A landslide disaster real-time monitoring method fusing temporal pixel difference

Abstract

This invention discloses a real-time landslide disaster monitoring method that integrates temporal pixel differences, belonging to the field of data processing technology. The method includes: establishing a reference image and performing geometric registration and radiometric normalization on the real-time image; generating a difference map pixel by pixel, and marking candidate points based on adaptive thresholds set according to local texture and noise; performing connectivity analysis on the candidate points, and marking suspected landslide areas using area and intensity thresholds; detecting and dynamically correcting isolated abrupt pixels, and establishing a database of repeatedly occurring isolated points to trigger cross-validation; calculating a depth map through binocular vision or monocular motion recovery, and generating a depth difference map by comparing it with historical depth maps; fusing the pixel difference map and the depth difference map to generate a deformation intensity map, and inputting it into a multi-agent reinforcement learning framework to dynamically optimize UAV flight path planning. The progressive perception from pixel to three-dimensional space, adaptive thresholds, multi-source fusion, and dynamic resource scheduling significantly improve the accuracy and real-time performance of landslide monitoring.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method for real-time monitoring of landslide disasters by integrating temporal pixel differences. Background Technology

[0002] Landslides are characterized by their suddenness and destructiveness, posing a serious threat to people's lives, property, and infrastructure. Real-time and accurate monitoring of landslide deformation and early warning are crucial means of disaster prevention and mitigation. With the rapid development of UAV remote sensing technology, landslide monitoring methods based on UAV imagery have become a research hotspot in this field due to their advantages such as mobility, high resolution, and relatively low cost.

[0003] Existing landslide monitoring methods based on UAV imagery can be mainly classified into the following categories:

[0004] The first category is displacement measurement methods based on feature matching. These methods extract feature points (such as SIFT, SURF, ORB, etc.) from images, perform feature matching between time-series images, calculate the displacement of these feature points, and then infer surface deformation. However, these methods have the following drawbacks: feature point extraction depends on the richness of image texture, making it difficult to extract a sufficient number of reliable feature points in texture-poor areas such as vegetated areas and bare soil areas; the feature matching process is computationally intensive, making it difficult to meet real-time monitoring requirements; and the displacement measurement results are only sparse point clouds, unable to fully characterize the boundaries and internal changes of the deformed area.

[0005] The second category is change detection methods based on deep learning. These methods train convolutional neural networks to perform semantic segmentation or change classification on images and identify deformed regions. However, these methods require a large number of labeled samples for training, while landslides are low-probability events, making sample acquisition difficult; the model's generalization ability is limited, and its performance fluctuates greatly under different terrain and lighting conditions; the network inference computation is large, making edge deployment difficult and hindering real-time processing.

[0006] The third category is displacement field measurement methods based on digital image correlation. These methods obtain dense displacement fields by calculating the correlation coefficients between image blocks. However, these methods are sensitive to changes in illumination, and differences in radiation can cause correlation calculations to fail; a fixed window size makes it difficult to simultaneously achieve both measurement accuracy and spatial resolution; and they lack effective mechanisms for suppressing isolated noise, easily generating a large number of false detection points.

[0007] The fourth category is monitoring methods based on multi-source data fusion. These methods attempt to fuse optical imagery, LiDAR point clouds, InSAR data, etc., in order to obtain more comprehensive deformation information. However, it is difficult to unify the spatiotemporal references of different data sources, and registration errors directly affect the fusion effect; fusion strategies are mostly post-processing fusion, making real-time online fusion difficult to achieve; and multi-source data acquisition equipment is costly, hindering large-scale application.

[0008] The above methods still have the following common technical problems in practical applications:

[0009] In terms of perception, existing methods are mostly limited to two-dimensional deformation analysis and lack the ability to simultaneously monitor elevation changes (such as uplift and subsidence). Before instability, landslides are often accompanied by three-dimensional deformation characteristics such as uplift at the leading edge or subsidence at the trailing edge. Relying solely on planar displacement information is insufficient to fully characterize the deformation evolution process, resulting in a single basis for early warning and insufficient confidence.

[0010] In terms of processing logic, existing methods often use fixed or global statistical thresholds, failing to adequately consider local image texture characteristics and temporal noise distribution. In areas with complex textures, fixed thresholds can easily misjudge texture fluctuations as deformations; in areas with flat textures, subtle deformations are easily masked by noise. Furthermore, isolated noise points are difficult to distinguish from genuine deformation points, leading to high false positive and false negative rates.

[0011] In terms of resource allocation, existing UAV monitoring routes are mostly pre-set fixed routes, making it impossible to dynamically adjust the observation frequency and resolution based on the real-time deformation intensity. There is a coexistence of insufficient monitoring in high-risk areas and resource waste in low-risk areas, making it difficult to maximize monitoring benefits under limited resource constraints.

[0012] In terms of system architecture, existing methods isolate the processing stages and lack a closed-loop feedback mechanism. Information flows unidirectionally between stages such as image preprocessing, change detection, deformation analysis, and early warning decision-making. Processing parameters cannot be adaptively optimized based on subsequent results, making it difficult for system performance to continuously improve over time.

[0013] In view of the above, this application is hereby submitted. Summary of the Invention

[0014] To address the aforementioned problems in existing technologies, a real-time monitoring method for landslide disasters that integrates temporal pixel differences is provided, with the aim of resolving at least one of the above problems.

[0015] The technical solution to achieve the purpose of this invention is as follows:

[0016] This invention provides a method for real-time monitoring of landslide disasters by fusing temporal pixel differences, which includes the following steps:

[0017] Step S1: Collect historical remote sensing image sequences of the target area and establish a reference image that includes the acquisition time, illumination conditions, and UAV pose parameters; for each frame of image acquired in real time, first obtain the current pose through the Beidou navigation system and inertial navigation system on the UAV, and perform geometric registration with the reference image using affine transformation or perspective transformation; at the same time, perform radiometric normalization on the image according to the difference in illumination conditions to eliminate false differences caused by illumination changes.

[0018] Step S2 involves performing pixel-by-pixel subtraction between the registered real-time image and the reference image to generate a pixel difference map ΔI(x,y) = |I_t(x,y) - I_ref(x,y)|. For multispectral or multichannel images, the first principal component after principal component analysis or the calculated spectral angle can be used to measure the difference. An adaptive threshold T(x,y) is set, which is dynamically adjusted based on the local texture complexity and historical noise level of the image. Each pixel is iterated over; if ΔI(x,y) > T(x,y), the pixel is marked as a candidate change point.

[0019] Step S3: Perform connected region analysis on the candidate change points generated in step S2, and calculate the area S_i of each connected region; preset the minimum deformation unit area threshold S_min; if S_i≥S_min and the mean of all pixels in the region exceeds the preset intensity threshold, then the region is determined to be a real deformation region and marked as a suspected landslide area; isolated change points with an area smaller than S_min are regarded as noise or micro-motion interference and ignored.

[0020] In step S4, during the pixel difference map generation process, isolated abrupt change pixels are detected simultaneously, i.e., a single pixel value significantly higher than its neighborhood mean (e.g., exceeding 3 times the standard deviation). For such pixels, one of the following correction strategies is adopted: the first correction method is to replace with the maximum value of the neighborhood, which is suitable for preserving edge abrupt changes; the second correction method is to replace with the minimum value of the neighborhood, which is suitable for suppressing salt-and-pepper noise; the third correction method is to replace with the neighborhood mean, which is suitable for smoothing. The corrected pixels participate in the region clustering verification in step S3. In addition, for isolated abrupt change points that repeatedly appear in multiple consecutive frames of images, a database of suspicious points is established to trigger subsequent multi-angle cross-validation.

[0021] Step S5: For the suspected landslide area identified in Step S3, further extract depth change information; use the binocular vision system on the UAV, or use monocular vision combined with structure-of-motion motion reconstruction technology, to calculate the dense optical flow field between adjacent frames, and then solve the depth map (optical path line) of the surface object; after registering the current depth map with the historical depth map, compare it pixel by pixel to generate a depth difference map ΔD(x,y).

[0022] Step S6: The pixel difference map generated in step S3 and the depth difference map generated in step S5 are fused to generate a deformation intensity map. The value of each pixel in the intensity map represents the overall deformation intensity at that location. The deformation intensity map is used as a reward signal input to the multi-agent reinforcement learning framework to dynamically optimize the flight path planning of the UAV swarm: higher sampling frequencies are allocated to areas with high deformation intensity (e.g., shortening the revisit period and lowering the flight altitude to obtain higher resolution); sparse sampling is used in areas with low deformation intensity; emergency encrypted observation is triggered for areas where deformation intensity rises rapidly; the flight path planning also considers the remaining battery power of the UAVs, communication constraints, etc., to maximize the monitoring benefits under resource constraints.

[0023] Compared with the prior art, the beneficial effects of the present invention include:

[0024] (1) A progressive deformation perception system from pixel level to region level and then to three-dimensional space was constructed, which significantly improved the accuracy and reliability of landslide identification. The pixel-level difference analysis in step S2 can capture the slight color and texture changes on the ground surface and detect deformation precursors 3-5 days earlier than the traditional feature point matching method; the regional clustering verification in step S3 reduces the false alarm rate from 15%-20% in the existing technology to below 3% by using dual thresholds of area and intensity, combined with morphological analysis and multi-frame consistency verification; the planar-three-dimensional dual verification mechanism in step S5 improves the overall accuracy of landslide judgment to over 98% and reduces the false alarm rate to below 1% by analyzing the spatial correlation between depth change and planar deformation.

[0025] (2) An intelligent processing logic with spatial and temporal adaptation was established, which fundamentally solved the technical problem that fixed thresholds are difficult to adapt to complex scenes. Step S2 dynamically calculates the threshold for each pixel based on the local texture complexity and historical noise level, avoiding misjudging natural fluctuations as deformations in textured areas and ensuring that weak deformations are not drowned out by noise in textured areas; Step S4 effectively removes salt-and-pepper noise and impulse interference while preserving edge and texture details, resulting in a significant improvement in image quality; The suspicious point database and multi-angle cross-validation mechanism in Step S4 reduce the misjudgment rate caused by isolated points by more than 80%;

[0026] (3) Dynamic flight path planning for UAV swarms driven by deformation intensity was realized, maximizing monitoring benefits under limited resource constraints. Step S6 fuses planar deformation information and three-dimensional deformation information into a deformation intensity map, which is used as a reward signal input to the multi-agent reinforcement learning framework. This reduces the revisit cycle of high-risk areas from several hours to minutes, increases the monitoring frequency by more than 5 times, reduces invalid coverage in low-risk areas by 60%, and improves overall monitoring efficiency by 3 times. At the same time, the automatic triggering mechanism for areas with rapidly increasing deformation intensity reduces the response time of emergency encrypted observation to seconds, enabling the acquisition of high-resolution data in the early stages of deformation acceleration.

[0027] (4) A multi-loop feedback mechanism was constructed, enabling the system to continuously learn and self-evolve. The quality verification results of step S1 are fed back to the benchmark selection and registration parameter adjustment, so that the preprocessing quality continues to improve with the accumulation of data; the detection results of step S2 are used to update the historical noise model, so that the threshold adapts to sensor aging and environmental changes, and the detection performance can be improved by more than 10% after long-term operation; the post-task evaluation results of step S6 are used to update the reinforcement learning strategy network, so that the route planning capability is continuously optimized; step S8 stores the early warning results and post-verification data into the historical benchmark library, which is used to periodically optimize the adaptive threshold model and parallel scheduling strategy. After the system has been running for 3 months, the registration success rate and radiation consistency can be improved by more than 15%, forming a virtuous cycle of "the more it is used, the smarter it becomes";

[0028] (5) Deep fusion and mutual verification of multi-source heterogeneous data were achieved. Step S1 integrates data from the BeiDou Navigation Satellite System and the Inertial Navigation System to obtain prior information with centimeter-level position and 0.1° attitude accuracy, reducing the registration calculation workload by more than 60%. Step S2 provides various spectral difference measurement methods such as principal component analysis, spectral angle measurement, and multi-channel joint difference, enabling the system to select the optimal detection method for different land cover types and change characteristics, and has higher sensitivity to landslide precursors such as vegetation fading, soil moisture changes, and mineral alteration. Step S5 integrates binocular vision and monocular motion recovery structure dual modes, increasing the success rate of depth data acquisition from 70% in the single mode to more than 95%, and expanding the effective coverage of depth measurement by 30% through prior terrain constraints. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1This is a schematic diagram of the main process of a real-time monitoring method for landslide disasters that integrates temporal pixel differences. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0032] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0033] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0034] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0035] The present invention will be further described in detail below with reference to embodiments.

[0036] like Figure 1 As shown, this invention provides a real-time monitoring method for landslide disasters by fusing temporal pixel differences, which includes the following steps:

[0037] Step S1: Collect historical remote sensing image sequences of the target area and establish a reference image that includes acquisition time, illumination conditions, and UAV pose parameters. For each frame of image acquired in real time, first obtain the current pose through the BeiDou navigation system and inertial navigation system on the UAV, and then perform geometric registration with the reference image using affine transformation or perspective transformation to eliminate pixel offset caused by changes in UAV attitude and ensure that subsequent difference calculations are performed in the same spatial coordinate system. At the same time, perform radiometric normalization on the image according to the difference in illumination conditions to eliminate spurious differences caused by changes in illumination.

[0038] Step S2 involves performing pixel-by-pixel subtraction between the registered real-time image and the reference image to generate a pixel difference map ΔI(x,y) = |I_t(x,y) - I_ref(x,y)|. For multispectral or multichannel images, the first principal component after principal component analysis or the calculated spectral angle can be used to measure the difference. An adaptive threshold T(x,y) is set, which is dynamically adjusted based on the local texture complexity and historical noise level of the image. Each pixel is iterated over; if ΔI(x,y) > T(x,y), the pixel is marked as a candidate change point.

[0039] Step S3: Perform connected component analysis on the candidate change points generated in step S2, and calculate the area (number of pixels) S_i of each connected component; preset the minimum deformation unit area threshold S_min (calibrated according to the UAV flight altitude and image resolution, for example, corresponding to a 0.5m × 0.5m area on the ground); if S_i ≥ S_min and the mean value of all pixels in the area exceeds the preset intensity threshold, then the area is determined to be a real deformation area and marked as a suspected landslide area; isolated change points with an area smaller than S_min are regarded as noise or micro-motion interference and are ignored; this step effectively eliminates misjudgments caused by single pixel noise, birds, shadow movement, etc. through spatial clustering verification;

[0040] In step S4, during the pixel difference map generation process, isolated abrupt change pixels are detected simultaneously, i.e., a single pixel value significantly higher than its neighborhood mean (e.g., exceeding 3 times the standard deviation). For such pixels, one of the following correction strategies is adopted: the first correction method is to replace with the maximum value of the neighborhood, which is suitable for preserving edge abrupt changes; the second correction method is to replace with the minimum value of the neighborhood, which is suitable for suppressing salt-and-pepper noise; the third correction method is to replace with the neighborhood mean, which is suitable for smoothing. The corrected pixels participate in the region clustering verification in step S3. In addition, for isolated abrupt change points that repeatedly appear in multiple consecutive frames of images, a database of suspicious points is established to trigger subsequent multi-angle cross-validation.

[0041] Step S5: For the suspected landslide area identified in Step S3, further extract depth change information; using the binocular vision system on the UAV, or by combining monocular vision with structure of motion recovery (SFM) technology, calculate the dense optical flow field between adjacent frames, and then solve the depth map (optical path line) of the surface object; after registering the current depth map with the historical depth map, compare it pixel by pixel to generate a depth difference map ΔD(x,y); if ΔD(x,y) exceeds a preset threshold, and the overlap between the changed area and the planar deformation area identified in Step S3 exceeds 50%, then confirm that the area has undergone three-dimensional deformation, and the confidence of the landslide is further improved; the depth change information can also be used to estimate the volume change of the landslide body;

[0042] Step S6: The pixel difference map generated in step S3 and the depth difference map generated in step S5 are fused to generate a deformation intensity map. The value of each pixel in the intensity map represents the overall deformation intensity at that location. The deformation intensity map is used as a reward signal input to the multi-agent reinforcement learning framework to dynamically optimize the flight path planning of the UAV swarm: higher sampling frequencies are allocated to areas with high deformation intensity (e.g., shortening the revisit period and lowering the flight altitude to obtain higher resolution); sparse sampling is used for areas with low deformation intensity; emergency encrypted observation is triggered for areas where deformation intensity rises rapidly; the flight path planning also considers the remaining battery power of the UAVs, communication constraints, etc., to maximize the monitoring benefits under resource constraints.

[0043] Step S7: To meet the real-time processing requirements of the above steps, a parallel computing architecture is used to process the image data in blocks. The image is divided into several continuous data blocks by row or column, and each data block maintains continuity in physical storage and logical processing to ensure that deformed areas at the edges of blocks can be completely identified. A GPU or multi-core CPU is used to perform pixel subtraction in step S2, connected component analysis in step S3, outlier detection in step S4, and optical flow calculation in step S5 on each data block in parallel. After each data block is processed, the results are merged through a boundary information synchronization mechanism to avoid fragmentation of deformed areas or repeated calculations caused by block division. The parallel computing architecture increases the processing frame rate to 5 to 8 times that of the original serial method, meeting the real-time processing requirements of over 20fps.

[0044] Step S8 integrates the planar deformation determination from Step S3, the three-dimensional deformation verification from Step S5, and the evolution trend of the deformation intensity map to generate multi-level early warning information: Blue warning (suspected deformation, requires attention); Yellow warning (deformation confirmed, intensive observation recommended); Red warning (rapid deformation, immediate evacuation). The early warning information includes quantitative indicators such as the deformation area boundary, deformation intensity value, depth change, and confidence level. The early warning results are fed back to the flight path planning module in Step S6, triggering intensive observation of high-risk areas. Simultaneously, the early warning results and post-event verification data (whether a landslide occurred, the scale of the landslide) are stored in the historical benchmark database for periodic optimization of the adaptive threshold model in Step S2 and the parallel scheduling strategy in Step S7, forming a closed-loop evolution.

[0045] It should be noted that this invention addresses common technical problems in landslide disaster monitoring, such as single perception dimension, rigid threshold setting, fixed resource allocation, and open-loop system architecture. It proposes a multi-level closed-loop monitoring system driven by temporal pixel differences. This system achieves multi-dimensional and accurate identification of surface deformation by constructing a progressive perception framework from pixel level to regional level and then to three-dimensional space. It solves the problem of distinguishing noise from deformation in complex environments by introducing a dynamic threshold mechanism with both spatial and temporal adaptive perspectives. It achieves dynamic optimization of UAV swarm resources through deformation intensity map-driven multi-agent reinforcement learning. Finally, it endows the system with continuous learning and self-evolution capabilities through multiple closed-loop feedback throughout the entire process.

[0046] It should be noted that the pixel difference map in step S2 provides an initial judgment of planar deformation and serves as the starting point for all subsequent steps; the regional clustering verification in step S3 performs spatial consistency screening on the initial judgment results, outputting high-confidence landslide suspected areas; the optical flow ranging in step S5 performs three-dimensional verification of the suspected areas to confirm whether there are elevation changes; and step S8 generates multi-level early warnings based on the aforementioned progressive evidence. This progressive logic ensures that each judgment step is supported by sufficient evidence, avoiding misjudgments based on a single dimension.

[0047] Step S2 performs difference calculations at the pixel level; step S3 aggregates pixels at the region level, assigning geometric meaning; step S5 maps two-dimensional pixels to three-dimensional depth space, assigning elevation information; step S6 maps the deformation intensity map to physical space, guiding UAV flight path planning. This mapping relationship allows the system's understanding of deformation to gradually upgrade from microscopic pixels to the macroscopic physical world.

[0048] The abnormal pixel correction mechanism in step S4 improves the quality of input data, laying the foundation for subsequent accurate analysis; the parallel computing architecture in step S7 ensures processing speed, enabling high-quality algorithms to run in real time. Together, they achieve both high-quality and fast processing results.

[0049] Step S6 dynamically adjusts the sampling strategy based on the deformation intensity map output from steps S3 and S5, achieving denser observations as deformation intensifies. Newly acquired high-density data is analyzed through steps S2 to S5, updating the deformation intensity map again, forming a closed-loop optimization process of perception → planning → re-perception. This closed loop ensures that monitoring resources are always focused on the areas requiring the most attention.

[0050] Step S8 stores the warning results and post-verification data in a historical benchmark database, which is used to periodically optimize the adaptive threshold of step S2 (e.g., learning the optimal threshold through reinforcement learning) and the parallel scheduling strategy of step S7 (e.g., dynamically adjusting the block size according to the data load). The system continuously evolves over time, and its performance is constantly improved.

[0051] The above-mentioned multiple logical relationships are nested and mutually supportive: causal progression ensures the reliability of decision-making, spatial mapping gives physical meaning, data-computing collaboration balances quality and efficiency, perception-planning closed loop optimizes resource allocation, and data-model evolution enables continuous improvement.

[0052] Furthermore, in some embodiments of the present invention, step S1, which involves constructing a multimodal reference image library and a metadata database, acquiring high-precision pose information in real time, dynamically selecting the optimal reference image, performing layered geometric registration to eliminate pixel offset, performing radiometric normalization to eliminate illumination differences, and performing quality verification and feedback correction, includes the following steps:

[0053] Step S11: Acquire historical remote sensing image sequences of the target area, covering typical scenes under different seasons, lighting conditions, and vegetation cover. A sharpness evaluation algorithm based on the Laplacian operator and a quality assessment algorithm based on signal-to-noise ratio are used to comprehensively score each image, selecting the image with the highest imaging quality as the primary reference image. Simultaneously, a metadata database is established, containing at least the following dimensions: acquisition time, lighting conditions, UAV pose parameters, and image quality indicators. The lighting conditions include solar altitude angle, solar azimuth angle, and cloud cover; the UAV pose parameters include latitude and longitude coordinates, flight altitude, roll angle, pitch angle, and yaw angle; and the image quality indicators include sharpness, contrast, and signal-to-noise ratio. The primary reference image is associated and stored with the metadata database to construct a traceable and searchable reference image library.

[0054] Step S12: For each frame of imagery acquired in real time, centimeter-level latitude and longitude coordinates are obtained through the BeiDou satellite navigation system onboard the UAV, and high-frequency attitude data, including roll angle, pitch angle, and yaw angle, are obtained through the inertial navigation system. Extended Kalman filtering is used to fuse the BeiDou satellite navigation system data and the inertial navigation system data, eliminating drift errors and random noise from individual sensors, and outputting high-precision pose parameters for each frame of imagery. The position accuracy of these high-precision pose parameters is ≤5cm, and the attitude accuracy is ≤0.1°.

[0055] Step S13: Based on the pose parameters and illumination conditions of the real-time acquired images, retrieve the reference image with the highest matching degree from the reference image library. The matching degree calculation adopts a weighted comprehensive scoring method, specifically including: calculating the spatial proximity, which is the Euclidean distance between the real-time image position and the reference image position; calculating the imaging angle difference, which is the absolute value of the difference between the real-time image attitude angle and the reference image attitude angle; and calculating the illumination condition similarity, which is the absolute value of the difference between the real-time image solar altitude angle and the reference image solar altitude angle. After normalizing the above three indicators, a weighted sum is calculated, and the reference image with the highest score is selected as the registration target for the current frame. For scenarios where a suitable reference image cannot be matched, the current frame is added to the reference image library to achieve dynamic expansion of the reference library.

[0056] Step S14: Perform layered geometric registration between the real-time image and the selected reference image, which includes a coarse registration layer, a fine registration layer, and a terrain correction layer.

[0057] The coarse registration layer is based on the pose parameters output by the fusion of the BeiDou satellite navigation system and the inertial navigation system. It uses affine transformation to perform preliminary correction on real-time images, eliminating overall translation, rotation and scale changes caused by changes in the flight attitude of the UAV.

[0058] The fine registration layer extracts feature points from the real-time image and the reference image based on the coarse registration. The feature points are obtained by scale-invariant feature transformation or directional fast rotation descriptor. The random sampling consensus algorithm is used to remove mismatched points, and the perspective transformation matrix or polynomial transformation model is calculated to perform fine correction of local distortion.

[0059] The terrain correction layer is used to introduce a digital elevation model for areas where the terrain undulation exceeds a preset threshold. Based on the ground elevation corresponding to the pixel, differential correction is performed using collinear equations to eliminate projection differences caused by terrain.

[0060] Through the above-mentioned layered registration, the real-time image and the reference image are ensured to achieve pixel-level alignment in the same spatial coordinate system, and the registration accuracy is controlled within 1 pixel.

[0061] Step S15: Perform radiometric normalization on the registered image to eliminate radiometric differences caused by changes in illumination conditions. This includes global normalization, local normalization, and invariant feature enhancement.

[0062] Global normalization uses a histogram matching method to map the grayscale histogram of the real-time image to the histogram distribution of the reference image, thereby eliminating overall brightness differences.

[0063] Local normalization is used for shadow areas or areas with uneven illumination. It employs homomorphic filtering or the Retinex algorithm to decompose the image into illuminance and reflectance components, retaining only the reflectance component to eliminate the influence of uneven illumination.

[0064] Invariant feature enhancement extracts features that are insensitive to changes in illumination. These features include gradient direction and local binary patterns, which serve as supplementary inputs for subsequent difference calculations.

[0065] After normalization, the normalization effect is evaluated using radiation error evaluation indicators, including root mean square error and structural similarity. If the effect does not reach the preset threshold, a new reference image is selected or a multi-reference fusion strategy is adopted.

[0066] Step S16 verifies the quality of the registered and normalized images, including geometric accuracy verification and radiometric consistency verification.

[0067] Geometric accuracy verification involves randomly selecting several checkpoints and calculating the pixel deviation between the real-time image and the reference image. If the deviation exceeds 1 pixel, the process returns to step S14 to adjust the registration parameters or add control points.

[0068] Radiometric consistency verification involves calculating the structural similarity between the normalized image and the reference image. If the similarity is less than 0.95, the process returns to step S15 to adjust the normalization parameters or use multi-reference fusion.

[0069] Once the image passes verification, its pose parameters, registration parameters, and normalization parameters are stored in the metadata database for use in optimizing subsequent reference selection and parameter adjustments.

[0070] It should be noted that step S1 constructs a spatiotemporal benchmark construction system with multiple logical relationships through the above six sub-steps:

[0071] First, the reference image library and metadata database constructed in step S11 provide the data foundation for the entire process, while the high-precision pose parameters obtained in step S12 provide prior constraints for subsequent registration. Step S13 dynamically selects the optimal reference based on the real-time pose obtained in step S12 and the metadata database obtained in step S11, forming a "data → selection" mapping relationship. Step S14 uses the reference selected in step S13 and the pose obtained in step S12 as input, and achieves geometric alignment through three progressive processes: coarse registration, fine registration, and terrain correction. Step S15 performs radiometric normalization based on step S14 to eliminate the influence of illumination. Step S16 verifies the quality of the results from steps S14 and S15, and feeds the verification results back to the parameter adjustments in steps S13, S14, and S15, forming a closed-loop optimization. Simultaneously, step S16 stores the verified images and their parameters into the metadata database to optimize the reference selection strategy in step S13, forming a data accumulation and optimization cycle.

[0072] The above logical relationship upgrades step S1 from a simple image registration to an intelligent preprocessing system with dynamic selection, layered processing, and closed-loop optimization capabilities.

[0073] Existing technologies typically use a fixed single reference image or simply select a reference image in chronological order, without considering the matching degree between real-time acquisition conditions and the reference image. This results in the registration success rate being greatly affected by lighting and seasonal changes. In contrast, this invention constructs a reference image library containing multiple scenes and lighting conditions, and dynamically selects the optimal reference image based on real-time pose parameters and lighting conditions, significantly improving the adaptability and success rate of registration.

[0074] Existing technologies for utilizing pose information mostly rely on coarse positions provided by GPS or entirely on image feature matching. The former has low accuracy, while the latter is computationally intensive and susceptible to regions with poor texture. This invention integrates data from the BeiDou Navigation Satellite System and the Inertial Navigation System, and obtains prior information with centimeter-level position and 0.1° attitude accuracy through extended Kalman filtering. This provides strong constraints for subsequent registration, significantly reducing the computational load of feature matching and improving robustness.

[0075] Existing geometric registration techniques typically employ a single transformation model (such as affine or perspective transformation), which struggles to simultaneously handle overall offset and local terrain distortion. This invention adopts a layered registration strategy, sequentially performing coarse registration (affine transformation), fine registration (perspective transformation / polynomial), and terrain correction (DEM differential correction). This progressively eliminates various geometric errors, improving registration accuracy to within 1 pixel.

[0076] Existing radiation normalization technologies mostly employ global histogram matching, which has limited effectiveness in handling local shadows and uneven illumination. This invention uses a three-layer normalization approach: global histogram matching eliminates overall brightness differences, local homomorphic filtering / Retinex eliminates illuminance unevenness, and invariant feature enhancement provides supplementary information, effectively addressing complex lighting scenarios.

[0077] Most existing technologies lack automated quality verification mechanisms, relying on manual sampling, which is inefficient and cannot provide timely feedback for correction. This invention introduces dual automated verification of geometric accuracy and radiometric consistency, and feeds the verification results back to the parameters of the preceding steps to form a closed-loop optimization, enabling the preprocessing quality to continuously improve with the accumulation of data.

[0078] Furthermore, existing technologies process images in isolation without establishing a related metadata database, making it difficult to support subsequent dynamic selection and parameter optimization. This invention constructs a metadata database containing acquisition conditions, processing parameters, and quality indicators, providing data support for system self-learning and continuous optimization.

[0079] Compared to existing preprocessing methods, in step S1, the hierarchical registration strategy, combined with high-precision pose priors, improves the geometric registration accuracy from 3-5 pixels in existing technologies to within 1 pixel, laying a solid foundation for subsequent pixel-level difference analysis. Dynamic optimal reference selection and three-layer radiometric normalization enable the method to maintain a high registration success rate even under complex conditions such as drastic changes in lighting, shadow occlusion, and terrain undulations, adapting to data acquisition in sunny, cloudy, and different times of day. Utilizing BDS / INS pose prior information significantly narrows the search range for feature matching, reducing registration computation by more than 60%, meeting the real-time processing requirements of UAV edge devices. The closed-loop feedback mechanism continuously optimizes registration and normalization parameters with data accumulation; after three months of system operation, the registration success rate and radiometric consistency can be further improved by more than 15%. The metadata database records the acquisition conditions, processing parameters, and quality indicators for each frame of imagery, providing complete data support for subsequent algorithm optimization and problem backtracking.

[0080] Furthermore, in some embodiments of the present invention, step S2 includes the following steps: multispectral difference feature extraction, local texture complexity calculation, historical noise level estimation, adaptive threshold calculation, candidate change point marking, multi-scale difference verification, candidate change point output, and metadata recording.

[0081] Step S21: For the registered real-time image and the reference image, select the difference measurement method according to the image type:

[0082] For single-band grayscale images, the pixel-by-pixel grayscale difference is directly used to generate a grayscale difference map ΔI_gray(x,y), and its calculation formula is as follows:

[0083] ΔI_gray(x,y)=|I_t(x,y)-I_ref(x,y)|

[0084] Where I_t(x,y) represents the gray value of the real-time image at pixel coordinates (x,y), I_ref(x,y) represents the gray value of the reference image at the same coordinates, and ΔI_gray(x,y) represents the absolute value of the gray value difference of the pixel.

[0085] For multispectral or multichannel images, one or a combination of the following three methods can be used to extract spectral difference features:

[0086] The first method is principal component analysis. Principal component transformation is performed on the multispectral data, and the first principal component with the highest contribution rate is selected. This principal component typically contains more than 80% of the information. The difference plot ΔI_pca(x,y) of the first principal component is calculated using the following formula:

[0087] ΔI_pca(x,y)=|PCA_1(I_t(x,y))-PCA_1(I_ref(x,y))|

[0088] Where PCA_1 represents the first principal component transformation function.

[0089] The second method is the spectral angle method. The multispectral vector of each pixel is treated as a spectral curve, and the spectral angle ΔI_sam(x,y) between the real-time image and the reference image is calculated using the following formula:

[0090] ΔI_sam(x,y)=arccos((I_t·I_ref) / (||I_t||·||I_ref||))

[0091] Where I_t and I_ref are the multispectral vectors of the real-time image and the reference image at pixel (x,y), respectively, · represents the vector dot product, ||·|| represents the vector magnitude, and ΔI_sam(x,y) represents the spectral angle, with a value range of [0,π / 2].

[0092] The third method is the multi-channel joint difference method. After calculating the difference for each band separately, the weighted average or maximum value of the differences for each band is taken to generate a joint difference map ΔI_multi(x,y).

[0093] For specific scenarios, the optimal method or multiple methods are dynamically selected based on the spectral characteristics of ground features to obtain a comprehensive spectral difference map ΔI_spec(x,y). The fusion method can employ a weighted average or a confidence-based selection strategy.

[0094] Step S22: To adaptively adjust the threshold, the texture complexity of the region containing each pixel needs to be quantized. One or a combination of the following metrics can be used:

[0095] The first metric is local variance. Calculate the variance σ²(x,y) of the pixel values ​​within a sliding window centered at pixel (x,y).

[0096] The second metric is the local gradient magnitude. The Sobel operator is used to calculate the mean G(x,y) of the gradient magnitude within the window, reflecting the edge density.

[0097] The third metric is the local binary pattern entropy. The entropy value H(x,y) of the local binary pattern features within the window is calculated; a higher entropy indicates richer texture.

[0098] Normalize the selected index to the [0,1] interval to obtain the texture complexity factor C(x,y).

[0099] Step S23: Establish a statistical model of historical noise levels. For each pixel location (x, y), count the difference values ​​of the past N frames of images at that location (difference from the reference image), and calculate its standard deviation σ_noise(x, y) as an estimate of the noise level at that location.

[0100] For new regions or pixels without historical data, the global noise standard deviation σ_global is used as the initial value. The historical noise level is updated online with new data, using a sliding window or exponentially weighted moving average method to ensure that the estimate can adapt to environmental changes.

[0101] Step S24: Combining the local texture complexity C(x,y) and the historical noise level σ_noise(x,y), calculate the adaptive threshold T(x,y) for each pixel. The calculation formula is as follows:

[0102] T(x,y)=α·σ_noise(x,y)+β·C(x,y)·k

[0103] In this system, α is the basic threshold coefficient, ranging from 3 to 5. Its physical meaning is that the threshold is proportional to the noise level; the higher the noise, the higher the threshold, avoiding misinterpreting noise as change. β is the texture adjustment coefficient, ranging from 0.5 to 2, used to adjust the influence of texture complexity on the threshold. k is the scaling factor, related to the image quantization level, ranging from 1 to 10 for 8-bit images and from 10 to 100 for 16-bit images. In areas with high noise levels, the threshold is increased to suppress noise interference; in areas with complex textures, greater pixel differences are allowed to avoid misinterpreting natural fluctuations in the texture itself as change. For special areas, such as water bodies and shadow areas, a priori masks can be introduced to adjust the α and β parameters, or different calculation models can be used.

[0104] Step S25: Iterate through each pixel (x,y) and compare the comprehensive spectral difference map ΔI_spec(x,y) (or grayscale difference map) with the adaptive threshold T(x,y):

[0105] If ΔI_spec(x,y) > T(x,y), then mark the pixel as a candidate change point, record its difference value, and generate a binary candidate change map M_candidate(x,y) = 1;

[0106] If ΔI_spec(x,y)≤T(x,y), then mark it as a point with no change, and M_candidate(x,y)=0.

[0107] To improve detection reliability, neighborhood consistency is introduced to assist in the judgment: for a marked candidate point, the number of candidate points in its eight neighborhoods is checked. If it is less than the preset number (e.g., 3), it is regarded as an isolated point and is not marked for the time being, to be verified in subsequent multi-frames.

[0108] Step S26: To further reduce false detections, the differences are verified at different scales. A Gaussian pyramid is constructed on the original image. Steps S21 to S25 are repeated on the downsampled low-resolution image to obtain a low-resolution candidate change map. The low-resolution map is upsampled back to its original size using bilinear interpolation and then logically ANDed with the original candidate change map, retaining only the pixels that are marked at all scales. Multi-scale verification can effectively suppress high-frequency noise and subtle texture interference, improving the reliability of the detection results.

[0109] Step S27: Output the candidate change map M_candidate(x,y) and metadata such as the difference value ΔI_spec(x,y) and local threshold T(x,y) for each candidate point, for use in subsequent steps. Simultaneously, store this metadata in a temporary cache to update historical noise levels and optimize subsequent threshold calculations. The metadata includes at least: pixel coordinates, difference value, threshold, texture complexity factor, and detection timestamp.

[0110] It should be noted that step S2 constructs a refined pixel difference detection system through the above seven interrelated sub-steps:

[0111] First, the spectral difference feature extraction in step S21 is the foundation of the entire detection process. It transforms the raw image data into physically meaningful difference metrics, providing input for subsequent analysis. The local texture complexity calculation in step S22 and the historical noise level estimation in step S23 characterize the pixel's features from both the spatial and temporal domains, respectively. Together, they provide the basis for the adaptive threshold calculation in step S24.

[0112] Step S24 is the core sub-step, which combines the outputs of steps S22 and S23. It uses a mathematical model to transform texture and noise information into a unique threshold for each pixel, demonstrating both spatial and temporal adaptability. Step S25, based on the threshold from step S24, discriminates the difference map from step S21, generating candidate change points, which is the direct output of the detection result.

[0113] The multi-scale verification in step S26, as an optional enhancement step, further filters the data based on step S25, improving detection reliability through consistency checks at different scales. Step S27 not only outputs the final result but also feeds back the intermediate data generated during the detection process (such as difference values ​​and thresholds) to the historical noise model in step S23 for model updates, forming a closed-loop optimization.

[0114] The above sub-steps form a progressive relationship: Step S21 → Step S22 / Step S23 → Step S24 → Step S25 → Step S26 → Step S27; a collaborative relationship: Step S22 and Step S23 jointly support Step S24; a feedback relationship: the output of Step S27 is fed back to Step S23. This relationship upgrades Step S2 from a simple pixel-by-pixel subtraction to an intelligent difference detection module with multispectral fusion, spatiotemporal adaptation, multi-scale verification, and closed-loop optimization capabilities.

[0115] Existing technologies for image change detection typically employ fixed global thresholds or simple statistical thresholds (such as the mean plus a certain number of standard deviations). While these methods are applicable in uniform scenes, they are prone to generating numerous false detections in areas with complex textures, and may miss subtle changes in areas with flat textures. This invention introduces a local texture complexity factor and historical noise level estimation to calculate an adaptive threshold for each pixel. This allows the threshold to dynamically change with spatial location and time, significantly improving the accuracy and adaptability of the detection.

[0116] Existing technologies for processing multispectral images often employ single-band difference or simple band stacking, failing to fully utilize spectral information. This invention provides multiple methods for extracting spectral difference features, including principal component analysis, spectral angle measurements, and multi-channel joint difference. Furthermore, it allows for dynamic selection or fusion of the optimal method based on the scene, resulting in higher sensitivity for detecting different types of surface changes (such as vegetation fading, soil moisture changes, and mineral alteration).

[0117] Existing technologies for noise processing typically employ preprocessing filtering methods, such as Gaussian filtering or median filtering. While these methods suppress noise, they also blur edges and details. This invention, by establishing a historical noise level model, directly integrates noise estimation into threshold calculation, avoiding information loss caused by preprocessing filtering, and enabling the threshold to adapt to sensor aging and environmental changes.

[0118] Most existing technologies lack multi-scale verification mechanisms, and single-scale detection is easily affected by high-frequency noise and texture interference. This invention introduces multi-scale pyramid verification to check the consistency of changes at different resolutions, effectively suppressing false alarms and improving the reliability of detection results.

[0119] Existing change detection technologies mostly employ open-loop processing, with detection results not fed back for optimizing detection parameters. This invention uses detection results and intermediate data to update historical noise models, enabling the system to possess self-learning and adaptive capabilities, and its detection performance can be continuously improved over long-term operation.

[0120] Compared to existing change detection methods, in step S2, the combination of multispectral difference measurement and adaptive thresholding improves the accuracy of change detection to over 95% and reduces the false detection rate to below 3%. The advantages of this invention are particularly evident in scenarios where traditional methods are prone to errors, such as vegetated areas, shaded areas, and areas with complex textures. Dynamic thresholding based on local texture and noise levels allows the method to maintain stable performance across different terrains, including farmland, urban buildings, and mountainous forests, eliminating the need for manual parameter adjustments and significantly reducing on-site debugging workload. The provision of multiple spectral difference measurement methods enables the system to select the optimal detection method for different land cover types and change characteristics, exhibiting higher sensitivity to early landslide precursors (such as vegetation stress and soil moisture changes). All operations are pixel-level local computations, allowing for high parallelization; processing 1080P images on embedded platforms can achieve over 30fps, meeting the real-time monitoring needs of UAVs. The online update mechanism of the historical noise model allows the system to adapt to environmental changes, such as sensor aging and seasonal changes in light intensity, resulting in a more than 10% improvement in detection performance after long-term operation. It not only outputs the candidate point locations, but also includes metadata such as difference values, local thresholds, and texture complexity, providing rich features for subsequent region analysis and 3D verification, and improving the overall system's decision reliability.

[0121] Furthermore, in some embodiments of the present invention, step S3 includes the following steps: extraction of connected regions of candidate change points, calibration of minimum deformation unit area threshold, calculation of region intensity features, dual threshold determination and suspected area marking, morphological analysis and noise region removal, multi-frame consistency verification, and suspected area output and metadata recording:

[0122] Step S31: Obtain the binary candidate change map M_candidate(x,y) output in step S2, where pixels with a value of 1 represent candidate change points, and pixels with a value of 0 represent no change points. A connected component labeling algorithm is used to perform cluster analysis on the candidate change points, dividing adjacent candidate change points into the same connected region. The connected component labeling algorithm uses the eight-neighborhood connectivity criterion, that is, if two pixels are adjacent in the horizontal, vertical, or diagonal direction, they are determined to be in the same connected region. The entire candidate change map is traversed, all connected regions are extracted, and a unique identifier ID_i is assigned to each connected region, where i=1,2,…,N, and N is the total number of connected regions.

[0123] For each connected region R_i, count the number of pixels S_i it contains, i.e. the region area. At the same time, record the geometric features of each connected region, such as the bounding rectangle and centroid coordinates, for subsequent analysis.

[0124] Step S32: Preset the minimum deformation unit area threshold S_min. This threshold is comprehensively calibrated based on the UAV flight altitude H, image spatial resolution R, and geological disaster investigation specifications. The relationship between image spatial resolution R (unit: meters / pixel), flight altitude H (unit: meters), and camera focal length f (unit: millimeters) is as follows:

[0125] R = (H × pixel size) / f

[0126] Pixel size refers to the physical size of a single pixel in the camera's image sensor.

[0127] According to landslide disaster investigation specifications, the smallest identifiable deformation unit is defined as a square area with actual ground dimensions of L×L, where L ranges from 0.5m to 2.0m, determined based on the complexity of the terrain and the required monitoring accuracy. The formula for calculating the minimum deformation unit area threshold S_min is:

[0128] S_min=ceil((L / R)²)

[0129] Here, ceil represents the round-up function, ensuring that the threshold covers the smallest deformation unit. For example, when the flight altitude H=100m, the camera focal length f=35mm, and the pixel size=4μm, R≈0.0114m / pixel. If L=0.5m is set, then S_min≈1930 pixels.

[0130] The S_min value can be dynamically adjusted according to different geological conditions and monitoring needs. For highly sensitive areas, the L value can be reduced, and for low-risk areas, the L value can be increased.

[0131] Step S33: For each connected region R_i, calculate the statistical characteristics of the difference values ​​of candidate change points within the region. Step S2 generated the spectral difference value ΔI_spec(x,y) for each candidate point, and based on this, calculated the mean intensity value μ_i of region R_i; at the same time, calculated the standard deviation of intensity σ_i, reflecting the consistency of the changes within the region.

[0132] A preset intensity threshold T_intensity is set. This threshold is determined based on the image noise level and historical deformation data statistics. It is usually set to T_intensity=μ_noise+k·σ_noise, where μ_noise is the global noise mean, σ_noise is the global noise standard deviation, and k is the confidence coefficient, with a value range of 2 to 5.

[0133] Step S34: Perform dual thresholding on each connected region R_i. If the following two conditions are met simultaneously, the region is determined to be a real deformation region and marked as a suspected landslide area:

[0134] Condition 1: Region area S_i ≥ S_min; Condition 2: Region intensity mean μ_i ≥ T_intensity.

[0135] For areas that meet only the area condition but not the intensity condition, they may be large areas of low-intensity noise (such as large areas of slight differences caused by slow changes in illumination) and will not be marked as suspected areas for the time being, but their evolution trend needs to be recorded and monitored. For isolated points or small areas that meet the intensity condition but not the area condition, they are considered as noise or micro-motion disturbances and will be ignored.

[0136] To improve the reliability of the judgment, an auxiliary criterion of regional intensity consistency can be introduced: if the intensity standard deviation σ_i in the region is too large (such as exceeding T_intensity / 2), it indicates that the changes in the region are uneven, and there may be mixed ground features or local interference. Further confirmation is required by subsequent multi-frame verification.

[0137] Step S35: For the initially marked suspected landslide areas, further morphological analysis is performed to eliminate areas that clearly do not conform to the geometric characteristics of a landslide. Morphological analysis includes:

[0138] First, calculate the region's shape complexity index, such as roundness C_i = 4πS_i / P_i², where P_i is the region's perimeter. Natural landslides typically have irregular shapes and low roundness; while disturbances such as birds and moving shadows often form highly rounded clumps. If C_i exceeds a preset threshold (e.g., 0.8), it is considered a suspicious disturbance, and the confidence level needs to be lowered or it should not be marked temporarily.

[0139] Second, calculate the consistency between the principal axis direction of the region and the terrain slope aspect. Use a digital elevation model to obtain the terrain slope aspect of the region and compare it with the principal axis direction of the region's smallest bounding rectangle. If the angle between the two exceeds a preset threshold (such as 45°), it is considered inconsistent with the terrain characteristics and requires further verification.

[0140] Third, for long and narrow areas (length-to-width ratio exceeding 5:1), they may be false positives for linear features (such as roads and rivers), and need to be excluded in conjunction with the feature classification results.

[0141] Step S36: Verify the spatiotemporal consistency between the suspected region marking results of the current frame and the marking results of the previous multiple frames. For each suspected region, establish its time series tracking record and calculate its frequency of occurrence and area change trend in consecutive frames. If the suspected region appears at least M times in K consecutive frames (M≤K) and its area shows an increasing trend, it is determined to be a stable deformation region, and the confidence level is increased; if it appears occasionally and has no trend, it is considered temporary interference and is downgraded or ignored.

[0142] Multi-frame consistency verification effectively suppresses misjudgments caused by transient interference such as birds, vehicles, and cloud shadows, while capturing real deformations that are constantly evolving.

[0143] Step S37: Output the landslide suspected area marker map M_suspect(x,y). The value of each pixel in the map represents the category of its region: 0 indicates a non-suspected area, and a non-zero value indicates that the pixel belongs to a suspected area, with the value being the region ID. Simultaneously, generate metadata records for each suspected area, including at least: region ID, area S_i, mean intensity μ_i, standard deviation of intensity σ_i, centroid coordinates, circumscribed rectangle, circularity, principal axis direction, and multi-frame verification results. This metadata is stored in a temporary database for use in the subsequent 3D verification in step S5 and the early warning output in step S8.

[0144] Step S3 constructs a multi-level spatial clustering verification system through the above seven interrelated sub-steps:

[0145] First, step S31 extracts connected regions from the candidate change points output in step S2, aggregating discrete pixels into geometrically meaningful regions, achieving a leap from pixel-level to region-level analysis. Step S32 calibrates the minimum deformation unit area threshold based on UAV flight parameters and ground resolution, establishing a mapping relationship between the number of pixels and the actual ground size. Step S33 calculates the intensity statistical characteristics of the region, aggregating the spectral difference information from step S2 into an overall intensity representation of the region.

[0146] Step S34 is the core judgment step, which integrates the area information from step S31 and the intensity information from step S33. It uses dual thresholds to simultaneously verify the geometric scale and intensity of change of the region. The two are complementary and indispensable. Relying solely on the area threshold may misjudge large areas with weak changes, while relying solely on the intensity threshold may misjudge isolated strong points. The synergistic effect of the dual thresholds effectively reduces the false alarm rate.

[0147] The morphological analysis in step S35 further filters the data based on step S34, introducing shape and topographic features to eliminate geometric shapes that do not conform to the physical laws of landslides, thus advancing from "statistically valid" to "geologically reasonable". The multi-frame consistency verification in step S36 extends the single-frame analysis to the time domain, distinguishing between continuous deformation and instantaneous interference through spatiotemporal correlation testing, thereby improving the robustness of the judgment.

[0148] Step S37 not only outputs the final result but also records the rich metadata generated during the process, providing data support for subsequent 3D verification, trend analysis, and early warning decision-making. Simultaneously, this metadata can be fed back into the parameter optimization of S2 and S3 themselves, forming a closed-loop improvement.

[0149] The above sub-steps form a progressive relationship: Step S31 → Step S32 / Step S33 → Step S34 → Step S35 → Step S36 → Step S37; a collaborative relationship: Steps S32 and S33 jointly support Step S34, and Step S34 and S35 complement each other; a feedback relationship: the metadata of Step S37 can be used to optimize the threshold calibration of Step S32 and the verification parameters of Step S36. This relationship upgrades Step S3 from a simple connected component analysis to an intelligent region recognition module with multi-dimensional verification, morphological analysis, and spatiotemporal consistency verification.

[0150] Existing technologies typically perform only simple connected region analysis after completing pixel-level change detection, eliminating isolated points by setting a fixed threshold for the number of pixels. While this method can filter out some noise, it lacks in-depth consideration of large-area weak changes, morphological features, and spatiotemporal continuity, resulting in a high misjudgment rate.

[0151] First, existing technologies often use empirical fixed values ​​for setting area thresholds, such as directly setting 100 pixels, without considering the relationship between the drone's flight altitude, image resolution, and the actual ground size. This invention establishes an area threshold calibration model that correlates the number of pixels with the actual ground size, giving the judgment standard a clear physical meaning and allowing for dynamic adjustment based on monitoring accuracy requirements, thus making it more adaptable.

[0152] Second, existing technologies mostly use a single area threshold for judgment, ignoring the consistency of intensity variation within a region. This invention introduces a dual threshold mechanism, simultaneously verifying both the region area and the mean intensity, and supplementing it with the intensity standard deviation to assess the region's uniformity. This effectively distinguishes between large-area weak noise (such as slow changes in illumination) and true deformation, significantly reducing misjudgments.

[0153] Third, existing technologies lack analysis of regional morphology and cannot distinguish the geometric differences between natural landslides and circular disturbances (such as birds or cloud shadows) and linear features (such as roads). This invention introduces morphological indicators such as circularity and consistency of the principal axis direction with the terrain to eliminate areas that clearly do not conform to the geometric characteristics of landslides, making the judgment results more consistent with geological reality.

[0154] Fourth, existing technologies mostly process each frame independently, failing to fully utilize temporal information and exhibiting weak ability to distinguish between transient interference (such as vehicles or animals) and continuous deformation. This invention introduces multi-frame consistency verification, effectively suppressing transient interference through time series tracking and trend analysis, while simultaneously capturing continuously evolving deformations, thus improving the system's robustness and early warning reliability.

[0155] Fifth, existing technologies provide only limited output, offering only regional location information and lacking detailed descriptions of regional characteristics. This invention generates metadata records for each suspected area, including area, intensity, shape, terrain, and multi-frame verification results. This provides rich information support for subsequent 3D verification, mechanism analysis, and early warning decisions, achieving an upgrade from "location identification" to "feature description."

[0156] Compared to existing post-processing methods, step S3, through dual threshold verification of area and intensity, combined with morphological analysis and multi-frame consistency verification, reduces the false positive rate from 15%–20% in existing technologies to below 3%, significantly improving the reliability of deformation identification. The area threshold is correlated with the actual ground dimensions, making the judgment results interpretable, avoiding the uncertainty caused by empirical thresholds, and can be flexibly adjusted according to different monitoring needs, thus broadening its applicability. Morphological analysis introduces indicators such as roundness and terrain consistency, effectively eliminating non-geological interference such as birds, cloud shadows, and vehicles, making the identification results more consistent with the geological characteristics of landslides, laying the foundation for subsequent mechanism analysis. Multi-frame consistency verification extends single-frame static analysis to spatiotemporal dynamic analysis, capturing the development trend of deformation, distinguishing between instantaneous interference and continuous deformation, and providing a more reliable basis for early warning. Multidimensional metadata is generated for each suspected area, not only serving the current judgment but also providing data support for subsequent three-dimensional verification, trend prediction, and early warning classification, improving the overall system's information utilization efficiency. All operations are based on regional-level geometric and statistical calculations, which require minimal computation and can be processed in real time, meeting the needs of drone edge deployment.

[0157] Furthermore, in some embodiments of the present invention, step S4 includes isolated mutation pixel detection, isolated point classification and feature analysis, multi-strategy adaptive correction, suspicious point database construction and update, multi-angle cross-validation triggering, and feedback optimization of detection parameters, comprising the following steps:

[0158] Step S41: During the generation of the pixel difference map ΔI_spec(x,y) in step S2, isolated abrupt pixel detection is performed simultaneously. For each pixel (x,y), a neighborhood window W(x,y) of size (2w+1)×(2w+1) is constructed with it as the center. The window size w can be preset according to the image resolution and noise characteristics, with a value range of 1 to 3. The mean μ_local(x,y) and standard deviation σ_local(x,y) of all pixels in the neighborhood window except the center pixel are calculated. The isolated point determination threshold T_isolated=μ_local(x,y)+k·σ_local(x,y) is set, where k is the sensitivity coefficient, with a value range of 2 to 5, and the default value is 3. If the difference value ΔI_spec(x,y) of the center pixel is greater than T_isolated, then the pixel is determined to be an isolated abrupt pixel and marked as a point to be corrected.

[0159] To improve detection reliability, multi-scale detection can be introduced: detection is performed at multiple window sizes, and if a point is identified as an isolated point at any scale, it is marked. Multi-scale detection can adapt to isolated noise of different sizes.

[0160] Step S42: For each labeled isolated mutation pixel, further analyze its neighborhood features to provide a basis for selecting subsequent correction strategies. Feature analysis includes:

[0161] First, the Sobel operator is used to calculate the gradient magnitude and direction within the neighborhood window to determine whether the isolated point is located on an edge. If there is a clear edge in the neighborhood (the gradient magnitude exceeds a preset threshold), and the isolated point is located in the direction of edge extension, it is likely a sudden change point on a real edge rather than noise.

[0162] Second, analyze the sign difference between isolated points and their neighboring pixels. Count the number of pixels in the neighborhood that are larger than and smaller than the center pixel. If most neighboring pixels are much smaller than the center pixel, they may be bright spots in salt-and-pepper noise; if most are much larger than the center pixel, they may be dark spots.

[0163] Third, calculate the temporal stability of isolated points. Query the suspicious point database (see step S44). If the isolated marker repeatedly appears at this location in the past several frames, it may be a genuine minor deformation or fixed interference.

[0164] Step S43: Based on the feature analysis results of S42, dynamically select a correction strategy to correct isolated mutation pixels. The correction strategy library includes the following three methods:

[0165] The first correction method is neighborhood maximum value replacement. The maximum value of all pixels within the neighborhood window is used as the correction value. This method is suitable for preserving edge abrupt changes; when an isolated point is located on an edge and is an edge enhancement point, using the maximum value can maintain edge strength.

[0166] The second correction method is neighborhood minimum value replacement. The minimum value of all pixels within the neighborhood window is taken as the correction value. This method is suitable for suppressing bright spots in salt-and-pepper noise; when isolated points are identified as bright spot noise, using the minimum value can effectively eliminate them.

[0167] The third correction method is neighborhood mean replacement. The mean (or median) of all pixels within the neighborhood window is used as the correction value. This method is suitable for smoothing; when isolated points have no obvious edge features and atypical salt-and-pepper noise, using the mean can smooth out anomalies.

[0168] The strategy selection rules are as follows: if the isolated point is located at the edge and the gradient direction is consistent, then the maximum value correction is used; if the isolated point is a significant bright or dark spot and the surrounding pixels are uniform, then the minimum value correction is used; otherwise, the mean value correction is used. The strategy selection rules can be trained and optimized based on actual data.

[0169] Step S44: Establish a suspected point database D_suspect to record isolated aberration pixel information that repeatedly appears in multiple consecutive frames of images. Each record in the database contains at least: pixel coordinates (x, y), first occurrence timestamp t_first, most recent occurrence timestamp t_last, occurrence count, history_strategy, and confidence label (0 indicates pending verification, 1 indicates confirmed noise, and 2 indicates confirmed real deformation).

[0170] For the locations in the current frame that are marked as isolated mutation pixels, perform the following update operation:

[0171] If the location already exists in the database, update t_last to the current time, increment count by 1, and record the correction strategy used this time. If the count exceeds the preset threshold N_suspect (e.g., 3 times), multi-angle cross-validation is triggered.

[0172] If the location does not exist in the database, a new record is created, with t_first and t_last both being the current time, count=1, and label=0.

[0173] For records that have not appeared for a long time (the difference between the current time and t_last exceeds the preset expiration time T_expire), they can be automatically cleaned up or archived to control the database size.

[0174] Step S45: When a pixel location is repeatedly marked as an isolated mutation point in multiple consecutive frames and the count reaches a preset threshold N_suspect, multi-angle cross-validation is triggered. Cross-validation includes the following methods:

[0175] First, multispectral cross-validation. Using multispectral information from this location, analyze whether the changes in each band are consistent. If abrupt changes occur only in a few bands, it may be sensor noise; if consistent changes occur in multiple bands, it may represent actual surface changes.

[0176] Second, multi-view cross-validation. Drones or other monitoring equipment are deployed to image the area from different perspectives to obtain three-dimensional observation data. If anomalies are observed at the location from all different perspectives, view-dependent interference (such as shadows or specular reflections) is eliminated.

[0177] Third, high-resolution magnification observation. Use drones to lower their flight altitude or use zoom cameras to perform high-resolution imaging of the area to observe for any minute traces of surface deformation (such as cracks or discoloration).

[0178] Fourth, compare with prior knowledge base. Query the geological knowledge map to determine if there are known easily deformable geological bodies (such as weak interlayers or fault fracture zones) at the location. If so, increase the confidence that it is a real deformation.

[0179] The cross-validation results are used to update the confidence labels in the suspicious point database: if it is confirmed as noise, label=1 is marked, and a fixed correction strategy (such as mean filtering) is applied to this position in subsequent frames; if it is confirmed as a real deformation, label=2 is marked, and no further correction is made in subsequent frames, retaining the original difference value for deformation identification; if it still cannot be confirmed, label=0 is retained, and monitoring continues.

[0180] Step S46: Feed the cross-validation results and long-term statistical data back to the outlier detection stage in step S41 to optimize the detection parameters. This includes:

[0181] Based on the confirmed noise samples, statistical analysis is performed on the statistical characteristics of the noise (such as mean, variance, and spatial distribution), and the sensitivity coefficient k for outlier detection is dynamically adjusted. If the noise detection rate is high, the value of k can be appropriately increased to reduce sensitivity; if true deformations are missed, the value of k can be decreased.

[0182] Based on the confirmed real deformation samples, analyze their neighborhood feature patterns, optimize the strategy selection rules in step S43, and make the correction strategy more accurately protect the real deformation points.

[0183] Step S4 constructs a closed-loop pixel quality enhancement system through the above six interrelated sub-steps:

[0184] First, step S41 performs real-time scanning of the pixel difference map to detect isolated, abruptly changed pixels that may be noise or anomalies. This serves as the input and starting point for the entire step. Step S42 performs feature analysis on these isolated points, extracting information such as neighborhood gradient, sign distribution, and temporal stability to provide a basis for subsequent differential processing.

[0185] Step S43 dynamically selects a correction strategy based on the analysis results of step S42, and performs targeted correction on isolated points. The corrected pixels directly participate in the subsequent regional clustering verification (step S3), ensuring the continuity of the overall processing flow.

[0186] Step S44 establishes a database of suspicious points, tracks and records recurring isolated points, and associates isolated events of a single frame into a time series. Step S45 initiates multi-angle cross-validation when conditions are triggered, comprehensively utilizing multispectral, multi-view, high-resolution, and prior knowledge to make a comprehensive judgment on suspicious points and distinguish noise from real deformation.

[0187] Step S46 feeds back the cross-validation conclusions and long-term statistical data to the detection parameters in step S41 and the strategy selection rules in S43, forming a closed-loop optimization that continuously improves system performance over time.

[0188] The above sub-steps form multiple logical relationships: a causal progression (step S41 → step S42 → step S43 → subsequent processing, step S44 → step S45 → step S46); a collaborative relationship (the feature analysis in step S42 provides the basis for strategy selection in step S43, and the database in step S44 provides the conditions for triggering step S45); a feedback relationship (the output of step S46 optimizes steps S41 and S43); and a spatiotemporal correlation (step S44 introduces the time dimension, and step S45 introduces multiple spatial dimensions). These relationships upgrade step S4 from a simple anomaly correction to a pixel quality enhancement module with intelligent discrimination, adaptive processing, multi-source verification, and continuous optimization capabilities.

[0189] Existing technologies typically employ fixed filtering methods, such as median filtering, mean filtering, or Gaussian filtering, when processing isolated noise points. This results in a uniform processing of the entire image, failing to distinguish between noise and abrupt changes at true edges, easily leading to edge blurring or loss of detail. This invention, by first detecting isolated abrupt changes and then dynamically selecting a correction strategy based on their neighborhood characteristics, achieves precise localization and differentiated processing of anomalies, effectively protecting edges and details while suppressing noise.

[0190] In existing technologies, most methods only process single-frame images, without considering the temporal performance of the same location, and cannot distinguish between frequently occurring fixed interferences (such as high-voltage power line towers and reflective objects) and genuine deformations. This invention constructs a database of suspicious points, tracks repeatedly occurring isolated points, and triggers multi-angle cross-validation, which can effectively identify and eliminate fixed interferences while capturing continuously evolving subtle deformations.

[0191] Existing technologies for cross-validation often rely on manual review or single data supplementation, which is inefficient and highly subjective. This invention designs multiple verification methods, including multispectral, multi-view, high-resolution, and prior knowledge base methods, and automatically triggers verification based on the frequency of suspicious points, thus achieving automation and intelligence in the verification process.

[0192] Existing technologies typically use fixed parameter settings, lacking adaptive adjustment capabilities and making it difficult to adapt to changes in different scenarios and environments. This invention introduces a feedback optimization mechanism, using cross-validation results to dynamically adjust the detection sensitivity coefficient and correct the strategy selection rules, enabling the system to possess self-learning and adaptive capabilities, resulting in continuous performance improvement over long-term operation.

[0193] Compared to existing anomaly handling methods, step S4, through isolated point detection and multi-strategy adaptive correction, effectively removes salt-and-pepper noise and impulse interference while preserving edge and texture details, significantly improving image quality and providing a more reliable data foundation for subsequent deformation recognition. The suspicious point database and multi-angle cross-validation mechanism effectively distinguish between real, minute deformations and fixed interference and recurring noise, reducing the false positive rate caused by isolated points by more than 80%. The feedback optimization mechanism allows detection parameters and correction strategies to automatically adjust with environmental changes, maintaining stable performance under different lighting and terrain conditions without frequent manual intervention. Multi-frame tracking analysis, considering isolated point events in a time series, captures gradually increasing precursors of weak deformations, providing support for early warning. Cross-validation results are fed back to the detection stage, allowing the system to accumulate experience and, after long-term operation, memorize noise patterns and deformation characteristics of specific scenes, continuously improving performance. Isolated point detection and correction are local operations, highly parallelizable, and processing speed meets real-time requirements; the suspicious point database is controllable in size, with low query and update overhead.

[0194] Furthermore, in some embodiments of the present invention, step S5 includes the following steps: depth measurement mode selection and data acquisition, dense optical flow field calculation and disparity map generation, depth map calculation and quality control, historical depth map registration and spatiotemporal benchmark unification, depth difference map generation and significance analysis, plane-depth change correlation verification, multi-frame depth temporal analysis and trend extraction, and depth information output and metadata recording.

[0195] Step S51: For the suspected landslide area identified in Step S3, dynamically select the depth measurement mode based on the sensor configuration of the UAV and the on-site environmental conditions. The depth measurement modes include the following two:

[0196] The first mode is the binocular vision stereo matching mode. When the UAV is equipped with a rigorously calibrated binocular camera system, it simultaneously acquires left and right views to form a stereo image pair. The baseline length B and focal length f of the binocular camera are known parameters. During the acquisition process, the exposure time of the left and right views is ensured to be consistent to reduce the impact of lighting differences on matching accuracy.

[0197] The second mode is the monocular vision motion reconstruction structure mode. When the drone is equipped with only a monocular camera, depth information is reconstructed using motion reconstruction structure technology through a continuous frame sequence during the drone's flight. This mode requires sufficient parallax in the drone's flight trajectory, and typically employs a cross-path or circular path design to ensure multi-view coverage.

[0198] The mode selection rules are as follows: the binocular vision mode is preferred because it has higher depth accuracy and better real-time performance; when the binocular system is unavailable or binocular matching fails due to weather or lighting conditions, it automatically switches to the monocular motion recovery structure mode; for key suspected areas with high precision requirements, both modes can be enabled at the same time, and reliability can be improved through subsequent fusion.

[0199] Step S52: For the selected mode, calculate the dense correspondence between corresponding image pairs and generate a disparity map. Specifically, this includes:

[0200] In binocular vision mode, a deep learning-based stereo matching algorithm (such as PSMNet, GANet) or a traditional optimization algorithm (such as SGM) is used to calculate the disparity map d(x,y) of the left and right views. The value of each pixel in the disparity map represents the difference in horizontal coordinates of corresponding points in the left and right views, in pixels.

[0201] In monocular motion recovery structure mode, optical flow estimation algorithms (such as FlowNet, RAFT, or the traditional Lucas-Kanade pyramid method) are used to calculate the dense optical flow field (u(x,y), v(x,y)) between adjacent frames, where u is the horizontal displacement and v is the vertical displacement. Combining the UAV pose information (obtained from step S12) and camera intrinsic parameters, the depth is calculated using the principle of triangulation.

[0202] To improve the robustness of optical flow calculation, geometric constraints are introduced in texture-poor regions (such as bare soil and water surfaces): the search range of optical flow is constrained by using prior terrain information provided by digital elevation models; or a piecewise smoothing assumption is adopted to fill the depth of textureless regions through variational optimization methods.

[0203] Step S53: Convert the disparity map or optical flow field into a physical depth map D_cur(x,y), representing the distance from the ground object to the camera at the current moment. The conversion formula is as follows:

[0204] For binocular vision mode: D_cur(x,y) = (f×B) / d(x,y)

[0205] Where f is the camera focal length (in pixels), B is the baseline length (in meters), d(x,y) is the disparity value (in pixels), and D_cur(x,y) is the depth value (in meters).

[0206] For monocular motion recovery structure mode: based on the principle of multi-view geometry, the depth value is obtained by solving the triangulation equation. The specific formula is: D_cur(x,y)=g(u(x,y),v(x,y),P_t,P_{t+1}).

[0207] Where g represents the triangulation function based on the pose parameters P_t and P_{t+1}.

[0208] After generating the depth map, quality control is performed. The depth confidence score C_depth(x,y) for each pixel is calculated. Confidence metrics include: matching cost of the disparity map or optical flow field (lower cost, higher confidence), local texture richness (richer texture, higher confidence), and consistency with the neighborhood depth (smaller deviation from the neighborhood mean, higher confidence). A confidence threshold T_conf is set; pixels below the threshold are marked as depth-unreliable regions and excluded from subsequent comparisons.

[0209] Step S54: Read the most recent high-quality depth map D_ref(x,y) for this area from the historical depth database. This depth map has been registered with the reference image and stored in the metadata database of step S1. Register the current depth map D_cur(x,y) with the historical depth map D_ref(x,y) to eliminate the depth map offset caused by UAV pose differences. The registration method adopts feature matching based on the depth map itself or the mapping relationship based on the registered image in step S14: since the real-time image has been registered with the reference image in step S1, the depth map coordinates can be directly mapped using the registration parameters.

[0210] After registration, ensure that D_cur and D_ref correspond pixel-by-pixel in the same spatial coordinate system.

[0211] Step S55: Compare the registered current depth map with the historical depth map pixel by pixel to generate a depth difference map ΔD(x,y), calculated using the following formula:

[0212] ΔD(x,y)=D_cur(x,y)-D_ref(x,y)

[0213] For pixels with a depth confidence level lower than T_conf, ΔD(x,y) is marked as invalid.

[0214] A significance analysis was performed on the depth difference map, including:

[0215] First, we analyze the mean and standard deviation of statistical depth differences to determine the overall trend of change.

[0216] Second, extract areas of significant change, i.e., pixels where the absolute value of ΔD(x,y) exceeds the preset depth change threshold T_depth. T_depth is set according to measurement accuracy and geological requirements, and is usually 2 to 5 times the standard deviation of depth measurement, or 0.1m to 0.5m according to landslide early warning specifications.

[0217] Third, perform connectivity analysis on the significantly changed regions, calculate the area, average depth change, and maximum depth change of each connected region, and generate a set of depth-changed regions R_depth.

[0218] Step S56: Perform spatial correlation analysis between the depth change region R_depth generated in step S55 and the suspected landslide area R_suspect identified in step S3. Calculate the overlap between the two: Overlap = |R_depth∩R_suspect| / |R_suspect|. If the overlap exceeds a preset threshold T_overlap (e.g., 50%), the suspected area is determined to have undergone three-dimensional deformation, and the landslide confidence level is raised to a high level.

[0219] For regions with overlap below the threshold, analyze the reasons: it may be a false detection of the plane (no change in depth) or a failure of depth measurement (such as poor texture leading to unreliable depth). For the former, reduce the confidence of the region; for the latter, trigger the enhanced measurement in step S57.

[0220] Step S57: For areas where 3D deformation has been confirmed, further multi-frame depth temporal analysis is performed. Using multi-period depth maps from the historical depth database, a depth time series D_t(x,y) is constructed for each pixel, t=1,2,…,T. Temporal analysis algorithms (such as linear regression, Kalman filtering, or LSTM) are used to extract depth change trends, including:

[0221] First, the rate of change v_depth(x,y) = dD / dt reflects how fast the deformation is.

[0222] Second, the acceleration a_depth(x,y) = d²D / dt² reflects whether the deformation is accelerating or decelerating.

[0223] Third, the cumulative change ΔD_accum(x,y) = D_T(x,y) - D_1(x,y) reflects the total deformation amplitude.

[0224] For areas where the rate of change continues to increase or the cumulative change exceeds the warning threshold, a three-dimensional deformation warning signal is generated.

[0225] Step S58: Output three-dimensional deformation information, including at least: depth difference map ΔD(x,y), depth change region R_depth and its geometric parameters (area, average depth change, maximum depth change), depth change rate field v_depth(x,y), depth change acceleration field a_depth(x,y), and three-dimensional deformation confidence. Simultaneously, this information is fused with the planar deformation information from step S3 to generate a comprehensive deformation field, which is stored in a spatiotemporal database for use in subsequent route planning in step S6 and early warning output in step S8.

[0226] Step S5 constructs a multi-dimensional, multi-modal, and multi-level three-dimensional deformation information extraction system through the above eight interrelated sub-steps:

[0227] First, step S51 dynamically selects the optimal depth measurement mode based on hardware conditions and the environment, demonstrating the intelligence of resource adaptation. Steps S52 and S53 sequentially complete dense correspondence calculation and depth resolution to form the current depth map, which is the foundation for 3D information extraction. Step S54 registers the current depth map with historical depth maps to establish a unified spatiotemporal benchmark.

[0228] Step S55 generates a depth difference map and performs a significance analysis to identify areas where the depth changes. This is the core output of 3D deformation recognition. Step S56 correlates and verifies the depth change areas with suspected planar areas. The two corroborate each other, forming a planar-3D dual confirmation mechanism, which significantly improves the reliability of the judgment.

[0229] Step S57 introduces a time dimension, performing temporal analysis and trend extraction on depth changes, upgrading from a single change to a dynamic evolution process, enabling the identification of dangerous precursors such as accelerated deformation. Step S58 integrates all information, outputting rich 3D deformation features to provide data support for subsequent decision-making.

[0230] The above sub-steps form multiple logical relationships: causal progression (step S51 → step S52 → step S53 → step S54 → step S55 → step S56 → step S57 → step S58); collaborative relationships (collaboration between binocular and monocular modes in step S52, collaboration between depth map and confidence level in step S53, collaboration between plane and depth in step S56); feedback relationships (trend analysis in step S57 can be fed back to threshold adjustment in step S55, and the output of step S58 can be used to optimize mode selection in step S51); and spatiotemporal correlation relationships (spatial registration is achieved in step S54, and temporal analysis is achieved in step S57, forming a complete four-dimensional perception capability). These relationships upgrade step S5 from a simple depth map comparison to an intelligent three-dimensional deformation perception module with mode adaptation, plane-stereo collaboration, and temporal evolution analysis.

[0231] Existing technologies for 3D deformation monitoring typically employ a single depth measurement method, such as fixed binocular vision or periodic laser scanning, lacking the ability to dynamically adjust the measurement mode according to site conditions. When a single method fails (e.g., binocular vision fails to match in a weakly textured area), depth information cannot be acquired, leading to monitoring interruption. This invention integrates both binocular vision and monocular motion-reconstruction structure (MRS) modes, intelligently switching or fusing them according to environmental conditions to ensure depth data acquisition in various scenarios, thus improving the system's robustness and adaptability.

[0232] Existing technologies for depth change detection typically compare two depth maps directly, ignoring the quality differences between the depth maps themselves, leading to a large number of false detections in low-confidence regions. This invention introduces a depth confidence evaluation mechanism to mark and exclude depth values ​​in regions with low matching costs and poor texture, significantly reducing noise and false detection rates in depth difference maps.

[0233] Existing landslide monitoring technologies often process planar deformation and three-dimensional deformation separately, using imagery for planar deformation and LiDAR or InSAR for three-dimensional deformation, lacking collaborative verification between the two. This invention achieves mutual verification of planar and depth information through spatial correlation analysis between suspected planar areas and depth change areas. Three-dimensional deformation is only confirmed when the two are consistent, effectively avoiding the problem of misjudgment from a single data source.

[0234] Existing deep change analysis techniques mostly involve two-phase comparisons, lacking a temporal characterization of the deformation process and making it difficult to identify dangerous signals such as accelerated deformation. This invention introduces multi-frame deep temporal analysis to extract dynamic features such as change rate, acceleration, and cumulative change, enabling the capture of deformation development trends and providing richer evidence for early warning.

[0235] Existing technologies often overlook the challenge of texture-poor regions in depth measurement accuracy. This invention optimizes depth calculation in weakly textured regions by introducing prior constraints from a digital elevation model and a piecewise smoothing assumption, thus expanding the applicability of depth measurement.

[0236] Compared with existing 3D deformation monitoring technologies, step S5 features adaptive switching between binocular and monocular dual modes, ensuring depth information acquisition even under complex conditions such as weak texture and varying lighting. The success rate of depth data acquisition increases from 70% in single-mode to over 95%. A depth confidence assessment mechanism is introduced to effectively eliminate low-quality depth pixels, improving the signal-to-noise ratio of the depth difference map by more than three times and reducing the false detection rate to below 2%. Spatial correlation verification cross-references suspected planar areas with depth change areas, increasing the overall accuracy of landslide detection to over 98% and reducing the false alarm rate to below 1%. Multi-frame temporal analysis provides dynamic features such as change rate and acceleration, enabling early identification of accelerated deformation and issuing warning signals an average of 7 to 15 days earlier than traditional two-phase comparison methods. Through prior terrain constraints and piecewise smoothing optimization, the effective coverage of depth measurement is expanded by 30%, solving the failure problem of traditional stereo vision in areas such as bare soil and water bodies. It not only outputs depth differences, but also provides multi-dimensional features such as confidence level, rate of change, and acceleration, providing more comprehensive data support for landslide mechanism analysis and early warning classification.

[0237] Furthermore, in some embodiments of the present invention, step S6 includes the following steps: multi-source deformation information fusion and deformation intensity map generation, multi-agent reinforcement learning framework construction, dynamic route planning and sampling strategy optimization, multi-machine collaboration and conflict avoidance, resource constraint management and emergency handling, and closed-loop feedback and strategy evolution.

[0238] Step S61: The pixel difference map ΔI_spec(x,y) output from step S3 is fused with the depth difference map ΔD(x,y) output from step S5 to generate a deformation intensity map S(x,y) that comprehensively reflects the surface deformation intensity. The fusion strategy employs an adaptive weighted fusion method, dynamically adjusting the weights based on the confidence level and dimensional differences between the two difference maps. Specifically, it includes the following sub-steps:

[0239] First, the pixel difference map and depth difference map are normalized, mapping their values ​​to a unified [0,1] interval. The formula for normalizing the pixel difference map ΔI_norm(x,y) is as follows:

[0240] ΔI_norm(x,y)=ΔI_spec(x,y) / I_max

[0241] Where I_max is the maximum value of the pixel difference map or a preset saturation threshold. The formula for the depth difference map normalization ΔD_norm(x,y) is:

[0242] ΔD_norm(x,y)=|ΔD(x,y)| / D_max

[0243] Where D_max is the maximum value of the depth difference map or a preset saturation threshold, and the absolute value is taken to reflect the magnitude of the depth change.

[0244] Second, the confidence score C_I(x,y) of the pixel difference map is derived from the local texture complexity and historical noise level in the adaptive threshold discrimination process of step S2; the confidence score C_D(x,y) of the depth difference map is derived from the depth confidence score C_depth(x,y) of step S53. The fusion weights are obtained after normalizing both.

[0245] Third, a weighted fusion method is used to generate deformation intensity values. For regions with only pixel difference information but no depth information (such as those marked as invalid due to low depth confidence), the pixel difference value is directly used as the intensity value, i.e., S(x,y)=ΔI_norm(x,y). For regions with rich depth information but low pixel difference values ​​(such as those where deep deformation has not yet caused changes in surface texture), the contribution of depth information is retained.

[0246] Fourth, the generated deformation intensity map undergoes spatial smoothing and enhancement. Gaussian filtering is used to eliminate isolated noise, while contrast stretching is used to enhance the distinguishability of high-risk areas. The final deformation intensity map S(x,y) is output, where the value of each pixel represents the overall deformation intensity at that location, ranging from [0,1]. A larger value indicates a more significant deformation and a higher risk.

[0247] Step S62: Construct a multi-agent reinforcement learning framework for UAV swarms to dynamically optimize flight path planning. This framework includes the following core elements:

[0248] First, the state space includes global state and local state. The global state is the deformation intensity map S_t(x,y) at the current moment, which reflects the deformation distribution of the entire monitoring area; the local state is the state of each UAV, including the current position p_i(t), remaining battery power e_i(t), communication quality q_i(t), and current mission mode m_i(t) (cruising / encryption / return to home).

[0249] Second, the action space for each UAV includes: flight direction θ_i(t) (continuous or discrete), flight speed v_i(t) (continuous or discrete), flight altitude h_i(t) (affecting image resolution), and sampling mode a_i(t) (normal sampling, high-resolution sampling, full-band sampling). The action space design must meet the UAV's dynamic constraints to ensure feasibility.

[0250] Third, the reward function is the core of guiding the agent to learn and optimize the strategy. It is designed as a multi-objective weighted sum, including information collection reward R_info, coverage efficiency reward R_cover, energy consumption penalty R_energy, communication constraint penalty R_comm, and cooperation reward R_coop.

[0251] Information collection reward R_info encourages drones to collect information from areas with high deformation intensity. Coverage efficiency reward R_cover encourages drones to quickly cover large areas and avoid redundant coverage. It uses information entropy increment or area novelty as a metric. Energy consumption penalty R_energy penalizes flight energy consumption, related to flight distance, altitude change, and hovering time, guiding energy conservation. Communication constraint penalty R_comm penalizes drones when they fly out of communication range or when signal quality falls below a threshold, ensuring the feasibility of swarm collaboration. Collaboration reward R_coop encourages drones to cooperate and avoid conflicts and redundant coverage. It can use a potential field method or a differential reward mechanism.

[0252] Fourth, multi-agent learning algorithms. A centralized training and distributed execution framework, such as MAPPO or QMIX, is adopted. During the training phase, the central server collects the experience trajectories of all UAVs and uses global information to perform joint policy optimization; during the execution phase, each UAV makes decisions based solely on local observations (its own state and local deformation intensity map) to achieve real-time response.

[0253] S63: Based on the aforementioned multi-agent reinforcement learning framework, each UAV selects the optimal action at each decision moment (e.g., after each frame acquisition or at each interval Δt) based on its current state, generating the flight trajectory for the next stage. The specific optimization objective is:

[0254] Under constraints of remaining power, communication, and flight performance, the goal is to maximize long-term cumulative rewards, i.e., maximize monitoring benefits. These benefits are reflected in: higher sampling frequency and resolution in high-risk areas, maintaining basic coverage in low-risk areas, while minimizing energy consumption and communication interruptions.

[0255] Step S64: When performing flight path planning in a drone swarm, multi-drone collaboration must be considered to avoid conflicts and redundant coverage. This includes:

[0256] First, a communication-based collaborative mechanism. The drone broadcasts its own position, current mission area, and planned trajectory in real time via its onboard communication module. Upon receiving the broadcast, each drone updates its local neighborhood information. If a potential conflict is detected (such as intersecting flight paths or approaching the same area), the drones adjust their actions through negotiation: adjusting flight direction, altitude, or speed to maintain a safe distance.

[0257] Second, a collision avoidance strategy based on potential fields. When selecting actions, an artificial potential field penalty term is introduced. When the distance between the two machines is less than a safety threshold, a repulsive force is applied to guide separation.

[0258] Third, dynamic task area division. Based on the current deformation intensity map and the location of each UAV, a region segmentation algorithm (such as Voronoi diagram) is used to dynamically allocate the main responsibility area for each UAV, reducing overlapping coverage and improving overall efficiency. The region segmentation results are input into the policy network as part of the local observations.

[0259] Step S65: During the route planning process, monitor the remaining battery power e_i(t) and communication quality q_i(t) of each UAV in real time to ensure mission feasibility and safety.

[0260] First, power constraint management. Set a power safety threshold E_safe. When e_i(t) is lower than the threshold, the drone will automatically trigger the return-to-home procedure and hand over the current task to a neighboring drone. When multiple drones are low on power at the same time, priority will be given to ensuring coverage of high-risk areas.

[0261] Second, communication constraint management. A communication quality threshold Q_min is set. When q_i(t) is lower than the threshold, the UAV attempts to adjust its position to restore communication. If it cannot restore communication, it will autonomously execute a simplified route (such as returning along the original path) according to preset rules until communication is restored.

[0262] Third, emergency response. When an emergency encrypted observation is triggered, if there are no drones with sufficient power or communication capabilities available, a backup drone will be dispatched from the base station to provide support, or the mission priority of other drones will be adjusted to ensure timely response to high-risk areas.

[0263] Step S66: Compare the actual monitoring data (such as collected images and detected deformed areas) after each task execution with the planned expectations to evaluate the efficiency of route planning. Evaluation indicators include: high-risk area coverage, revisit cycle compliance rate, energy efficiency, and early warning timeliness. The evaluation results are fed back to the multi-agent reinforcement learning framework as new training samples to periodically update the policy network, enabling the system to continuously optimize based on accumulated experience.

[0264] Meanwhile, metadata such as the evolution of deformation intensity maps and UAV response effects are stored in the database to optimize the design of fusion weights, threshold parameters, and reward functions, thereby achieving closed-loop evolution of the entire system.

[0265] Step S6 constructs an intelligent dynamic flight path planning system for UAV swarms through the above six interrelated sub-steps:

[0266] First, step S61 fuses the planar deformation information (pixel difference map) and the three-dimensional deformation information (depth difference map) into a unified deformation intensity map, which forms the basis for all subsequent decisions. The fusion process considers the confidence levels of both types of information to ensure the reliability of the intensity map.

[0267] Step S62 constructs a multi-agent reinforcement learning framework, using the deformation intensity map as a core component of the reward signal, and establishes a causal logic of "deformation intensity-driven planning". The state space includes global deformation information and the individual state of the UAV, the action space covers flight control and sampling modes, and the reward function integrates multiple objectives such as information acquisition, efficiency, energy consumption, communication, and cooperation, forming a complete definition of the optimization problem.

[0268] Step S63 implements dynamic route planning within the reinforcement learning framework, transforming the hierarchical results of the deformation intensity map into specific sampling strategies (frequency, resolution), and introducing an emergency triggering mechanism to achieve a dynamic response of "the stronger the deformation, the denser the observation".

[0269] Step S64 introduces a multi-machine collaboration and conflict avoidance mechanism to ensure that the cluster maintains overall coordination while pursuing its own goals, avoiding resource waste and security risks, and embodying the collaborative logic of "individual-group".

[0270] Step S65 manages resource constraints such as power and communication, and handles emergencies to ensure the feasibility of the planning scheme under actual physical constraints, reflecting the balance logic of "optimization-constraint".

[0271] Step S66 forms a closed-loop feedback loop through post-task evaluation and strategy updates, enabling planning capabilities to continuously improve with data accumulation, reflecting the growth logic of "execution-learning-evolution".

[0272] Existing technologies for drone monitoring route planning mostly employ pre-set fixed routes or trigger-based adjustments based on simple rules (such as repeating scans if deformation is detected). Fixed routes cannot adapt to dynamically changing deformation risks, resulting in insufficient monitoring in high-risk areas and wasted resources in low-risk areas; simple rules lack global optimization, making it difficult to maximize efficiency under multi-drone collaboration and resource constraints.

[0273] It should be noted that technical features that are not fully explained will be addressed using conventional technical methods.

[0274] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A landslide disaster real-time monitoring method fusing temporal pixel differences, characterized in that, Includes the following steps: Step S1: Collect historical remote sensing image sequences of the target area to establish a reference image; for each frame of image acquired in real time, obtain the current pose for geometric registration and perform radiometric normalization according to the illumination conditions; Step S2: The registered real-time image and the reference image are subtracted pixel by pixel to generate a pixel difference map; the local texture complexity is a texture complexity factor calculated based on the local variance, gradient magnitude or local binary mode entropy; the historical noise level is the noise level estimated based on the standard deviation of the difference value at the pixel position in the historical frame; an adaptive threshold is set based on the local texture complexity of the image and the historical noise level, and pixels exceeding the threshold are marked as candidate change points. Step S3: Perform connectivity analysis on the candidate change points and calculate the area of ​​each connectivity region; A minimum deformation unit area threshold is preset, and areas whose area exceeds the threshold and whose average pixel difference within the area exceeds the preset intensity threshold are marked as suspected landslide areas; Step S4: Simultaneously detect isolated mutation pixels and select a correction strategy based on neighborhood features for correction. The isolated mutation pixels are isolated mutation pixels in the pixel difference map generated in step S2. The corrected pixels participate in the regional clustering verification in step S3. For isolated mutation points that appear repeatedly in multiple consecutive frames, establish a suspicious point database and trigger multi-angle cross-verification. It comprehensively utilizes multispectral, multi-view, high resolution and prior knowledge to comprehensively judge suspicious points and distinguish noise from real deformation. Step S5: For the suspected landslide area, the depth map is calculated using binocular vision or monocular motion reconstruction technology, and a depth difference map is generated by comparing it with the historical depth map. The historical depth map is a depth map that is stored in the historical depth database in advance. Step S6: Normalize the pixel difference map and the depth difference map respectively, and then weight and fuse them to generate a deformation intensity map; The deformation intensity map is used as a reward signal input to the multi-agent reinforcement learning framework. The reward signal includes information acquisition reward calculated based on the deformation intensity map. The flight path planning of the UAV swarm is dynamically optimized so that areas with high deformation intensity can obtain higher sampling frequency and resolution. 2.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, characterized in that, Step S1 further includes: constructing a reference image library covering multiple seasons and lighting conditions and storing associated metadata; fusing BeiDou and inertial navigation data to obtain high-precision pose; dynamically selecting the optimal reference image based on pose and lighting conditions; performing layered geometric registration through coarse registration, fine registration, and terrain correction; performing three-layer radiometric normalization through global histogram matching, local homomorphic filtering, and invariant feature extraction; and verifying the quality of the registration and normalization results and feeding back optimization parameters. 3.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, characterized in that, Step S2 further includes: extracting spectral differences for multispectral images using principal component analysis, spectral angle measures, or multi-channel joint difference; calculating the texture complexity factor using local variance, gradient magnitude, or local binary mode entropy; estimating the noise level by statistically analyzing the standard deviation of historical frame difference values; calculating the adaptive threshold; performing multi-scale pyramid verification on the marked candidate points; outputting candidate point metadata and feeding back to update the noise model.

4. The method for real-time monitoring of landslide disasters based on time-series pixel differences according to claim 1, characterized in that, Step S3 further includes: extracting connected regions using the eight-neighbor connectivity criterion; calibrating the minimum deformation unit area threshold based on flight altitude and resolution; calculating the mean and standard deviation of regional differences; determining suspected landslide areas using dual thresholds of area and intensity; removing interference through morphological analysis of circularity and topographic aspect consistency; improving confidence through multi-frame spatiotemporal consistency verification; and outputting a suspected area marker map and metadata. 5.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, characterized in that, Step S4 further includes: setting an isolated point determination threshold based on the neighborhood mean and standard deviation; analyzing the gradient characteristics, sign distribution, and temporal stability of isolated points; selecting the neighborhood maximum, minimum, or mean value for correction based on the feature analysis results; establishing a suspicious point database to record repeatedly occurring isolated points; triggering multispectral, multi-view, high-resolution, and prior knowledge base cross-validation when the threshold is reached; and feeding back the validation results to optimize detection parameters and correction strategies. 6.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, characterized in that, Step S5 further includes: dynamically selecting binocular vision or monocular motion recovery structure mode according to sensor configuration; calculating dense disparity map or optical flow field, and introducing DEM prior constraints in texture-poor regions; solving the depth map and calculating depth confidence; generating a depth difference map after registering the current depth map with historical depth maps; extracting depth change regions and performing spatial correlation analysis with suspected planar regions to confirm 3D deformation; extracting change rate and acceleration through multi-period depth time series analysis; and outputting 3D deformation information. 7.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, characterized in that, Step S6 further includes: normalizing the pixel difference map and depth difference map and then fusing them according to confidence weights to generate a deformation intensity map; constructing a multi-agent reinforcement learning framework, with the state space including the deformation intensity map and the UAV state, the action space including flight direction, speed, altitude and sampling mode, and the reward function comprehensively considering multiple objectives such as information collection, coverage efficiency, energy consumption, communication and cooperation; maximizing long-term cumulative rewards under the constraints of power and communication; achieving multi-aircraft cooperation through communication sharing, potential field collision avoidance and dynamic region partitioning; real-time monitoring of power and communication and handling of emergency events; and post-task evaluation feedback to update the policy network. 8.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, wherein, It also includes step S7: using a parallel computing architecture to process the image into blocks, performing pixel subtraction, connected region analysis, outlier detection and optical flow calculation in parallel through GPU or multi-core CPU, and merging the results through a boundary synchronization mechanism. 9.The landslide disaster real-time monitoring method of fusing temporal pixel difference according to claim 1, characterized in that, It also includes step S8: generating multi-level early warning information by comprehensively considering planar deformation determination, three-dimensional deformation verification, and deformation strength evolution trend; The early warning result is fed back to step S6 to trigger encrypted observation, and is stored in the historical benchmark library along with the post-verification data for optimizing the threshold model and scheduling strategy.

10. A method for real-time monitoring of landslide disasters based on fusion of temporal pixel differences according to any one of claims 1 to 9, characterized in that, Steps S1 to S8 form a multi-loop feedback system: Step S1 provides feedback on quality verification to optimize the selection of the benchmark and registration parameters; Step S2 provides feedback on candidate point metadata to update the noise model; Step S3 provides feedback on suspected area metadata to optimize the threshold calibration; Step S4 provides feedback on cross-validation results to optimize the detection parameters and correction strategies. Step S5 provides feedback on trend analysis to adjust the depth change threshold; Step S6 provides feedback on evaluation results to update the reinforcement learning strategy; Step S8 provides feedback on early warning results to optimize the threshold model and scheduling strategy.

Images