CN-122024180-A - Landslide hazard real-time monitoring method integrating time sequence pixel differences

CN122024180ACN 122024180 ACN122024180 ACN 122024180ACN-122024180-A

Abstract

The invention discloses a landslide hazard real-time monitoring method integrating time sequence pixel differences, and relates to the technical field of data processing. The method comprises the steps of establishing a reference image, carrying out geometric registration and radiation normalization on a real-time image, generating a difference image pixel by pixel, setting a self-adaptive threshold value to mark candidate points according to local textures and noise, carrying out communication area analysis on the candidate points, marking a landslide suspected area through an area and intensity threshold value, detecting isolated abrupt pixels and dynamically correcting, establishing a database for triggering cross verification on the repeatedly-occurring isolated points, resolving a depth image through a binocular vision or monocular movement recovery structure, comparing the depth image with a historical depth image to generate a depth difference image, fusing the pixel difference image and the depth difference image to generate a deformation intensity image, and inputting a multi-agent reinforcement learning frame to dynamically optimize unmanned aerial vehicle route planning. Progressive sensing from pixels to three-dimensional space, self-adaptive threshold value, multi-source fusion and dynamic resource scheduling, and accuracy and instantaneity of landslide monitoring are remarkably improved.

Inventors

XIANG BO
JIA YANG
DING YULIN
WANG ZHONGWEN
LIU ZIQIANG

Assignees

四川省公路规划勘察设计研究院有限公司
西南交通大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. A landslide hazard real-time monitoring method integrating time sequence pixel differences is characterized by comprising the following steps: Step S1, acquiring a historical remote sensing image sequence of a target area to establish a reference image, acquiring a current pose for geometric registration for each frame of image acquired in real time, and carrying out radiation normalization according to illumination conditions; Step S2, performing pixel-by-pixel subtraction operation on the registered real-time image and the reference image to generate a pixel difference image, setting a self-adaptive threshold according to the local texture complexity and the historical noise level of the image, and marking the pixels exceeding the threshold as candidate change points; Step S3, analyzing the communication areas of the candidate change points, and calculating the areas of the communication areas, presetting a minimum deformation unit area threshold value, and marking the areas with areas exceeding the threshold value and pixel difference average values exceeding a preset intensity threshold value in the areas as landslide suspected areas; step S4, synchronously detecting isolated abrupt change pixels and selecting a correction strategy to correct according to neighborhood characteristics, establishing a suspicious point database for isolated abrupt change points which repeatedly appear in continuous multiframes, and triggering multi-angle cross verification; S5, resolving a depth map of a suspected landslide region through binocular vision or monocular movement recovery structure technology, and comparing the depth map with a historical depth map to generate a depth difference map; And S6, fusing the pixel difference map and the depth difference map to generate a deformation intensity map, inputting the deformation intensity map as a reward signal into a multi-agent reinforcement learning frame, and dynamically optimizing unmanned aerial vehicle cluster route planning to enable a region with high deformation intensity to obtain higher sampling frequency and resolution.
2. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference is characterized in that step S1 further comprises the steps of constructing a reference image library covering multi-season multi-illumination conditions and storing metadata in a correlated mode, fusing Beidou and inertial navigation data to obtain high-precision pose, dynamically selecting an optimal reference image according to the pose and illumination conditions, conducting rough registration, fine registration and terrain correction hierarchical geometric registration, extracting three-layer radiation normalization through global histogram matching, local homomorphic filtering and invariant features, conducting quality verification on registration normalization results, and feeding back optimization parameters.
3. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference of claim 1 is characterized in that step S2 further comprises the steps of extracting spectrum differences from multispectral images by adopting principal component analysis, spectrum angle measurement or multichannel joint difference values, calculating texture complexity factors through local variance, gradient amplitude or local binary pattern entropy, estimating noise levels by statistics of standard deviation values of historical frame difference values, calculating self-adaptive thresholds, carrying out multiscale pyramid verification on marked candidate points, outputting candidate point metadata and feeding back updated noise models.
4. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference is characterized in that step S3 further comprises the steps of extracting a connected region by adopting an eight-neighborhood connected criterion, calibrating a minimum deformation unit area threshold value according to flying height and resolution, calculating a region difference mean value and standard deviation, judging a landslide suspected region through a region and intensity double threshold value, conducting morphological analysis to remove interference through circularity and topography slope consistency, improving confidence through multi-frame space-time consistency verification, and outputting a suspected region marking graph and metadata.
5. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference of claim 1 is characterized in that step S4 further comprises the steps of setting an isolated point judgment threshold value according to a neighborhood mean value and a standard deviation, analyzing gradient characteristics, symbol distribution and time stability of isolated points, selecting a neighborhood maximum value, a minimum value or mean value to replace according to characteristic analysis results for correction, establishing a suspected point database to record repeatedly-appearing isolated points, triggering multi-spectrum, multi-view, high-resolution and priori knowledge base cross verification when the threshold value is reached, and feeding verification results back to optimize detection parameters and correction strategies.
6. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference is characterized in that step S5 further comprises the steps of dynamically selecting a binocular vision or monocular movement recovery structure mode according to sensor configuration, calculating a dense parallax map or an optical flow field, introducing DEM priori constraint in a texture depletion region, calculating a depth map, calculating depth confidence, generating a depth difference map after registering a current depth map and a historical depth map, extracting a depth change region, performing spatial correlation analysis with a plane suspected region to confirm three-dimensional deformation, extracting change rate and acceleration through multi-period depth time sequence analysis, and outputting three-dimensional deformation information.
7. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference is characterized in that step S6 further comprises the steps of generating a deformation intensity map through weighting fusion according to confidence after normalization of a pixel difference map and a depth difference map, constructing a multi-agent reinforcement learning frame, enabling a state space to comprise the deformation intensity map and an unmanned plane state, enabling an action space to comprise a flight direction, a speed, an altitude and a sampling mode, enabling a reward function to comprehensively consider information acquisition, coverage efficiency, energy consumption, communication and cooperative multi-objective, maximizing long-term cumulative rewards under electric quantity and communication constraint, achieving multi-machine cooperation through communication sharing, potential field collision prevention and dynamic area division, monitoring electric quantity communication in real time and processing emergency events, and evaluating a feedback update strategy network after tasks.
8. The landslide hazard real-time monitoring method based on the time sequence pixel difference fusion of claim 1, further comprising the step of S7, performing image blocking processing by adopting a parallel computing architecture, performing pixel subtraction, connected region analysis, isolated point detection and optical flow calculation in parallel by a GPU or a multi-core CPU, and merging results by a boundary synchronization mechanism.
9. The landslide hazard real-time monitoring method based on the time sequence pixel difference fusion is characterized by further comprising the steps of S8, comprehensively judging plane deformation, verifying three-dimensional deformation and generating multi-level early warning information by means of deformation strength evolution trend, feeding early warning results back to S6, triggering encryption observation, and storing the early warning results and post verification data in a historical reference library for optimizing a threshold model and a scheduling strategy.
10. The landslide hazard real-time monitoring method based on the fusion time sequence pixel difference of any one of claims 1 to 9 is characterized in that steps S1 to S8 form a multiple closed loop feedback system, wherein the multiple closed loop feedback system comprises quality verification feedback optimization benchmark selection and registration parameters of step S1, candidate point metadata feedback updating noise models of step S2, suspected region metadata feedback optimization threshold calibration of step S3, cross verification result feedback optimization detection parameters and correction strategies of step S4, trend analysis feedback adjustment depth change threshold of step S5, evaluation result feedback updating reinforcement learning strategies of step S6, early warning result feedback optimization threshold models and scheduling strategies of step S8.

Description

Landslide hazard real-time monitoring method integrating time sequence pixel differences Technical Field The invention relates to the technical field of data processing, in particular to a landslide hazard real-time monitoring method integrating time sequence pixel differences. Background Landslide disasters have the characteristics of strong burst and large damage, and form serious threats to lives, properties and infrastructure of people. The method is a key means for preventing and reducing disasters, and accurately monitoring landslide deformation in real time and early warning in advance. With the rapid development of unmanned aerial vehicle remote sensing technology, landslide monitoring methods based on unmanned aerial vehicle images have become research hotspots in the field due to the advantages of flexibility, high resolution, relatively low cost and the like. The existing landslide monitoring method based on unmanned aerial vehicle images is mainly divided into the following categories: The first is a displacement measurement method based on feature matching. The method performs feature matching between time sequence images by extracting feature points (such as SIFT, SURF, ORB and the like) in the images, calculates the displacement of the feature points, and further deduces the surface deformation. However, the method has the defects that the feature point extraction depends on the image texture richness, a sufficient number of reliable feature points are difficult to extract in texture-poor areas such as vegetation coverage areas and bare soil areas, the calculation amount of the feature matching process is large, the real-time monitoring requirement is difficult to meet, the displacement measurement result is only sparse point cloud, and the boundary and internal change of a deformation area cannot be comprehensively described. The second category is a change detection method based on deep learning. The method carries out semantic segmentation or change classification on the image by training a convolutional neural network, and identifies a deformation region. However, the method needs a large number of marked samples for training, landslide disasters belong to small probability events, the samples are difficult to obtain, the model generalization capability is limited, the performance fluctuation is large under different landforms and illumination conditions, the network reasoning calculation amount is large, the edge end deployment is difficult, and the real-time processing is difficult to realize. The third category is displacement field measurement methods based on digital image correlation. The method obtains a dense displacement field by calculating the correlation coefficient between the image blocks. However, the method is sensitive to illumination change, the related calculation is invalid due to radiation difference, the measurement precision and the spatial resolution are difficult to be simultaneously considered by the fixed window size, an effective inhibition mechanism for isolated noise is lacking, and a large number of false detection points are easy to generate. The fourth category is monitoring methods based on multi-source data fusion. The method tries to fuse the optical image, liDAR point cloud, inSAR data and the like so as to obtain more comprehensive deformation information. However, space-time references of different data sources are difficult to unify, registration errors directly affect fusion effects, most of fusion strategies are post-processing fusion, real-time online fusion is difficult to achieve, and multi-source data acquisition equipment is high in cost and difficult to popularize and apply on a large scale. The method still has the following common technical problems in practical application: in the sensing dimension, the existing method is mostly limited to planar two-dimensional deformation analysis, and lacks synchronous monitoring capability on elevation changes (such as humps and sinkers). Before unstability, the landslide body often accompanies three-dimensional deformation characteristics such as leading edge bulge or trailing edge subsidence, and the deformation evolution process is difficult to completely describe only by relying on plane displacement information, so that early warning basis is single and confidence coefficient is insufficient. In terms of processing logic, the threshold setting of the conventional method is mostly a fixed value or a global statistic value, and noise distribution of local texture characteristics and time dimension of the image is not fully considered. In the area of complex texture, the fixed threshold value is easy to misjudge the fluctuation of the texture as deformation, and in the area of flat texture, the weak deformation is easy to be submerged by noise. Meanwhile, the isolated noise points and the real deformation points are difficult to effectively distinguish, so that the false detection rate and the omiss