A landslide disaster hidden point intelligent identification method based on unmanned aerial vehicle multispectral remote sensing

Abstract

The application discloses a landslide disaster hidden danger point intelligent identification method based on unmanned aerial vehicle multispectral remote sensing, relates to the technical field of geological disaster monitoring and unmanned aerial vehicle remote sensing intelligent identification, and comprises the following steps: 1) data acquisition and preprocessing, 2) trajectory disturbance feature mining, 3) spectral feature enhancement and space-time filtering, 4) double-mode coupling analysis and intelligent identification, 5) closed-loop verification and flight route optimization; in the application, the original value of trajectory offset is reserved, a fusion data set is constructed, the problems of data dimension fragmentation and one-way error processing are solved, and the data utilization rate is improved; through multidisciplinary cross modeling and coupling knowledge base construction, the subject blank is filled, and an unmanned aerial vehicle-environment-geology dynamic mutual feedback mechanism is established; with the help of double-mode feature enhancement, counterfactual reasoning and flight route optimization, the identification efficiency of early concealed landslide hazards is effectively improved, the hazard development stage can be distinguished, and the high-precision and early identification demand is met.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and UAV remote sensing intelligent identification technology, specifically a method for intelligent identification of landslide hazard points based on UAV multispectral remote sensing. Background Technology

[0002] Landslides, as a highly destructive geological hazard, are widely distributed in mountainous areas, reservoir perimeters, and seismically active zones worldwide, posing a serious threat to people's lives and property, infrastructure such as transportation and water conservancy, and the ecological environment. Therefore, achieving early and accurate identification of landslide hazard points is of vital practical significance for early warning, formulating prevention and control strategies, and carrying out engineering remediation.

[0003] Traditional methods for identifying landslide hazards mainly rely on manual patrols and single-point sensor monitoring. However, manual patrols are greatly limited by terrain conditions, making it difficult to achieve comprehensive coverage in complex areas such as high mountains, canyons, and drawdown zones. Furthermore, they are inefficient, highly subjective, and prone to overlooking potential hazards. While single-point sensors can achieve continuous monitoring of local areas, their coverage is limited, making it difficult to meet the needs of large-scale hazard investigation. In addition, data acquisition costs are high, and the timeliness is poor, making it impossible to capture the dynamic changes of landslide hazards in a timely manner.

[0004] With the development of remote sensing technology, UAV multispectral remote sensing technology has gradually become an important means of landslide disaster monitoring due to its advantages such as flexibility, high resolution, low cost, and repeatable monitoring. As a data acquisition carrier, UAVs can penetrate into complex terrain areas that are difficult for humans to reach. The multispectral sensors they carry can acquire reflectance / radiation information of surface targets in multiple bands such as visible light, near-infrared, and short-wave infrared. Through specific spectral combinations such as normalized difference vegetation index (NDVI) and moisture index, they can sensitively characterize microenvironmental anomalies in landslide hazard areas, such as rock fracturing, soil moisture changes, and vegetation withering, providing a rich data source for hazard identification.

[0005] Although UAV multispectral remote sensing technology has shown promising prospects in landslide monitoring, existing methods for identifying landslide hazard points based on this technology still have systemic limitations. These limitations are interconnected and collectively restrict the accuracy of identifying early and hidden hazards: existing technologies have failed to effectively mine the dynamic correlation information between "UAV-environment-geology." The core issues lie in the limitations of understanding data correlation, the singularity of the technical framework, and the lack of interdisciplinary collaboration, specifically manifested in three mutually influential levels:

[0006] At the technical implementation level, there are problems of fragmented data dimensions and one-way error processing. Existing technologies treat UAV trajectory data (navigation system output) and multispectral data (sensor output) as independent data streams. Trajectory data is only used for geometric correction of image stitching, and multispectral data is only used for spectral feature analysis of ground features, failing to recognize the inherent connection between the two and their value in landslide hazard identification. Meanwhile, minor trajectory deviations generated when UAVs fly along preset routes are usually considered navigation errors and need to be corrected using techniques such as Kalman filtering and differential GPS (e.g., UAV manufacturers' SDK development documents define trajectory deviations >1m as "navigation anomalies" and trigger automatic return-to-home mechanisms; geological monitoring standards also require trajectory accuracy to be controlled within 0.5m). However, neither of these considerations addresses the secondary use of the deviation data, ignoring the potential geological environmental information it may contain. These two limitations together lead to low data utilization and missed opportunities to identify hazards through multi-dimensional data correlation.

[0007] At the cognitive level, there are gaps in interdisciplinary collaboration. Existing landslide remote sensing research is mostly limited to a single-chain reasoning of "remote sensing spectrum → ground object attributes," failing to address the cross-integration of aerodynamics, geotechnical mechanics, and remote sensing informatics. Aerodynamic research primarily focuses on UAV attitude control, without exploring the reaction of the geological environment to flight trajectories; while geotechnical engineering analyzes hidden terrain through borehole sampling, it hasn't considered using UAV remote sensing for indirect detection. This disciplinary barrier makes it difficult for current technologies to understand the deep mechanism by which "hidden terrain gradients affect UAV flight states, and subsequently correlate with trajectory offsets and spectral characteristics," becoming the fundamental reason for limitations in technological implementation.

[0008] The aforementioned limitations ultimately lead to insufficient effectiveness of existing intelligent identification methods: due to the failure to integrate multi-dimensional data associations, existing methods mostly focus on extracting potential hazard points from multispectral data with obvious characteristics. For landslide hazards in the early latent stage, where spectral features are covered by surface vegetation or soil, the identification effect is poor, making it difficult to meet the requirements of high-precision and early identification.

[0009] In view of this, an intelligent identification method for landslide hazard points based on UAV multispectral remote sensing is provided to overcome the above problems. Summary of the Invention

[0010] The purpose of this invention is to provide an intelligent identification method for landslide hazard points based on UAV multispectral remote sensing, so as to solve the problems mentioned in the background art.

[0011] To address the aforementioned technical problems, this invention provides an intelligent identification method for landslide hazard points based on UAV multispectral remote sensing, comprising the following steps:

[0012] 1) Data acquisition and preprocessing: Optimize the UAV system, retain all original trajectory offset values, preset standard strip flight paths and allow the UAV to adaptively fine-tune its attitude, perform radiometric correction and geometric coarse correction on multispectral images, use them for sub-second alignment of trajectory point clouds and spectral pixels, and build a fused dataset;

[0013] 2) Trajectory disturbance feature mining: calculate the distance of trajectory deviation from the preset route, mark the disorder area, construct the offset energy field model, invert the dynamic pressure distribution of airflow, link with the geotechnical mechanics knowledge base, and delineate underground hidden danger risk areas;

[0014] 3) Spectral feature enhancement and spatiotemporal filtering: Extract dominant spectral indices to delineate intuitive anomalous areas, capture minute spectral anomalies through spatiotemporal-spectral joint filtering, and improve signal strength using a weak signal enhancement algorithm;

[0015] 4) Dual-modal coupling analysis and intelligent recognition: Construct a coupled knowledge base, establish a counterfactual reasoning deep learning network, input relevant features, output pixel-level landslide probability map and development stage, and execute hierarchical early warning decision-making;

[0016] 5) Closed-loop verification and route optimization: Perform multi-source data cross-validation on high-risk areas, update the coupled knowledge base in reverse according to the actual test results, and automatically generate trajectory offset entropy values ​​in high-risk areas to guide encrypted routes.

[0017] Furthermore, in step 1), the UAV system adopts a vertical take-off and landing fixed-wing UAV, equipped with a high-precision RTK, multispectral sensor and IMU attitude sensor. It only issues a warning when the trajectory deviation is greater than 1m, but does not force a return to home for correction. It also records the attitude angle change sequence and environmental parameters simultaneously.

[0018] Further, in step 2), the trajectory deviation statistics and disorder zone marking include calculating the distance of trajectory deviation from the preset route, the frequency of local curvature change, and the threshold area of ​​abnormal attitude angle fluctuation. When the pitch angle change is greater than 10° for 5 consecutive times, the disorder zone marking is triggered.

[0019] Furthermore, in step 2), the offset energy field model is constructed based on the UAV dynamics equations, and the vortex intensity formula is:

[0020] ;

[0021] in:

[0022] : Vortex intensity;

[0023] The area covered by the drone's flight path;

[0024] Line integral along the actual flight path of the UAV;

[0025] : Topographic gradient air pressure change;

[0026] UAV velocity vector field;

[0027] Time derivative.

[0028] Further, in step 3), the dominant spectral indices include NDVI drop, rock fragmentation index, and soil moisture index; the spatiotemporal-spectral joint filtering uses a sliding window to compare time-series images, with a sliding time window of 1-3 months and a spatial window of 3×3 pixels. The formula for the spectral anomaly index is:

[0029] ;

[0030] in:

[0031] Spectral anomaly index;

[0032] Average reflectance within the current time window;

[0033] Average reflectance within a historical time window: interval from the current window The historical time window mean reflectance, For time intervals;

[0034] Background reflectance value;

[0035] The spatial gradient operator acts on the average reflectivity;

[0036] Absolute value operator;

[0037] when Areas with a value >0.15 are marked as suspicious areas; the weak signal enhancement algorithm uses principal component analysis to extract spectral change gradient features.

[0038] Furthermore, in step 4), the knowledge base is coupled to integrate geological maps, geotechnical parameters, and historical landslide cases to establish a rule base for topographic slope curvature-loose body critical angle-trajectory disorder response; the input features of the counterfactual reasoning deep learning network include trajectory offset, attitude fluctuation entropy value, spectral anomaly index, topographic slope curvature, and environmental parameters.

[0039] Furthermore, the counterfactual reasoning deep learning network incorporates an ideal behavior benchmark comparison branch. It generates ideal trajectories and spectral reference values ​​under undisturbed geological conditions through simulation, and calculates the significance of the deviation between the actual data and the benchmark. The formula is as follows:

[0040] ;

[0041] in:

[0042] Significance of Deviation: The overall output result represents the degree of difference between the actual collected data and the ideal benchmark data; The larger the value, the more significant the deviation of the actual data from the ideal state without geological disturbance, implying a higher contribution weight caused by geological disturbance; conversely, it may be more affected by environmental noise.

[0043] Actual feature vector: A vector composed of multi-dimensional observation data actually collected by the UAV, containing all feature parameters input to the counterfactual reasoning network;

[0044] Ideal eigenvector: An ideal baseline data vector generated through simulation under undisturbed geological conditions, compared with... One-to-one correspondence between dimensions;

[0045] The dimension-wise difference between the actual feature vector and the ideal feature vector reflects the degree to which each feature parameter deviates from the ideal state.

[0046] Squaring and Summation The sum of the squared differences of each dimension is obtained by summing them up to get the total squared deviation of all feature dimensions from the ideal state, which comprehensively reflects the overall deviation of multi-dimensional features.

[0047] Square root Take the arithmetic square root of the total sum of squares, such that... The dimensions of the deviation are kept consistent with the original dimensions of the characteristic parameters to intuitively understand the absolute degree of the deviation.

[0048] Furthermore, the characteristic is that: in step 4), the tiered early warning decision includes:

[0049] Low risk: Small changes in a single mode, i.e., trajectory offset ≤0.3m or spectral gradient <5% of background value;

[0050] Medium risk: The trajectory is disordered and vortex region nested within the spectral slightly deteriorated region, i.e., the offset is >0.3m and the spectral exponential gradient is >10% of the background value;

[0051] High risk: Strong coupling anomaly and a combination of terrain attributes including steep slopes on the free surface, where strong coupling anomaly refers to the nesting of drastic trajectory fluctuations, high entropy areas, and continuous spectral deterioration.

[0052] Furthermore, in step 5), the multi-source data cross-validation includes triggering low-altitude supplementary measurements or ground drilling verification in high-risk areas, and fusing LiDARDEM differential and InSAR time-series deformation fields to verify micro-displacement trends; in intelligent feedback optimization, the disorder probability of encrypted routes in high-risk areas is reduced to half of the original route interval.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] 1. Solved the problem of "data dimension fragmentation and one-way error processing" at the technical implementation level: By retaining all original values ​​of trajectory offset, the tiny offsets that were originally regarded as "navigation errors" are transformed into independent geological exploration signals, breaking through the traditional paradigm that trajectory data is only used for geometric correction; by aligning trajectory point clouds and spectral pixels at the sub-second level to construct a fused dataset, the fragmentation between trajectory data and multispectral data is broken, the utilization rate of data information is greatly improved, and multi-dimensional data association is realized to identify potential risks.

[0055] 2. Filling the "multidisciplinary gap" in cognition: By constructing a migration energy field model based on UAV dynamic equations, inverting the dynamic pressure distribution of airflow and linking it with the geotechnical mechanics knowledge base, the physical link of "implicit terrain gradient → airflow vortex → trajectory migration" is revealed; by constructing a coupled knowledge base integrating geological maps, geotechnical mechanics parameters, and historical landslide cases, a rule base of terrain slope curvature-loose body critical angle-trajectory disorder response is established, breaking down disciplinary barriers and constructing a dynamic mutual feedback mechanism of "UAV-environment-geology", providing a physical and geological rational explanation for the correlation between trajectory and spectral characteristics.

[0056] 3. Improved efficiency in identifying early and hidden hazards: Temporal-spatial-spectral joint filtering captures minute spectral anomalies and combines them with principal component analysis to enhance weak signals, penetrating vegetation / soil cover; trajectory offset is used to locate underground stress accumulation zones, complementing the weak spectral signals to identify early hazards that are "spectrally concealed but stress-active"; a counterfactual reasoning deep learning network is introduced to eliminate environmental noise interference and reduce misjudgments by calculating the significant deviation between actual data and an ideal benchmark without geological disturbance; and high-risk areas are guided by trajectory offset entropy values ​​to densify flight paths, increasing data density and distinguishing hazard development stages (stress accumulation period, deformation development period, and high-risk instability period), meeting the needs for high-precision, early-stage identification. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the principle of an intelligent identification method for landslide hazard points based on UAV multispectral remote sensing according to the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 The present invention provides a technical solution:

[0060] This method proposes four major technical systems: trajectory offset geological resource utilization, dual-modal feature penetration masking, counterfactual reasoning deviation quantification, and flight path dynamic entropy value driving, forming a closed-loop solution:

[0061] Turning trajectory deviation into a valuable asset: Converting redundant navigation errors into indirect detection signals of underground stress / loose body distribution, and correlating terrain micro-disturbances with trajectory turbulence through aerodynamic inversion.

[0062] Penetrating the dual cover of vegetation and soil: temporal-spatial-spectral joint filtering captures weak spectral changes, trajectory offset locates stress accumulation areas, and complementary identification of hidden hazards.

[0063] Counterfactual reasoning module innovation: Introduces ideal behavior benchmarks without geological disturbances, quantifies the significance of actual trajectory / spectral deviations, and avoids misjudgment due to environmental noise.

[0064] Adaptive trajectory entropy for flight routes: Dynamically encrypt flight routes based on trajectory disorder features, with non-preset uniform coverage, significantly improving data density in high-risk areas.

[0065] Furthermore, it does not rely on cutting-edge, high-cost equipment (such as LiDAR to penetrate vegetation), but instead cleverly overcomes the bottleneck of early hidden hazard identification through in-depth mining of conventional data and modeling of physical laws.

[0066] See Figure 1 As shown, an embodiment of an intelligent identification method for landslide hazard points based on UAV multispectral remote sensing is presented:

[0067] 1. Data Acquisition and Preprocessing: Original Trajectory Offset Preservation and Precise Spatiotemporal Registration:

[0068] 1.1 Unmanned Aerial Vehicle (UAV) System Optimization: A vertical takeoff and landing (VTOL) fixed-wing UAV is adopted (balancing flexibility in complex terrain and wide-area coverage), equipped with high-precision RTK (trajectory accuracy ≤0.1m), multispectral sensors (including visible light / near infrared / shortwave infrared), and an IMU attitude sensor. Furthermore: UAV firmware parameters are modified to retain all original trajectory offset values ​​(non-filtered correction), issuing warnings only when the offset is >1m without forcing a return-to-home correction, thus avoiding the smoothing out of geological signals; simultaneously recording attitude angle mutation sequences (high-frequency fluctuations in pitch / roll / yaw) and environmental parameters (wind speed / air pressure).

[0069] 1.2 Flight path design: A standard strip flight path is preset, but the UAV is allowed to adapt to slight terrain undulations / air vortices to fine-tune its attitude and retain the real trajectory, forming the original trajectory disorder data.

[0070] 1.3 Spatiotemporal precise registration: Multispectral images are subjected to radiometric correction and geometric coarse correction (with initial trajectory position assistance), and the trajectory point cloud and spectral pixels are aligned in sub-second intervals to construct a fused dataset (3D coordinates + spectral reflectance + time series).

[0071] Trajectory offset is collected as an independent geological exploration data source, overturning the paradigm that "trajectory is only used for geometric correction"; unfiltered raw data recording breaks through the traditional one-way control of trajectory accuracy by vendor SDKs / specifications.

[0072] 2. Trajectory perturbation feature mining: airflow vortex modeling and migration energy geological correlation:

[0073] 2.1 Trajectory Deviation Statistics and Disorder Zone Marking:

[0074] 2.1.1 Calculate the distance of trajectory deviation from the preset route, the frequency of local curvature abrupt changes, and the threshold area of ​​abnormal attitude angle fluctuations (such as pitch angle abrupt changes > 10° triggering the disorder zone marker 5 times consecutively).

[0075] 2.1.2 Constructing the Migration Energy Field Model: Based on the UAV dynamic equations (Bernoulli equation + terrain DEM driven), the dynamic pressure distribution of airflow induced by terrain micro-disturbances is inverted to locate the frequent vortex / turbulence areas (airflow disturbance caused by loose rock and soil leading to UAV attitude adjustments). Vortex intensity formula:

[0076] ;

[0077] in:

[0078] Vortex intensity (quantifies the strength of airflow vortices / turbulence, used to characterize the degree of airflow turbulence caused by topographic micro-disturbances);

[0079] : The spatial area covered by the drone's flight path (i.e., the total area of ​​the current analysis region);

[0080] Line integral along the actual flight path of the UAV (the integration range is the current flight path trajectory being analyzed).

[0081] Topographic gradient pressure variation (spatial pressure gradient, characterizing the spatial distribution difference of air pressure caused by slight topographic undulations) (for gradient operators)

[0082] UAV velocity vector field (a vector containing the magnitude and direction of velocity, describing the motion state of the UAV along its flight path).

[0083] Time derivative (time increment during integration).

[0084] 2.2. Geotechnical Mechanics Knowledge Base Connection: Integrate parameters such as loose body thickness, internal friction angle, and cohesion, and couple vortex / disordered zones with potential stress concentration zones (anticline axis / fault fracture zone) to delineate underground hidden danger risk zones.

[0085] The trajectory deviation is interpreted as the physical response of the airflow disturbed by loose rock and soil through aerodynamic modeling, and the correlation between trajectory disorder and the distribution of underground loose layers is established; the energy field quantification surpasses the coverage limitations of traditional single-point sensors, and low-cost initial screening of hidden dangers over a large area is achieved.

[0086] This section improves upon the existing UAV SDK trajectory forced correction mechanism and terrain DEM single terrain description function.

[0087] 3. Spectral feature enhancement and spatiotemporal filtering: Capturing weak signals penetrating vegetation / soil:

[0088] 3.1 Extraction of dominant spectral indices: Based on the sudden drop in NDVI, rock fragmentation index (SWIR band combination), and soil moisture index (abnormally high humidity area), visually abnormal areas are delineated.

[0089] 3.2 Hidden Signal Mining:

[0090] 3.2.1 Spatiotemporal-Spectral Joint Filtering: A sliding window (1-3 months temporal window, 3×3 pixels spatial window) is used to compare time-series images, focusing on sudden, minor spectral anomalies (<5% reflectance abrupt changes but with a continuous cumulative trend), and suppressing noise interference such as seasonal vegetation changes. Algorithm formula:

[0091] ;

[0092] in:

[0093] Spectral Anomaly Index (a comprehensive measure of the spatiotemporal anomaly of spectral reflectance, used to identify minute spectral changes);

[0094] : Average reflectance within the current time window (mean reflectance of multispectral image pixels within a sliding time window of 1-3 months);

[0095] Average reflectance within a historical time window (interval with the current window) The historical time window mean reflectance, (for time intervals)

[0096] Background reflectance value (average reflectance of stable ground features within the analysis area, used as a baseline reference value);

[0097] The spatial gradient operator acts on the average reflectance (characterizing the rate of change of reflectance in space, highlighting local spatial anomalies).

[0098] Absolute value operator (ensures that the relative value of the change over time is non-negative);

[0099] when Areas with a value >0.15 are marked as suspicious.

[0100] 3.2.2 Weak signal enhancement algorithm: Principal component analysis (PCA) is used to extract spectral change gradient features to enhance the signal intensity of early hidden soil moisture changes (early signs of groundwater level rise) and intensified rock weathering.

[0101] Existing methods rely on fixed thresholds to miss small spectral changes of less than 5%, while spatiotemporal gradient accumulation anomalies are used to penetrate vegetation / soil cover to identify potential hazards; PCA dimensionality reduction and fusion of multi-band changes surpasses the insufficient sensitivity of a single exponent.

[0102] This section improves upon existing methods for identifying single dominant spectral indices and single-phase fixed thresholds in the NDVI and SWIR bands.

[0103] 4. Dual-modal coupling analysis and intelligent recognition: Counterfactual reasoning and hierarchical early warning:

[0104] 4.1 Construction of Coupled Knowledge Base: Integrate geological maps, geotechnical parameters, and historical landslide cases to establish a rule base for topographic slope curvature, loose body critical angle, and trajectory disorder response (e.g., vortex disorder is easily triggered in loose deposits on steep slopes).

[0105] 4.2 Counterfactual Reasoning Deep Learning Network:

[0106] 4.2.1 Input features: trajectory offset, attitude fluctuation entropy, spectral anomaly index, terrain slope curvature, and environmental parameters.

[0107] 4.2.2 Module Design: An ideal behavior benchmark comparison branch is added—the ideal trajectory and spectral reference values ​​under undisturbed geological conditions are generated through simulation, and the significance of the deviation between the actual data and the benchmark is calculated. The formula (a core indicator used to quantify the significance of the deviation between the actual observed data and the undisturbed ideal benchmark) is as follows:

[0108] ;

[0109] in:

[0110] Significance of deviation: The overall output result represents the degree of difference between the actual collected data and the ideal benchmark data. The larger the value, the more significant the deviation of the actual data from the ideal state without geological disturbance, implying a higher contribution weight caused by geological disturbance (such as underground stress accumulation, loose body distribution, etc.); conversely, it may be more affected by environmental noise (such as random airflow, natural vegetation growth).

[0111] Actual Feature Vector: A vector composed of multi-dimensional observation data actually collected by the UAV, containing all feature parameters input to the counterfactual reasoning network, with specific dimensions corresponding to:

[0112] Trajectory offset: The three-dimensional spatial distance (in meters) between the actual flight trajectory and the preset flight path;

[0113] Attitude fluctuation entropy: A quantitative indicator of the degree of disorder in the high-frequency fluctuations of pitch, roll, and yaw angles of UAVs (dimensionless, reflecting attitude stability).

[0114] Spectral anomaly index: a gradient of minute changes in spectral reflectance obtained after spatiotemporal filtering and PCA enhancement (as shown in the formula). Calculation results (dimensionless).

[0115] Terrain slope and curvature: The surface slope (°) and curvature (1 / m, reflecting the degree of terrain undulation) of the target area; Environmental parameters: Environmental interference factors such as wind speed (m / s) and air pressure (hPa) during actual flight.

[0116] Ideal eigenvector: An "ideal baseline data vector under undisturbed geological conditions" generated through simulation, compared with... The dimensions are in a one-to-one correspondence, and their values ​​are generated based on the following assumptions:

[0117] There is no underground stress disturbance and no loose rock and soil mass distribution (i.e., no geological disaster-inducing factors).

[0118] It is only affected by purely environmental factors such as uniform airflow and stable terrain (excluding geologically driven abnormal interference).

[0119] Specifically, these include: ideal trajectory offset (theoretically close to 0, containing only normal navigation noise), ideal attitude fluctuation entropy (low entropy value, reflecting stable attitude), ideal spectral anomaly index (close to 0, with no significant abrupt change in reflectance), ideal terrain parameters (pure natural terrain attributes, without geological disturbance superposition), and ideal environmental parameters (benchmark values ​​consistent with the actual environment).

[0120] The difference between the actual and ideal feature vectors in each dimension reflects the degree to which each feature parameter deviates from the ideal state. For example, if the actual trajectory offset of a point is 0.8m and the ideal offset is 0.1m, then the difference in this dimension is 0.7m, representing the potential impact of geological disturbance on the trajectory of that point.

[0121] Squaring and Summation The sum of squares on each dimension difference (eliminating the influence of positive and negative signs and amplifying significant differences) yields the total square of all feature dimensions deviating from the ideal state, comprehensively reflecting the overall deviation of multi-dimensional features.

[0122] Square root Take the arithmetic square root of the total sum of squares, such that... The dimensions of the deviation are kept consistent with the original dimensions of the characteristic parameters (such as length, dimensionless exponent, etc.), which makes it easier to intuitively understand the absolute degree of the deviation.

[0123] The objective is to quantify the contribution weight of geological disturbances.

[0124] 4.2.3 Network Structure: Improve the U-Net or LSTM spatiotemporal model (trajectory sequence + spectral image input) to output pixel-level landslide probability map and development stage (offset-dominated: stress accumulation period; spectral-dominated: deformation development period; dual-dominated: high-risk instability period).

[0125] 4.3 Tiered Early Warning Decision-Making:

[0126] 4.3.1 Low risk: Small changes in a single mode (trajectory offset ≤ 0.3m or spectral gradient < 5% of background value);

[0127] 4.3.2, Medium Risk: Trajectory disordered vortex region nested within spectral micro-deterioration region (offset > 0.3m + spectral exponential gradient > 10% of background value);

[0128] 4.3.3 High risk: Strong coupling anomaly (severe trajectory fluctuations, high entropy nesting, and continuous spectral deterioration) + combination of terrain attributes such as steep slopes on the free surface.

[0129] Trajectory offset is deeply integrated with spectral variation, terrain attributes, and environmental data into the input network; counterfactual benchmark bias quantification is applied to landslide identification to scientifically eliminate false alarms caused by environmental interference; a multidisciplinary coupled rule base drives model interpretability, breaking through the black box of traditional deep learning.

[0130] 5. Closed-loop verification and route optimization: Trajectory entropy drives intelligent iteration.

[0131] 5.1 Cross-validation of multi-source data: Low-altitude supplementary measurement (UAV surrounding oblique photography of crack details) or ground drilling verification in high-risk areas (inclinometers are deployed first in the center of the trajectory disorder); fusion of LiDARDEM differential and InSAR time-series deformation field to verify micro-displacement trends.

[0132] 5.2 Intelligent feedback optimization: The measured results are used to update the coupled knowledge base in reverse (correcting the trajectory response mode of loose body thickness / strength); the trajectory offset entropy value is automatically generated in high-risk areas to guide the densification of flight paths (the flight path interval in areas with high disorder probability is reduced to 1 / 2 of the original), and the efficiency and accuracy of non-uniform coverage are significantly improved.

[0133] The flight path adaptively adjusts dynamically in disordered areas, replacing the redundant coverage of the traditional fixed grid; trajectory entropy (uncertainty quantification) is used to guide resource allocation and optimize cost-effectiveness.

[0134] Summarize:

[0135] I. It solved the problems of "fragmented data dimensions and unidirectional error processing" at the technical implementation level:

[0136] Existing technologies process UAV trajectory data and multispectral data separately (the trajectory is only used for geometric correction, and the spectrum is only used for ground feature analysis), and treat minor trajectory deviations as navigation errors that need to be corrected (ignoring their geological information value), resulting in low data utilization.

[0137] Original trajectory offset preservation and geological resource utilization: Modify the UAV firmware parameters to retain all original trajectory offset values ​​(non-filtered correction), break through the one-way control of trajectory accuracy by the manufacturer's SDK and monitoring specifications, and transform the tiny offset (<1m) that was originally regarded as "navigation error" into an independent geological exploration signal, and establish the correlation between trajectory data and underground stress and loose body distribution.

[0138] Precise spatiotemporal fusion of dual-modal data: Through sub-second spatiotemporal registration (timestamp error between trajectory point cloud and spectral pixel <0.1s), trajectory offset, attitude fluctuation sequence and multispectral reflectance data are deeply coupled to construct a fusion dataset of "three-dimensional coordinates + spectral features + time series", which completely breaks the traditional paradigm of "separation of trajectory and spectrum".

[0139] The utilization rate of data information has increased exponentially, and trajectory deviation has been transformed from "redundant error" into "geological exploration carrier," achieving a breakthrough in identifying potential hazards through multi-dimensional data correlation.

[0140] II. Filling the "interdisciplinary gaps" in cognition:

[0141] Existing technologies are limited to a single-chain reasoning of "remote sensing spectrum → ground object attributes," lacking the cross-integration of aerodynamics, geotechnical mechanics, and remote sensing informatics, and thus cannot understand the deep mechanism of "hidden terrain - flight status - trajectory / spectral correlation."

[0142] This invention fills this gap through interdisciplinary modeling:

[0143] Integration of aerodynamics and geotechnical mechanics: Based on Bernoulli's equation, a "misaligned energy field model" is constructed to invert the dynamic pressure distribution of airflow caused by topographic micro-disturbances. The model associates the trajectory disordered area with loose soil and rock, and underground stress concentration zones (anticline axis, fault fracture zone), revealing the physical link of "hidden topographic gradient → airflow vortex → trajectory misalignment".

[0144] Construction of a multidisciplinary coupled knowledge base: Integrating geological maps, geotechnical parameters (thickness of loose body, internal friction angle), and historical landslide cases, a rule base of "topographic slope curvature - critical angle of loose body - trajectory disorder response" is established to achieve a closed-loop interpretation of "remote sensing spectral features + trajectory dynamic features + geomechanical properties".

[0145] By breaking down disciplinary barriers and establishing a dynamic feedback mechanism between "drones, environment, and geology," a physical and geologically plausible explanation is provided for the correlation between trajectory deviation and spectral characteristics.

[0146] Third, it solves the problem of insufficient effectiveness of existing methods in identifying early and hidden hidden hazards:

[0147] Existing methods, due to the lack of integration of multi-dimensional data, struggle to identify early latent hazards (with weak or obscured spectral characteristics) that are covered by surface vegetation / soil.

[0148] This invention achieves a breakthrough through dual-modal penetration and intelligent reasoning:

[0149] Dual-modal feature penetration double masking:

[0150] Spectral side: "Temporal-spatial-spectral joint filtering" is used to capture small cumulative changes in reflectance of less than 5% (such as early soil moisture anomalies, early signs of vegetation withering), combined with PCA to enhance weak signals and penetrate vegetation / soil cover;

[0151] Trajectory-side: By locating underground stress accumulation zones (such as airflow vortex zones caused by loose body disturbance) through trajectory offset, and complementing weak spectral signals, early hidden dangers that are "spectrally concealed but stress-active" can be accurately identified.

[0152] Counterfactual reasoning to quantify bias: Introducing an "ideal benchmark without geological disturbance" and calculating the significance of the deviation between the actual trajectory / spectrum and the benchmark through a formula, eliminating environmental noise interference, avoiding misjudging isolated spectral / trajectory anomalies as potential hazards, and improving identification accuracy.

[0153] Dynamic entropy-driven optimization of flight routes: Based on the characteristics of trajectory disorder, the flight routes in high-risk areas are dynamically encrypted (the interval between high-probability disorder areas is reduced to 1 / 2 of the original), non-uniform coverage increases data density, and targeted enhancement of the detailed capture of early hidden dangers.

[0154] The detection rate of early hidden dangers covered by vegetation / soil is improved, and the development stage of hidden dangers can be distinguished (stress accumulation period / deformation development period / high-risk instability period), meeting the needs of high-precision and early identification.

[0155] therefore:

[0156] This invention does not rely on cutting-edge, high-cost equipment, but achieves three major breakthroughs through "deep mining of conventional data + multidisciplinary physical law modeling":

[0157] Trajectory offset has transformed from an "error correction target" to a "geological exploration signal," resulting in a revolutionary increase in data value.

[0158] Interdisciplinary integration constructs a causal chain of "hidden terrain-airflow-trajectory-spectrum," filling a gap in the field's understanding;

[0159] Dual-modal penetration and intelligent reasoning break through the bottleneck of early hidden hazard identification, providing precise tools for disaster prevention and mitigation in complex areas such as the Three Gorges Reservoir area and the southwestern mountainous areas.

Claims

1. A method for intelligent identification of landslide hazard points based on UAV multispectral remote sensing, characterized in that, Includes the following steps: 1) Data acquisition and preprocessing: Optimize the UAV system, retain all original trajectory offset values, preset standard strip flight paths and allow the UAV to adaptively fine-tune its attitude, perform radiometric correction and geometric coarse correction on multispectral images, use them for sub-second alignment of trajectory point clouds and spectral pixels, and build a fused dataset; 2) Trajectory disturbance feature mining: Calculate the distance of trajectory deviation from the preset route, the frequency of local curvature change, and the threshold area of ​​abnormal attitude angle fluctuation. When the pitch angle change is >10° for 5 consecutive times, the turbulence area is marked. Construct an offset energy field model based on the UAV dynamic equation, invert the dynamic pressure distribution of airflow, associate with the geotechnical mechanics knowledge base, and delineate the underground hidden danger risk area. The offset energy field model is constructed based on the dynamic equations of the UAV, and the vortex intensity formula is: ； in: : Vortex intensity; The area covered by the drone's flight path; Line integral along the actual flight path of the UAV; : Topographic gradient air pressure change; UAV velocity vector field; Time derivative; 3) Spectral feature enhancement and spatiotemporal filtering: Explicit spectral indices, including NDVI drop, rock fragmentation index, and soil moisture index, are extracted to delineate intuitive anomaly areas. Minor spectral anomalies are captured by spatiotemporal-spectral joint filtering. The spatiotemporal-spectral joint filtering uses a sliding window to compare time-series images. The sliding time window is 1-3 months and the spatial window is 3×3 pixels. A weak signal enhancement algorithm is used to improve signal strength. 4) Dual-modal coupling analysis and intelligent recognition: Construct a coupled knowledge base, establish a counterfactual reasoning deep learning network based on U-Net or LSTM spatiotemporal model, input relevant features, output pixel-level landslide probability map and development stage, and execute hierarchical early warning decision-making; The construction of the coupled knowledge base includes: integrating geological maps, geotechnical parameters, and historical landslide cases to establish a rule base for topographic slope curvature, loose body critical angle, and trajectory disorder response; The counterfactual reasoning deep learning network incorporates an ideal behavior benchmark comparison branch, which generates ideal trajectories and spectral reference values ​​under conditions of no geological disturbance through simulation, and calculates the significance of the deviation between the actual data and the benchmark. 5) Closed-loop verification and route optimization: Perform multi-source data cross-validation on high-risk areas, update the coupled knowledge base in reverse according to the actual test results, and automatically generate trajectory offset entropy values ​​in high-risk areas to guide encrypted routes.

2. The intelligent identification method for landslide hazard points based on UAV multispectral remote sensing as described in claim 1, characterized in that: In step 1), the UAV system adopts a vertical take-off and landing fixed-wing UAV, equipped with a high-precision RTK, multispectral sensor and IMU attitude sensor. It only issues a warning when the trajectory deviation is greater than 1m, but does not force a return to home for correction. It also records the attitude angle change sequence and environmental parameters simultaneously.

3. The intelligent identification method for landslide hazard points based on UAV multispectral remote sensing as described in claim 1, characterized in that: In step 3), the dominant spectral indices include NDVI drop, rock fragmentation index, and soil moisture index; the spatiotemporal-spectral joint filtering uses a sliding window to compare time-series images, with a sliding time window of 1-3 months and a spatial window of 3×3 pixels. The formula for the spectral anomaly index is: ； in: Spectral anomaly index; Average reflectance within the current time window; Average reflectance within a historical time window: interval from the current window The historical time window mean reflectance, For time intervals; Background reflectance value; The spatial gradient operator acts on the average reflectivity; Absolute value operator; when Areas with a value >0.15 are marked as suspicious areas; the weak signal enhancement algorithm uses principal component analysis to extract spectral change gradient features.

4. The intelligent identification method for landslide hazard points based on UAV multispectral remote sensing as described in claim 1, characterized in that: In step 4), the input features of the counterfactual reasoning deep learning network include trajectory offset, attitude fluctuation entropy, spectral anomaly index, terrain slope curvature, and environmental parameters. The attitude fluctuation entropy value is a quantitative indicator of the disorder of high-frequency fluctuations in the pitch, roll, and yaw angles of the UAV. The environmental parameters are the wind speed and air pressure environmental interference factors during actual flight.

5. The intelligent identification method for landslide hazard points based on UAV multispectral remote sensing as described in claim 4, characterized in that: The counterfactual reasoning deep learning network incorporates an ideal behavior benchmark comparison branch. It generates ideal feature vectors under undisturbed geological conditions through simulation. These ideal feature vectors are generated based on the assumptions of no underground stress disturbance, no loose soil or rock distribution, and influence only by uniform airflow and stable terrain—pure environmental factors. Their dimensions correspond one-to-one with the actual feature vectors, specifically including ideal trajectory offset, ideal attitude fluctuation entropy, ideal spectral anomaly index, ideal terrain parameters, and ideal environmental parameters. The significance of the deviation between the actual data and the benchmark is calculated using the following formula: ； in: Significance of Deviation: The overall output result represents the degree of difference between the actual collected data and the ideal benchmark data; The larger the value, the more significant the deviation of the actual data from the ideal state without geological disturbance, implying a higher contribution weight caused by geological disturbance; conversely, it may be more affected by environmental noise. Actual feature vector: A vector composed of multi-dimensional observation data actually collected by the UAV, containing all feature parameters input to the counterfactual reasoning network; Ideal eigenvector: An ideal baseline data vector generated through simulation under undisturbed geological conditions, compared with... One-to-one correspondence between dimensions; The dimension-wise difference between the actual feature vector and the ideal feature vector reflects the degree to which each feature parameter deviates from the ideal state. Squaring and Summation The sum of the squared differences of each dimension is obtained by summing them up to get the total squared deviation of all feature dimensions from the ideal state, which comprehensively reflects the overall deviation of multi-dimensional features. Square root Take the arithmetic square root of the total sum of squares, such that... The dimensions of the deviation are kept consistent with the original dimensions of the characteristic parameters to intuitively understand the absolute degree of the deviation.

6. The intelligent identification method for landslide hazard points based on UAV multispectral remote sensing as described in claim 1, characterized in that: In step 4), the tiered early warning decision includes: Low risk: Small changes in a single mode, i.e., trajectory deviation ≤ 0.3m or spectral anomaly gradient < 5% of background value; Medium risk: A region with turbulent trajectory and vortex pattern nested within a region of slight spectral deterioration, i.e., trajectory offset > 0.3m and spectral anomaly index gradient > 10% of background value; High risk: Strong coupling anomaly and steep slope combination terrain attributes, where strong coupling anomaly refers to the nested persistent spectral deterioration of the high entropy area of ​​the trajectory fluctuation; the high entropy area of ​​the trajectory fluctuation is the region with a high degree of disorder of high frequency fluctuation of UAV pitch angle, roll angle and yaw angle, and persistent spectral deterioration is the region where the spectral anomaly index is consistently greater than 0.

15.

7. The intelligent identification method for landslide hazard points based on UAV multispectral remote sensing as described in claim 1, characterized in that: In step 5), multi-source data cross-validation includes triggering low-altitude supplementary measurements or ground drilling verification in high-risk areas, and fusing LiDARDEM differential and InSAR time-series deformation fields to verify micro-displacement trends. In the intelligent feedback optimization, the probability of disorder in high-risk area encrypted routes is reduced to half of the original route interval.

Images