) is used to remove small isolated noise points; the edges extracted by DOM are logically ORed with the terrain abrupt change lines extracted by DEM to obtain a more complete and accurate binary mask of the disaster body boundary. Boundary vectorization and morphological parameter calculation: The final binary boundary mask is contour-tracked to extract its outer contour polygon, which is then vectorized and stored as a vector boundary. Based on this vector boundary and the original DEM, the key morphological parameters of the potential hazard point are calculated: area. (Calculated using the polygon area formula), perimeter The length of the major axis / minor axis, the main direction, and the average thickness of the sliding body (estimated by the difference in elevation between the DEM inside and outside the boundary) are calculated. The optimized vector boundary and the calculated morphological parameters are then updated into the hazard point information to complete the interpretation.

[0027] The intelligent disaster risk prediction module includes a risk assessment unit, a time prediction unit, and an environmental remediation unit; The risk assessment unit uses a random forest algorithm to assess the geological hazard risk level based on the morphological parameters of hazard points and a historical disaster database. The steps include: Feature vector construction and data preparation: Extracting morphological parameters, such as area, from the hazard point information output by the image intelligent interpretation module. ,perimeter Average slope Relative height difference curvature Simultaneously, historical disaster cases under similar geological environments are matched from the historical disaster database, and their corresponding morphological parameters and final disaster level labels (usually divided into low, medium, and high risk levels) are extracted. All parameters are standardized to eliminate the influence of dimensions, and a feature vector is constructed for each sample. and corresponding risk level labels ; Random Forest Model Training and Prediction: A random forest classification model is constructed, which consists of an ensemble of multiple decision trees. During the training phase, feature vectors and labels from historical samples are used to generate a training subset for each tree via bootstrapping. When a node splits, the optimal splitting feature and threshold are selected from a randomly chosen feature subset. The splitting criterion is typically the Gini index. (in For the node The proportion of samples of each class is minimized; the final classification is determined by voting among multiple trees; for the current potential hazard point, its feature vector is input into the trained random forest model, each tree outputs a class probability, and finally, the voting results of all trees are combined to obtain the probability distribution of the potential hazard point belonging to each risk level. ; Risk level determination and output: based on probability distribution The level corresponding to the highest probability is taken as the preliminary risk level of the potential hazard. Simultaneously, a confidence threshold can be set. If the maximum probability is lower than the threshold, it is marked as requiring manual review, and the risk level is adjusted accordingly. Its probability vector is output to the time prediction unit and the environment correction unit; The time prediction unit uses an LSTM time series model to predict the possible time points of a disaster and generates a disaster timeline, including the following steps: Time series dataset construction: For each disaster event that has occurred in the historical disaster database, relevant dynamic factors within a certain period before its occurrence are extracted, including but not limited to: temporal changes in the morphological parameters of potential hazards (which can be obtained from multiple monitoring data periods), and periodic environmental factors (such as monthly cumulative rainfall). Average soil moisture and triggering events (such as peak ground acceleration). Each sample is constructed as a time step of... Sequence data ,in for The feature vector at time point corresponds to the disaster occurrence time (which can be converted into a time interval from the end of the sequence). ); LSTM Model Construction and Training: A Long Short-Term Memory (LSTM) network model is constructed, the core of which is a memory unit containing a forget gate, an input gate, and an output gate; at time step... The LSTM unit is based on the current input The hidden state of the previous moment and cell state Calculate the new cell state and hidden state ;Specifically: The forget gate controls the degree to which the previous state is retained: ; Input gate control for new information input: ; Candidate cell status: ; Cell status update: ; Output gate: ; Final hidden state ; The model utilizes the last hidden state of the sequence. Predicting time intervals using regression in fully connected layers Mean squared error loss is used during training: (in (sample size) in These represent the times of the forget gate, input gate, and output gate, respectively. The output; This is the weight matrix for the corresponding gate; This is the bias vector for the corresponding gate; Use the Sigmoid activation function; This represents the current state of the candidate cells. The cell states at the previous and current time points; Disaster timing prediction and timeline generation: For current potential hazard points, collect their latest data. The dynamic feature sequence at each time step is input into a trained LSTM model to predict the estimated time interval from the current moment to the possible occurrence of a disaster. Combined with the current time Calculate and predict the time of disaster occurrence Repeat this process for all potential hazard points within the area, according to their... The data is sorted sequentially to generate a regional disaster risk timeline, and the predicted time points and corresponding confidence intervals are output. The environmental correction unit combines real-time rainfall, soil moisture, and seismic activity data to dynamically correct the prediction model, including the following steps: Real-time environmental data acquisition and fusion: Real-time acquisition of rainfall intensity in the area where potential hazards are located via external sensor networks or meteorological / seismic station interfaces. Soil volumetric moisture content and seismic activity index Environmental data, along with the static morphological parameters and historical risk levels of the potential hazard points, will be used to analyze these data. and prediction time interval Spatiotemporal alignment and fusion are performed to construct a composite feature vector for correction. ; Calculation of dynamic correction coefficients based on environmental factors: Establishing a quantitative model of the impact of environmental factors on disaster risk; for example, calculating environmental correction coefficients using weighted linear combinations or nonlinear mapping functions (such as the Sigmoid function). : ; in These are the critical thresholds for each factor. The weighting coefficient is learned from historical data and is used to adjust the risk probability or prediction time. Dynamic updating of model parameters and correction of prediction results: adjusting the correction coefficients Introducing risk assessment and time prediction models; for risk assessment, it can be... The features can be incorporated into the random forest model for re-prediction, or the risk probability can be directly adjusted. (Normalization required) For time-based forecasting, the forecast time interval can be adjusted: ,in To adjust the parameters, the final output is the risk level after environmental correction. and the revised predicted disaster timing And update the risk prediction report.

[0028] The dynamic monitoring and control module includes a route planning unit, a sensor control unit, and a monitoring and scheduling unit; The route planning unit dynamically optimizes the UAV's flight path and altitude using a genetic algorithm based on risk level and terrain complexity, including the following steps: Path and altitude joint optimization based on an improved genetic algorithm: First, possible flight sequences (i.e., the order in which waypoints are visited) and the relative altitude of each flight segment are jointly encoded to form chromosomes, and a population containing multiple chromosomes is initialized; then, the fitness of each chromosome is calculated. This is the reciprocal of the objective function value; a larger value indicates a better solution. Iterative evolution is then performed: a roulette wheel selection method is used to select parent chromosomes based on fitness; sequential crossover (OX) and two-point mutation operations are performed on the selected parents to generate new offspring chromosomes; simultaneously, a specialized mutation operator is designed to fine-tune flight altitude, ensuring optimal performance in complex terrain (high altitude). (Above flight altitude) satisfy ( Based on the safety height, (This is an adjustment coefficient). A certain proportion of elite individuals are retained in each generation and directly enter the next generation.

[0029] Decoding and Flight Plan Generation: When the algorithm reaches the preset maximum number of iterations or the fitness converges, the chromosome with the highest fitness is selected for decoding; the decoding process restores the chromosome sequence to the specific waypoint visiting order. and flight altitude of each segment The path is smoothed according to the performance constraints of the UAV (such as minimum turning radius and maximum climb rate) to generate the final, executable flight trajectory file (such as KML or mission file) and output to the flight control system and monitoring and scheduling unit. The sensor control unit adaptively adjusts the LiDAR scanning frequency and camera exposure parameters based on vegetation cover and weather conditions, including the following steps: Real-time environmental perception and demand analysis: Receives real-time results from airborne vision systems, meteorological sensors, and preliminary DOM / DEM analysis; key inputs include: vegetation cover estimated from real-time video streams. (Calculated by the proportion of green pixels), light intensity (lux), atmospheric visibility And the preset terrain category corresponding to the flight position; based on these parameters, assess the current data acquisition needs: high vegetation cover or low light conditions require enhanced LiDAR penetration and image signal-to-noise ratio; high dynamic change areas require higher sampling frequency; Rule-based and model-based adaptive parameter tuning: Establishing a rule base for sensor parameter tuning and a lightweight prediction model; for LiDAR sensors, the pulse repetition frequency... Based on vegetation cover and flight speed adjust: ,in As the reference frequency, For adjustment coefficients, high Appropriately increase To increase point cloud density; for optical cameras, exposure parameters are dynamically calculated using an exposure value (EV) model: the target EV is determined by the illumination intensity. The decision was made by adjusting the shutter speed. Aperture value and sensitivity This is achieved through a combination of methods, prioritizing a shutter speed sufficient to avoid motion blur. , (Image plane movement speed); Command issuance and feedback verification: The calculated optimal sensor parameter set ( Scanning angle , , The data is encapsulated as control commands and sent to the airborne LiDAR and camera controller in real time via the data transmission link. After the command is executed, a small sample of data (such as point cloud fragments or preview images) is collected for rapid quality assessment. If the assessment results do not meet the preset quality standards, the coefficients in the deviation fine-tuning parameter model are adjusted again. Based on prediction results and real-time data, the monitoring and scheduling unit intelligently allocates monitoring resources and task priorities, including the following steps: Multi-source task request integration and queue construction: Receives task requests from multiple sources: periodic inspection tasks based on risk prediction reports (fixed areas); emergency detailed investigation tasks targeting sudden high-risk points (such as a sudden increase in risk level); and specific area supplementary testing or verification tasks for model learning needs. All tasks are structured and described according to submission time, spatial location, and task type (general survey, detailed investigation, verification) to form an initial task queue. ; Dynamic priority scoring model: For each task in the task queue Calculate a dynamic priority score The scoring model comprehensively considers: the final risk level of the risk points corresponding to the task. (Higher levels are awarded higher scores), predicting the timing of disasters. The weighted scoring formula considers factors such as urgency (higher score for closer tasks), task attributes (emergency tasks have higher weight than learning tasks), and resource suitability (e.g., whether currently available drones and sensors are suitable for the task). The simplified weighted scoring formula is: ; in For the task Priority score, For the current time, For the task Type weights, Resource matching degree represents the suitability of the currently available drones and sensor configurations for the task (quantified value, such as 0~1). These are the weighting coefficients for each item; according to right Sort the tasks in descending order to obtain the priority task queue. ; Resource allocation and task assignment: Based on the number, type, endurance, and sensor configuration of currently available drone platforms, a greedy or matching algorithm is used to allocate high-priority queues. The system assigns tasks to the most suitable combination of drones and sensors. Geographical proximity is considered during assignment to merge tasks and reduce relocation, while ensuring that the maximum endurance of a single drone is not exceeded. A detailed task package (including flight path, sensor parameters, and monitoring area) is generated for each drone and sent through the communication unit. At the same time, the system monitors the task execution status. If a task is interrupted, there is a sudden change in weather, or a new higher priority task is inserted, the task rescheduling is dynamically triggered.

[0030] The adaptive learning optimization module includes a difference analysis unit, a model update unit, and a performance evaluation unit; The difference analysis unit calculates the identification error and prediction deviation by comparing the predicted disaster location with the actual monitoring data, including the following steps: Data alignment and matching: After the system completes a monitoring task and acquires actual surface data (such as the latest DOM / DEM) or confirms the actual occurrence of geological disasters through field verification, spatiotemporal data alignment is first performed. The predicted bounding boxes or vector polygons of each hazard point in the previous prediction report are spatially overlaid and matched with the new data or verification results in a unified coordinate system. If there are new hazard bodies or significant changes in morphology near the predicted point, they are associated with the prediction results. If there are no changes, they are considered true negative samples. Multi-dimensional error index calculation: For each successfully matched prediction and actual sample pair, a set of quantitative error indices is calculated. The main indices include: Positional deviation: Calculate the Euclidean distance between the center point of the predicted boundary and the center point of the actual boundary. ; Identification accuracy: compared with the predicted disaster types With actual type If they match, it is recorded as correct; otherwise, it is recorded as incorrect. Boundary fit: Calculate the predicted polygon. Compared to actual polygons intersection ratio Prediction time bias: For disasters that have already occurred, calculate the prediction time. Compared with the actual time absolute time difference The error indices of all samples are summarized to form an error matrix; Deviation Feature Extraction and Analysis: Based on the calculated error matrix, statistical analysis is performed to extract systematic deviation features, and calculations are performed on all samples. average and standard deviation To evaluate the overall boundary recognition accuracy and stability of the model; to analyze whether there are significant differences in the distribution of error indicators under different terrain categories (such as steep slopes, vegetated areas), different disaster types, or different environmental conditions (such as after rain); The model update unit dynamically adjusts the weights of the deep learning network and the parameters of the prediction model based on a reinforcement learning mechanism; The performance evaluation unit periodically assesses model accuracy and system stability, triggering a model reconstruction and optimization process, including the following steps: Periodic comprehensive performance index calculation: Every fixed evaluation period (such as monthly or after a certain number of monitoring tasks are completed), a comprehensive evaluation of the performance of all core modules of the system is conducted; the evaluation index set includes: Image intelligent interpretation module: accuracy on the latest validation set. (in It is a true positive. (false positives) and recall rate (in (False negatives) and F1 score; Disaster risk intelligent prediction module: accuracy of risk level classification, mean absolute error of prediction time for disasters that have occurred. Overall system: average task response time, false alarm rate and false negative rate; calculate the current values ​​of these indicators and their moving averages and trends over multiple periods; Stability diagnosis and degradation detection: Set warning thresholds for various indicators (such as F1 score below 0.85, or MAE increasing for three consecutive cycles); use statistical process control methods such as control charts to detect whether there is statistically significant degradation in performance; Optimization process triggering and instruction generation: If the performance metrics remain consistently higher than the target and stable, the current optimization pace is maintained; if performance degradation is detected or a warning threshold is triggered, a model reconstruction instruction is generated. This instruction may include expanding the training dataset, introducing new negative or hard samples, triggering a re-search or adjustment of the CNN or LSTM backbone network structure (such as increasing network depth), resetting part of the reinforcement learning agent's experience pool, guiding it to explore new policy spaces, and sending the generated instructions to the data storage management module and related modules to initiate a systematic retraining and optimization process until the performance recovers to a satisfactory level.

[0031] The model update unit is based on a reinforcement learning mechanism and includes the following steps: State awareness phase: Real-time acquisition of the error between predicted disaster location and actual monitoring data, changes in environmental variables, and system identification accuracy indicators to construct a state vector, including the following steps: The system collects outputs from the difference analysis unit, environmental monitoring data streams, and system operation logs in real time. Specifically, this includes the average intersection-union ratio (IUU) generated by comparing the recognition results of the image intelligent interpretation module with the actual verification data in the last M monitoring tasks. and classification accuracy The disaster risk intelligent prediction module's time-averaged absolute error for predicting events that have already occurred. Key environmental characteristics during the current cycle, such as average rainfall intensity. and vegetation cover change rate The system's key deep learning models (such as CNN feature extractors) contain intermediate layer activation statistics (e.g., mean, variance). These heterogeneous data are standardized and concatenated according to preset weights to form a state vector that comprehensively describes the system's current performance and its environment. ; Action decision-making phase: Based on the deep Q-network model, the optimal model parameter update strategy is selected according to the current state, including neural network weight adjustment and prediction model parameter correction, including the following steps: The state vector The input is fed into a deep Q-network (DQN), whose network parameters are: The network output is all possible adjustment actions given the current state. The corresponding Q value The Q-value represents an estimate of the long-term cumulative reward that can be obtained after performing the action, in the action space. The discretization design is a series of specific model parameter adjustment instructions; such as Multiply the learning rate of the last layer of the CNN classifier by a factor of 0.8; Increase the number of hidden layer units in the LSTM predictor by 10%; Adjusting the weighting coefficient of the slope feature in the random forest model increases... ; Keeping all parameters constant, the agent follows - Greedy strategy selects actions based on probability. Randomly explore the action space, with probability Select the action with the highest current Q value. The selected action command will be sent to the corresponding model executor; Reward Feedback Phase: A reward function is designed based on the improvement rate of recognition accuracy, the accuracy of early warning, and the system response time. A dynamic optimization strategy is implemented through temporal difference learning, including the following steps: Execute action After a complete monitoring and evaluation cycle, the system enters a new state. Calculate immediate rewards based on changes in performance metrics between the old and new states. reward function It is designed to encourage improved accuracy, accurate early warnings, and efficient responses, while penalizing meaningless increases in model complexity: ; in, It is a state The corresponding performance metrics, It is the change in model complexity (such as the number of parameters or the amount of computation). The weighting coefficients for each reward / penalty item, rewards With state transition information Store together in the experience replay pool middle; Iterative optimization phase: The network weights are gradually updated through multiple rounds of interactive training, and an experience replay mechanism is used to prevent overfitting, enabling the model to continuously learn and improve performance in complex environments. This includes the following steps: Periodically from the experience replay pool A small batch of empirical data is randomly sampled, and the parameters of the DQN are updated using a temporal difference (TD) learning objective. Target Q value The calculation is as follows: ; in, It is a discount factor. The parameters are those of the target network (a stable network that periodically replicates its parameters from the main network), and then, by minimizing the loss function... To update the parameters of the main DQN The experience replay mechanism breaks the temporal correlation between data and improves learning stability. Through repeated iterations, DQN gradually learns to select the optimal model parameter adjustment strategy for different system states in complex and ever-changing environments, thereby realizing continuous self-learning and performance improvement of the entire monitoring system's identification and prediction model.

[0032] The early warning and decision support module includes an early warning issuance unit and a report generation unit; When the risk level exceeds the preset threshold, the early warning issuing unit will automatically issue audible and visual alarms, SMS notifications, and platform push notifications. The report generation unit generates geological disaster monitoring reports and emergency plans based on the interpretation and prediction results.

[0033] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent control system for monitoring potential geological hazards using airborne LiDAR, characterized in that, The system includes: Data acquisition module: Real-time acquisition of laser point cloud data, positioning attitude data and surface image data of geological hazard-prone areas through airborne LiDAR sensors, GNSS / IMU system and optical camera, and output of multi-source raw monitoring data; Data processing and fusion module: Receives raw monitoring data from multiple sources, performs noise reduction, smoothing and flight strip adjustment on point clouds, generates digital elevation models and digital orthophotos, and outputs standardized terrain and image datasets. Image intelligent interpretation module: Based on standardized terrain and image datasets, combined with deep learning algorithms, it identifies geological hazard features and extracts boundaries, outputting the location, type, and morphological parameters of potential hazard points; The intelligent disaster risk prediction module: Based on the location, type, and morphological parameters of potential hazards, combined with historical disaster data and real-time environmental factors, it uses random forest algorithm and time series analysis to predict the disaster risk level and occurrence time, and generates a risk prediction report; The dynamic monitoring and control module adaptively adjusts the monitoring frequency, sensor parameters, and flight path based on the risk prediction report, thereby achieving dynamic optimization of monitoring resources. The adaptive learning optimization module dynamically updates the disaster identification and prediction model parameters based on reinforcement learning algorithms by comparing the prediction results with the actual monitoring data, thereby improving the system's identification accuracy and prediction reliability. The early warning and decision support module automatically triggers multi-level early warnings and generates emergency plans when the disaster risk level exceeds a preset threshold.

2. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 1, characterized in that: The data acquisition module includes a LiDAR acquisition unit, a positioning and attitude determination unit, and an image acquisition unit; The LiDAR acquisition unit acquires three-dimensional point cloud data of geological disaster-prone areas in real time through an airborne lidar sensor, and generates a high-density point cloud data stream based on multi-echo technology to penetrate vegetation cover. The positioning and attitude determination unit records the precise position and attitude information of the flight platform in real time through the GNSS / IMU integrated navigation system, and generates a positioning and attitude determination data stream; The image acquisition unit acquires true-color images of the ground surface using a synchronously mounted high-resolution optical camera, and generates an orthophoto data stream.

3. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 1, characterized in that: The data processing and fusion module includes a point cloud preprocessing unit, a digital elevation model / digital orthophoto generation unit, and a multi-source fusion unit. The point cloud preprocessing unit performs noise reduction, smoothing and flight strip adjustment on the original point cloud based on wavelet transform and Kalman filter algorithms to generate clean point cloud data. The digital elevation model / digital orthophoto generation unit generates digital elevation models and digital orthophotos based on purified point cloud data and optical images, through triangulation construction and orthorectification technology. The multi-source fusion unit integrates digital elevation models, digital orthophotos, and positioning data into a standardized terrain and image dataset with a unified coordinate system through spatiotemporal registration and data fusion algorithms.

4. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 3, characterized in that: The point cloud preprocessing unit is based on wavelet transform and Kalman filtering algorithms, and includes the following steps: Primary filtering stage: A moving average filter is used to perform preliminary smoothing on the original point cloud, filtering out high-frequency noise and isolated noise points; Intermediate filtering stage: A threshold denoising algorithm based on wavelet transform is used to adaptively filter point cloud signal components of different frequency bands to separate vegetation and surface reflection signals. Advanced filtering stage: The dynamic point cloud is predicted and corrected using a Kalman filter, and the flight strip adjustment and system error elimination are achieved by combining POS data; Consistency verification phase: The correlation coefficient analysis method is used to verify the consistency between the filtered point cloud and the original data to ensure that the terrain features are completely preserved.

5. The intelligent control system for monitoring potential geological hazards using airborne LiDAR according to claim 1, characterized in that: The image intelligent interpretation module includes a feature extraction unit, a disaster identification unit, and a boundary refinement unit; The feature extraction unit automatically extracts terrain slope, aspect, curvature, and texture features from digital elevation models and digital orthophotos based on a deep learning network; The disaster identification unit identifies geological disaster types such as landslides, collapses, and debris flows through a convolutional neural network classification model, and outputs the location and type labels of potential hazards. The boundary refinement unit uses edge detection and morphological algorithms to optimize the boundaries and calculate morphological parameters of the recognition results.

6. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 1, characterized in that: The intelligent disaster risk prediction module includes a risk assessment unit, a time prediction unit, and an environmental correction unit. The risk assessment unit uses a random forest algorithm to assess the geological disaster risk level based on the morphological parameters of potential hazard points and a historical disaster database. The time prediction unit uses an LSTM time series model to predict the possible time points of a disaster and generates a disaster timeline. The environmental correction unit combines real-time rainfall, soil moisture, and seismic activity data to dynamically correct the prediction model.

7. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 1, characterized in that: The dynamic monitoring and control module includes a route planning unit, a sensor control unit, and a monitoring and scheduling unit. The route planning unit dynamically optimizes the UAV's flight path and altitude using a genetic algorithm based on risk level and terrain complexity. The sensor control unit adaptively adjusts the LiDAR scanning frequency and camera exposure parameters according to vegetation cover and weather conditions. The monitoring and scheduling unit intelligently allocates monitoring resources and task priorities based on prediction results and real-time data.

8. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 1, characterized in that: The adaptive learning optimization module includes a difference analysis unit, a model update unit, and a performance evaluation unit. The difference analysis unit calculates the identification error and prediction deviation by comparing the predicted disaster location with the actual monitoring data; The model update unit dynamically adjusts the weights of the deep learning network and the parameters of the prediction model based on a reinforcement learning mechanism. The performance evaluation unit periodically assesses the model accuracy and system stability, triggering the model reconstruction and optimization process.

9. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 8, characterized in that: The model update unit is based on a reinforcement learning mechanism and includes the following steps: State perception phase: Real-time acquisition of the error between predicted disaster location and actual monitoring data, changes in environmental variables, and system identification accuracy indicators to construct a state vector; Action decision-making stage: Based on the deep Q-network model, the optimal model parameter update strategy is selected according to the current state, including neural network weight adjustment and prediction model parameter correction; Reward feedback phase: Design a reward function based on the improvement rate of recognition accuracy, the accuracy of early warning, and the system response time, and dynamically optimize the strategy through temporal difference learning; Iterative optimization phase: The network weights are gradually updated through multiple rounds of interactive training, and an experience replay mechanism is used to prevent overfitting, so as to achieve continuous self-learning and performance improvement of the model in complex environments.

10. The intelligent control system for monitoring geological disaster risks using airborne LiDAR according to claim 1, characterized in that: The early warning and decision support module includes an early warning release unit and a report generation unit; When the risk level exceeds a preset threshold, the early warning issuing unit automatically issues an audible and visual alarm, an SMS notification, and a platform push notification. The report generation unit generates geological disaster monitoring reports and emergency plans based on the interpretation and prediction results.

Images