Multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling regulation method

Abstract

The application discloses a multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling regulation method, relates to the technical field of atmospheric boundary layer adaptive sampling, and comprises the following steps: step one, determining an unmanned aerial vehicle sampling area, synchronously collecting unmanned aerial vehicle-borne sensor data and remote sensing data, and further obtaining preprocessed parameter data sequences; step two, extracting basic features and instantaneous mutation features from the preprocessed parameter data sequences, integrating the basic features and the mutation features of the parameters, and further obtaining a sensor feature matrix; step three, preprocessing remote sensing data, determining ground reflectivity data and satellite boundary layer height data, extracting remote sensing spatial feature vectors and remote sensing time feature vectors, and integrating to obtain a remote sensing feature matrix; and step four, determining dynamic weights by using an analytic hierarchy process, combining the sensor feature matrix and the remote sensing feature matrix to determine a multi-modal fusion feature vector, establishing a physical model, and planning a high-turbulence region route.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of atmospheric boundary layer adaptive sampling, and particularly relates to a multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling regulation method. BACKGROUND

[0002] As a key layer directly affected by the earth's atmosphere and underlying surface (land, water), the spatiotemporal dynamic evolution of the atmospheric boundary layer's temperature, humidity, pressure, turbulence intensity and aerosol distribution directly affects the accuracy of weather forecasting, air quality assessment and regional pollution tracing efficiency. Therefore, high-precision and high-timeliness sampling monitoring of the atmospheric boundary layer is one of the core needs in atmospheric science research and environmental governance practice.

[0003] With the development of unmanned aerial vehicle technology, unmanned aerial vehicles have gradually replaced traditional fixed sites and manned aircraft to become the mainstream platform for atmospheric boundary layer sampling due to their flexibility, maneuverability and close-range detection capabilities. However, current unmanned aerial vehicle atmospheric boundary layer sampling technology still has many technical bottlenecks, making it difficult to meet the needs of precise characterization and adaptive regulation of boundary layer dynamic characteristics. The specific problems are as follows:

[0004] Current unmanned aerial vehicle sampling technology relies on a single data source. The current technology for feature extraction of sampling data is limited to basic statistical features (such as mean and variance), which can only reflect the steady-state change trend of atmospheric parameters, but ignores the instantaneous mutation characteristics commonly found in atmospheric boundary layers. Current boundary layer models used to guide sampling rely on pure physical mechanisms (such as atmospheric similarity theory) for derivation, which can reflect the macroscopic laws of boundary layer evolution, but are not dynamically corrected with real-time multi-modal monitoring data. Due to factors such as underlying surface heterogeneity and local meteorological disturbances (such as sudden gusts), the prediction error of pure physical models is large, making it difficult to provide accurate basis for sampling height planning and turbulence region identification, and thus leading to a disconnect between sampling regulation and actual boundary layer state.

[0005] Therefore, there is an urgent need for a multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling regulation method to address the above problems. SUMMARY

[0006] The purpose of the present application is to provide a multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling regulation method to solve the technical problems of ignoring instantaneous mutation characteristics, fixed sampling regulation mode, and inability to adapt to dynamic changes in the boundary layer in the prior art.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] The multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling regulation method comprises:

[0009] Step one, determine the UAV sampling area, synchronize the collection of UAV sensor data and remote sensing data, and further obtain the preprocessed parameter data sequence;

[0010] Step two, extract the basic features and instantaneous mutation features from the preprocessed parameter data sequence, integrate the basic features and mutation features of each parameter, and further obtain the sensor feature matrix;

[0011] Step three, pre-process the remote sensing data, determine the ground reflectivity data and satellite boundary layer height data, extract the remote sensing spatial feature vector and remote sensing time feature vector, and integrate to obtain the remote sensing feature matrix;

[0012] Step four, determine the dynamic weight by AHP, combine the sensor and remote sensing feature matrix to determine the multi-modal fusion feature vector, establish the physical model, and plan the high turbulence area route.

[0013] Further, the UAV sensor data and remote sensing data are collected synchronously, and the specific method is:

[0014] The UAV is equipped with wind speed sensor, temperature and humidity sensor, air pressure sensor and particulate matter concentration sensor. The preset collection frequency and sampling spatial resolution maximum value are determined according to historical data and actual requirements. The sampling area is determined, the center of the sampling area is taken as the starting point, a rectangular flight area with a radius of o is set, the initial flight height range is [y1, y2], and the wind speed, temperature and humidity, air pressure and particulate matter concentration data are collected in real time in the sampling area.

[0015] The remote sensing data collection is started synchronously with the sensor sampling. One hyperspectral image is taken every h1 seconds, covering the spatial range of the sampling area, and the image resolution is h2. A time node is determined every t time length, and the satellite remote sensing data is received by the ground terminal at each time node. The boundary layer height value corresponding to the sampling area is extracted, and the boundary layer height data between adjacent two time nodes is obtained through interpolation processing.

[0016] Further, the basic features are extracted from the preprocessed parameter data sequence, and the specific method is:

[0017] The basic features include statistical features, trend features and fluctuation features;

[0018] The statistical features include calculating the arithmetic mean and variance of the preprocessed data sequence;

[0019] Linear fitting is performed on the preprocessed data sequence to construct a linear equation with time stamp as independent variable and parameter value as dependent variable. The slope in the equation is the trend feature;

[0020] The fluctuation feature is determined by the ratio of standard deviation to arithmetic mean.

[0021] Further, based on the pre-processed parameter data sequence, the instantaneous mutation feature is extracted, and the specific method is as follows:

[0022] For the obtained pre-processed data sequence, the difference between the data points corresponding to two adjacent time stamps is calculated, and then divided by the interval of the two time stamps, to obtain the parameter change gradient in the time period;

[0023] The absolute value of all gradients of the single parameter is counted, the maximum value is taken and multiplied by the sensitivity coefficient to obtain the dynamic threshold, wherein the sensitivity coefficient is equal to the wind speed standard deviation in the corresponding time period;

[0024] If the gradient absolute value of a certain time period exceeds the dynamic threshold, the time period is marked as a mutation moment, and the change amplitude of the parameter in the time period is calculated, and finally the mutation feature set containing the mutation moment and the mutation amplitude is formed .

[0025] Further, the remote sensing data is pre-processed to determine the ground reflectivity data and the satellite boundary layer height data, and the specific method is as follows:

[0026] For the unmanned aerial high-spectral remote sensing image, a plurality of ground control points are selected in the sampling area, and the image pixel coordinates are aligned with the actual geographic coordinates through the control points;

[0027] For the high-spectral remote sensing image, the dark current is first deducted, and then the FLAASH atmospheric correction model is used to eliminate the influence of atmospheric scattering and absorption on the radiation value, to obtain the ground reflectivity data, denoted as R(λ), λ represents wavelength;

[0028] For the satellite remote sensing boundary layer height data, the cloud coverage area is removed, and the Kriging interpolation method is used to fill the blank area to obtain the complete boundary layer height spatial distribution data of the sampling area;

[0029] The pre-processed satellite boundary layer height data H(x, y) is determined.

[0030] Further, the remote sensing spatial feature vector is extracted, and the specific method is as follows:

[0031] For hyperspectral reflectance data R(λ), the K-means clustering algorithm is used to extract the central reflectance a1 and area proportion a2 of each cluster, forming a spatial clustering feature Fl(λ) = {(a1(λ, 1), a2(λ, 1)), ..., (a1(λ, K), a2(λ, K))}, where a1(λ, K) represents the reflectance of the cluster center of the Kth aerosol type in the sampling area at wavelength λ, and a2(λ, K) represents the area proportion of the Kth aerosol type in the sampling area at wavelength λ. The partial derivatives of the height of each spatial point in the x and y directions are calculated, and the spatial gradient value is obtained by taking the square root of the sum of the squares, denoted as GH(x, y). The spatial clustering feature is integrated with the spatial gradient of the pixels to obtain the remote sensing spatial feature vector Rs = [ =GH(x,y), =Fl(λ)].

[0032] Furthermore, the remote sensing temporal feature vector is extracted, specifically using the following method:

[0033] Calculate the difference in boundary layer height between two adjacent time points, and then divide it by the time interval to obtain the time rate of change of boundary layer height, rh.

[0034] The cumulative change is obtained by taking the absolute value of the boundary layer height difference at all adjacent time points within a certain period and then summing them. The combined result of the rate of change over time and the cumulative change is denoted as the remote sensing temporal feature vector Rt=[ =rh(t), ], where t represents the t-th time period, and rh(t) represents the rate of change of time between the start and end times of the t-th time period.

[0035] Furthermore, the dynamic weights are determined using the analytic hierarchy process (AHP). The specific method is as follows:

[0036] A three-level hierarchical structure is constructed: the target layer O, which is used to determine the importance weights of multimodal features; the criterion layer C, which, in combination with the core requirements of atmospheric sampling, sets two criteria, C1 and C2. C1 is used to measure the contribution of features to characterizing key physical quantities of the boundary layer, and C2 is used to measure the guiding value of features for sampling control strategies; and the indicator layer P, which clarifies the sensor features and remote sensing features to be compared. Sensor features include P1 temperature, P2 humidity, P3 air pressure, and P4 particulate matter concentration, while remote sensing features include P5 spatial gradient of boundary layer height, P6 spatial clustering features of aerosols, P7 temporal change rate of boundary layer height, and P8 cumulative change of boundary layer height.

[0037] Based on the definition of the physical meaning of each feature in atmospheric physics, the 1-9 scaling method is used to construct the judgment matrix of the index layer features for the criteria layers C1 and C2 respectively. The matrix elements represent the importance scale of the row features to the column features.

[0038] According to each judgment matrix, the index layer weight is calculated, and finally the sensor feature subjective weight set and the remote sensing feature subjective weight set are formed.

[0039] Further, according to each judgment matrix, the index layer weight is calculated, and the specific method is:

[0040] Combined with the atmospheric sampling target, the importance of C1 and C2 is judged, for the judgment matrix that passes the consistency test, the weight of each feature under the corresponding criterion is calculated by using the normalization method,

[0041] According to the maximum eigenvalue, the consistency index is calculated, and further the consistency ratio CR is obtained, if CR < s, it indicates that the consistency of the judgment matrix is qualified, and the weight is reliable, if CR >= s, the scale of pairwise comparison needs to be adjusted combined with the atmospheric physics theory until CR < 0.1;

[0042] According to the index layer weight = criterion layer weight * weight of the feature under the corresponding criterion, finally the sensor feature subjective weight set and the remote sensing feature subjective weight set are formed.

[0043] As described above, since the above technical scheme is adopted, the beneficial effects of the present application are:

[0044] 1、The present application relies on the analytic hierarchy process to construct a three-level structure, and determines the feature weight based on the atmospheric physics theory as the core basis, the criterion layer is closely related to the boundary layer sampling physical target, the index layer feature corresponds to the core physical parameter, the judgment matrix is constructed by using the 1-9 scale method and consistency test, the subjective deviation of the traditional empirical weight is avoided, and it is ensured that the weight distribution conforms to the atmospheric physics law, and this dynamic weight mechanism makes the multi-modal fusion feature vector highlight the key features and suppress the redundant information, and the physical relevance and decision support ability of the fusion result are significantly improved;

[0045] 2、The present application adopts a hybrid architecture of "atmospheric similar theory physical model + LSTM time sequence network": the physical model provides the theoretical benchmark value of the boundary layer height and the turbulence intensity, the LSTM network corrects the model deviation in real time based on the multi-modal fusion feature, and can trigger rapid update when detecting instantaneous mutation, and ensures the accurate prediction of the model to the boundary layer dynamics, and on this basis, by improving the algorithm takes the turbulence intensity as the path cost, and preferentially plans the sampling route of the high turbulence and high value area, and dynamically adjusts the sampling parameters, which avoids the redundant sampling waste of the low turbulence area and the endurance of the unmanned aerial vehicle, and realizes the data encryption collection of the key area, greatly improves the sampling pertinence and data effectiveness, and provides high-quality and representative measured data for atmospheric boundary layer research;

[0046] 3、The application breaks the limitation of single data source in space-time coverage by synchronously collecting unmanned aerial vehicle sensor data and remote sensing data, the sensor data captures the dynamic changes of atmospheric parameters in a local area, the remote sensing data reflects the boundary layer height distribution and aerosol clustering characteristics in a regional scale, and the two complement each other in space-time, and not only the basic features are extracted to depict the steady evolution of parameters, but also the dynamic threshold method is used to capture the instantaneous mutation characteristics, so as to make up for the defects of traditional technologies which only pay attention to static characteristics, and finally the sensor and remote sensing feature matrix formed can comprehensively cover the multi-dimensional evolution information of the boundary layer "steady-instantaneous" and "local-regional", and provide complete data support for subsequent accurate analysis. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0048] Figure 1 The step diagram of the multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling control method of the present application is shown;

[0049] Figure 2 The step diagram of the method for extracting the sensor feature matrix of the present application is shown;

[0050] Figure 3 The step diagram of the method for extracting the remote sensing feature matrix of the present application is shown. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] Embodiment one, the multi-modal fusion unmanned aerial vehicle atmospheric boundary layer adaptive sampling control method as shown in Figure 1 specifically includes the following steps:

[0053] Step one, determine the sampling area of the unmanned aerial vehicle, synchronously collect the unmanned aerial vehicle sensor data and remote sensing data, and further obtain the preprocessed parameter data sequence;

[0054] The unmanned aerial sensor data collection carries wind speed sensors, temperature and humidity sensors, air pressure sensors, and particulate matter concentration sensors. According to historical data and actual requirements, the preset collection frequency and sampling spatial resolution maximum value are determined to determine the sampling area. Taking the center of the sampling area as the starting point, a rectangular flight area with a radius of 2 km is set. The initial flight height range is 500-1000 m (covering the estimated boundary layer height range). Real-time collection of wind speed, temperature and humidity, air pressure, and particulate matter concentration data is performed. The data is transmitted back to the ground control terminal in real time through the UAV image transmission link, and is stored locally at the same time.

[0055] Remote sensing data collection is synchronized with sensor sampling. Every h1 second, a hyperspectral image is taken, covering the spatial range of the sampling area. The image resolution is h2 (i.e., each pixel corresponds to a ground area of h2 x h2). Every t time interval determines a time node. The ground terminal receives satellite remote sensing data at each time node, including MODIS sensor boundary layer height data. The boundary layer height value corresponding to the sampling area is extracted, and the boundary layer height data between adjacent time nodes is obtained through interpolation processing.

[0056] Determine the sampling time interval of each sensor. For each data collection time period, select 5 to 10 data points as a sliding window (the window size can be dynamically adjusted according to the current atmospheric boundary layer turbulence intensity. The stronger the turbulence, the larger the window to enhance the smoothing effect). Calculate the arithmetic mean of all data points in the window as the smoothed data corresponding to the center timestamp of the window to form the preliminary preprocessed data sequence.

[0057] For the smoothed preliminary preprocessed data sequence, first calculate the arithmetic mean and standard deviation of all data points. Adopt the 3σ rule for outlier rejection - if the absolute value of the difference between a data point and the mean value exceeds 3 times the standard deviation, it is determined to be an outlier and is removed from the sequence. Perform preprocessing operations on wind speed, temperature and humidity, air pressure, and particulate matter concentration data to obtain the final preprocessed parameter data sequence. This sequence can truly reflect the dynamic changes of atmospheric parameters.

[0058] Step two, based on the preprocessed parameter data sequence, extract the basic features and instantaneous mutation features, integrate the basic features and mutation features of each parameter, and further obtain the sensor feature matrix.

[0059] Based on the preprocessed parameter data sequence, extract the basic features, including:

[0060] Statistical feature extraction: calculate the arithmetic mean (reflecting the overall level of atmospheric parameters over a period of time, such as average temperature) and variance of the preprocessed data sequence (the greater the variance, the more intense the parameter fluctuation), where the denominator is the length of the data sequence minus 1 (i.e., sample variance);

[0061] Trend feature extraction: linear fitting is performed on the preprocessed data sequence to construct a linear equation with time stamp as independent variable and parameter value as dependent variable. The slope of the equation is the trend feature (a positive slope indicates an upward trend of the parameter over time, such as a gradual increase in temperature; a negative slope indicates a downward trend, and the greater the absolute value of the slope, the more obvious the trend).

[0062] Variance coefficient feature extraction: the coefficient of variation is calculated by the ratio of standard deviation to arithmetic mean. This feature can eliminate the influence of parameter magnitude difference, facilitating the comparison of fluctuation intensity of different types of parameters (such as temperature), for example, a coefficient of variation of 0.2 indicates that the parameter fluctuation is relatively mild, providing a quantitative basis for subsequent judgment of boundary layer stability.

[0063] Based on the preprocessed parameter data sequences, instantaneous mutation features are extracted, including:

[0064] Gradient calculation: the difference between the data points corresponding to the adjacent two time stamps (sampling time points) is calculated, and then divided by the interval between the two time stamps (i.e., sampling interval), to obtain the parameter change gradient in that time period (the greater the absolute value of the gradient, the more intense the parameter change in that time period).

[0065] Dynamic threshold setting: the absolute values of all gradients of a single parameter are calculated, and the maximum value is multiplied by a sensitivity coefficient to obtain the dynamic threshold, where the sensitivity coefficient is equal to the standard deviation of wind speed in the corresponding time period (t is the time interval).

[0066] Mutation feature determination: if the absolute value of the gradient of a certain time period exceeds the dynamic threshold, the time period is marked as a mutation time, and the change amplitude of the parameter in that time period (i.e., the absolute value of the difference between adjacent data points) is calculated, finally forming a mutation feature set containing mutation time and mutation amplitude. This feature set can accurately capture the instantaneous dynamics of the boundary layer and supplement the shortcomings of the steady-state features, for example: when two mutation points are identified, t=100.2s and t=101.0s, the corresponding mutation amplitudes are a and b, respectively, then Fc={ (100.2, 8), (101.0, 10)}.

[0067] Integrate the basic features and mutation features of each parameter to form a feature vector of a single sensor. The arithmetic mean , variance , linear fitting slope , coefficient of variation and mutation feature set of the parameter data collected by each sensor are calculated. , combined in a fixed order as the feature vector of the sensor, denoted as [ , , , , F ], where i represents the i-th sensor;

[0068] The temperature feature vector , the humidity feature vector , the air pressure feature vector , and the particulate matter concentration feature vector are integrated to form the sensor feature matrix S .

[0069] Step three, pre-process the remote sensing data, determine the surface reflectivity data and satellite boundary layer height data, extract the remote sensing spatial feature vector and remote sensing time feature vector, and integrate to obtain the remote sensing feature matrix;

[0070] The remote sensing data is pre-processed by geometric and radiometric correction. Geometric correction: for unmanned aerial high-spectral remote sensing images, select multiple ground control points (such as obvious building corner points, road intersections) in the sampling area, and align the image pixel coordinates with the actual geographic coordinates through the control points;

[0071] Radiometric correction: for high-spectral remote sensing images, first deduct the dark current (the dark current signal of the sensor in the absence of light), then use the FLAASH atmospheric correction model to eliminate the influence of atmospheric scattering and absorption on the radiation value, and finally obtain accurate surface reflectivity data, denoted as R(λ), λ represents wavelength;

[0072] For satellite remote sensing boundary layer height data, remove the cloud coverage area (clouds will interfere with boundary layer height inversion), and use Kriging interpolation method (based on the height values of the surrounding n non-cloud pixels) to fill in the blank area to obtain the complete spatial distribution data of the boundary layer height of the sampling area.

[0073] Determine the satellite boundary layer height data H(x, y) after preprocessing (x and y are plane coordinates) to calculate the partial derivative of the height of each spatial point in the x and y directions, and obtain the spatial gradient value by square sum, denoted as GH(x, y);

[0074] For hyperspectral reflectance data R(λ), the K-means clustering algorithm (K=4 in this embodiment, preset according to the aerosol type of the sampling area, such as background aerosol, dust, industrial aerosol, and fog droplet) is used to extract the center reflectivity a1 of each type (for example, the industrial aerosol cluster center has a reflectivity of 0.18 at a wavelength of 550 nm) and the area ratio a2 (industrial aerosol ratio 12%), to form a spatial clustering feature Fl(λ)={(a1(λ,1),a2(λ,1)),...,(a1(λ,K),a2(λ,K))},a1(λ,K) represents the reflectivity of the Kth sampling area aerosol type cluster center at wavelength λ, and a2(λ,K) represents the area ratio of the Kth sampling area aerosol type at wavelength λ. The spatial clustering feature and the spatial gradient of the pixel are integrated to obtain the remote sensing spatial feature vector Rs=[ =GH(x,y), =Fl(λ)];

[0075] The time change rate is extracted, the difference between the boundary layer heights of two adjacent time points is calculated, and then divided by the time interval to obtain the time change rate rh of the boundary layer height (the change rate is positive, indicating that the boundary layer rises over time, such as the increase of solar radiation during the day leading to the rise of the mixing layer; the change rate is negative, indicating that the boundary layer decreases);

[0076] Cumulative change amount extraction: taking the absolute value of the difference between the boundary layer heights of all adjacent time points in a period of time, and then summing to obtain the cumulative change amount (the larger the cumulative change amount, the more frequent the fluctuation of the boundary layer height in the time period, and the more unstable the boundary layer), and the time change rate and the cumulative change amount are integrated to obtain a combination, which is denoted as the remote sensing time feature vector Rt=[ =rh(t), ],where t represents the tth time period, and rh(t) represents the time change rate of the starting time point and the ending time point of the tth time period;

[0077] The remote sensing spatial feature vector and the remote sensing time feature vector are combined to form a remote sensing feature matrix R .

[0078] Step four, determine the dynamic weight by the analytic hierarchy process, combine the sensor and the remote sensing feature matrix to determine the multi-modal fusion feature vector, establish a physical model, and plan the route of the high-turbulence area;

[0079] Based on atmospheric physics theory, a judgment matrix is constructed by the analytic hierarchy process (AHP) to determine the importance of different features;

[0080] Based on the physical target of atmospheric boundary layer sampling (precise description of boundary layer evolution and guidance for unmanned aerial vehicle sampling control), a three-level structure is constructed, and each level is based on atmospheric physics theory:

[0081] Objective layer (O): Determine the importance weight of multi-modal features (sensors + remote sensing), support multi-modal fusion and adaptive sampling regulation;

[0082] Criteria layer (C): Set 2 criteria (both derived from atmospheric physics application logic) based on atmospheric sampling core requirements:

[0083] C1: Contribution of features to the characterization of key physical quantities of the boundary layer (e.g., whether the features are directly related to core physical parameters such as boundary layer height, turbulence intensity, and aerosol distribution);

[0084] C2: Guidance value of features to sampling regulation strategy (e.g., whether the features directly affect the regulation decisions such as sampling path height planning and frequency adjustment);

[0085] Index layer (P): Clearly define the specific features to be compared (corresponding to the core features in the sensor and remote sensing feature matrix):

[0086] Sensor features (P1-P4): P1 (temperature), P2 (humidity), P3 (air pressure), P4 (particulate matter concentration);

[0087] Remote sensing features (P5-P8): P5 (spatial gradient of boundary layer height), P6 (aerosol spatial clustering feature), P7 (time variation rate of boundary layer height), P8 (cumulative change amount of boundary layer height).

[0088] According to the definition of the physical meaning of each feature by atmospheric physics, a judgment matrix of the index layer features is constructed for criteria C1 and C2 using the "1-9 scale method". The matrix elements represent the importance scale of the row features to the column features. Since there may be logical contradictions such as P1>P2, P2>P3 but P1<P3 in pairwise comparison, consistency check is needed to verify the rationality of the judgment matrix;

[0089] According to each judgment matrix, calculate the index layer weight (simplified calculation: normalize each column of the matrix, then sum by row, and calculate the eigenvalue approximation);

[0090] Based on the atmospheric sampling target, determine the importance of C1 (physical quantity characterization contribution) and C2 (regulation guidance value) (both equally important (scale 1), so the weights of C1 and C2 are both 0.5). For the judgment matrix that passes the consistency check, use the normalization method to calculate the weight of each feature under the corresponding criteria,

[0091] According to the maximum eigenvalue calculation consistency index, and further get the consistency ratio CR, if CR<0.1, it is explained that the judgment matrix consistency is qualified, and the weight is reliable; if CR≥0.1, the scale of pairwise comparison needs to be adjusted combined with the atmospheric physics theory (such as modifying the scale of "P4 and P1" from 3 to 2), until CR<0.1.

[0092] According to the index layer weight = criterion layer weight × the weight of the feature corresponding to the criterion, the final sensor feature subjective weight set and remote sensing feature subjective weight set are formed.

[0093] Based on the dynamic weight, the sensor features and remote sensing features are deeply fused to form a fusion feature vector that can comprehensively reflect the characteristics of the boundary layer. The sensor feature matrix and the remote sensing feature matrix are respectively weighted and summed according to the dynamic weight to obtain the weighted sensor feature and the weighted remote sensing feature. The two types of features are spliced to form the final multi-modal fusion feature vector, and the formula is as follows:

[0094] ;

[0095] Among them, indicates the fusion feature vector, weighted sensor and remote sensing features, indicates the i-th feature of the sensor, indicates the j-th feature of the remote sensing, indicates the dynamic weight of the i-th feature of the sensor, indicates the dynamic weight of the j-th feature of the remote sensing;

[0096] A physical mechanism + data driven hybrid architecture is used to construct a dynamic model of the boundary layer structure. Based on the atmospheric similarity theory, the theoretical values of the key parameters (height, turbulence intensity) of the boundary layer are calculated, wherein the theoretical value of the boundary layer height is equal to the Karman constant × the friction velocity / Coriolis parameter, and the theoretical value of the turbulence intensity is equal to 0.15 × (1-sampling height / boundary layer height The friction velocity and the Coriolis parameter are calculated as the basic physical quantities, which are directly derived from the wind speed, latitude and other data collected by the unmanned aerial vehicle sensor;

[0097] The unmanned aerial vehicle collects the temperature corresponding to different heights z at a fixed height interval within the initial flight height range to form a vertical temperature sequence {T(z1),..., T(za),..., T(zn)} (z1<z2<...<zn). The temperature gradient of adjacent heights is calculated, for example, the temperature gradient at z1 is equal to T(z2) minus T(z1) divided by the height difference between z1 and z2. According to the actual demand and historical data, set the temperature gradient threshold value QA. When the temperature gradient first appears at za and is greater than QA, it is judged that za+1 is the thickness of the mixing layer.

[0098] The output bias of the physical model is optimized using an LSTM temporal network. The multimodal fusion feature vector is used as input, and the correction coefficients for boundary layer height, turbulence intensity, and mixing layer thickness are used as output. The correction coefficients are multiplied by the theoretical values ​​of the physical model to obtain the correction values, thereby reducing the error between the prediction results and the measured data (such as microwave radiometer and satellite true values). Newly acquired multimodal data (after feature extraction and fusion) is input into the LSTM network every Q minutes to update the network parameters. When a transient change feature is detected (such as the change amplitude of a parameter exceeding the preset threshold, or more than two consecutive change points within 1 minute), the model is immediately triggered to update rapidly without waiting for the normal Q-minute cycle.

[0099] In this embodiment, the network parameters are set as follows: number of input layer neurons = fusion feature vector dimension (9-dimensional), 2 hidden layers (32 neurons per layer), 3 output layer neurons (corresponding to the correction coefficients for boundary layer height, turbulence intensity, and mixing layer thickness), activation function is ReLU, optimizer is Adam, and training epochs are 50. Training data: the fusion feature vectors and measured boundary layer data (real boundary layer height obtained through a microwave radiometer) from the first hour (14:00-15:00) are selected as the training set, and the data from the last hour (15:00-16:00) are selected as the test set.

[0100] Based on the boundary layer height predicted by the model, set the sampling height range (approximately 0.3-1.2 times the predicted height); at the height corresponding to the predicted mixed layer thickness and at the height where instantaneous abrupt changes occur (such as the 300m height where the particulate matter concentration abrupt changes), reduce the sampling interval to half of the initial value, focusing on capturing key layer data.

[0101] Improve The algorithm uses the turbulence intensity predicted by the model as the path cost—the higher the cost in areas with stronger turbulence, the higher the cost. It prioritizes the planning of sampling routes, increases sampling points in high-turbulence areas, and reduces redundant sampling in low-turbulence areas.

[0102] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0103] The preferred embodiments of the application disclosed above are only to facilitate the elucidation of the application. The preferred embodiments do not describe all the details of the application and limit the application to the specific embodiments. Obviously, many modifications and variations can be made in light of the teachings above. The description is chosen and described in order to provide the best illustration of the application and its practical application to those skilled in the art and to enable those skilled in the art to utilize the application in its best mode. The application is only limited by the claims and their full scope and equivalents.

Claims

1. A multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles (UAVs), characterized in that, include: Step 1: Determine the sampling area of ​​the UAV, and simultaneously collect data from the UAV's onboard sensors and remote sensing data to obtain the preprocessed parameter data sequence. Step 2: Extract basic features and transient mutation features from the preprocessed parameter data sequences, integrate the basic features and mutation features of each parameter, and further obtain the sensor feature matrix; The method for extracting the transient mutation features is as follows: for the obtained preprocessed data sequence, calculate the difference between the data points corresponding to two adjacent timestamps, and then divide it by the interval between the two timestamps to obtain the parameter change gradient within the time interval between the two timestamps. The absolute values ​​of all gradients for a single parameter are statistically analyzed. The maximum value is taken and multiplied by the sensitivity coefficient to obtain the dynamic threshold, where the sensitivity coefficient is equal to the standard deviation of wind speed within the corresponding time period. If the absolute value of the gradient exceeds the dynamic threshold in a certain time period, then that time period is marked as a mutation moment, and the magnitude of parameter change within that time period is calculated, ultimately forming a mutation feature set containing the mutation moment and mutation magnitude. ; Step 3: Preprocess the remote sensing data, determine the surface reflectance data and satellite boundary layer height data, extract the remote sensing spatial feature vector and remote sensing temporal feature vector, and integrate them to obtain the remote sensing feature matrix; Step 4: Determine the dynamic weights using the analytic hierarchy process (AHP), combine sensor and remote sensing feature matrices to determine the multimodal fusion feature vector, establish a physical model, and plan routes in high-turbulence areas.

2. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 1, characterized in that, The method for simultaneously collecting UAV-borne sensor data and remote sensing data is as follows: The drone is equipped with a wind speed sensor, a temperature and humidity sensor, an air pressure sensor, and a particulate matter concentration sensor. Based on historical data and actual requirements, the sampling frequency and the maximum value of the sampling spatial resolution are preset to determine the sampling area. A rectangular flight area with a radius of 0 is set with the center of the sampling area as the starting point and the initial flight altitude range [y1, y2]. Wind speed, temperature and humidity, air pressure, and particulate matter concentration data are collected in real time within the sampling area. Remote sensing data acquisition is started synchronously with sensor sampling. One hyperspectral image is captured every h1 seconds, covering the spatial range of the sampling area. The image resolution is h2. A time node is determined every t time interval. The ground terminal receives satellite remote sensing data at each time node, extracts the boundary layer height value corresponding to the sampling area, and obtains the boundary layer height data between two adjacent time nodes through interpolation. The actual values ​​of radius o, initial flight altitude range [y1, y2], hyperspectral imaging interval h1, image resolution h2, and duration t are calculated based on actual needs and historical data.

3. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 1, characterized in that, The basic features are extracted from the preprocessed data sequences of each parameter. The specific method is as follows: The basic characteristics include statistical characteristics, trend characteristics, and fluctuation characteristics; Statistical characteristics include calculating the arithmetic mean and variance of the preprocessed data sequence; Linear fitting is performed on the preprocessed data sequence to construct a linear equation with timestamp as the independent variable and parameter values ​​as the dependent variable. The slope of the equation is the trend feature. Fluctuation characteristics are determined by the ratio of standard deviation to arithmetic mean.

4. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 1, characterized in that, The remote sensing data is preprocessed to determine the surface reflectance data and the satellite boundary layer height data. The specific method is as follows: For UAV-borne hyperspectral remote sensing images, multiple ground control points are selected in the sampling area, and the image pixel coordinates are aligned with the actual geographic coordinates through the control points. For hyperspectral remote sensing images, dark current is first subtracted, and then the FLAASH atmospheric correction model is used to eliminate the influence of atmospheric scattering and absorption on the radiation value to obtain the surface reflectance data, denoted as R(λ), where λ represents the wavelength. For the boundary layer height data from satellite remote sensing, cloud-covered areas are removed, and blank areas are filled using Kriging interpolation to obtain complete spatial distribution data of the boundary layer height of the sampled area; Determine the preprocessed satellite boundary layer height data H(x, y).

5. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 1, characterized in that, The specific method for extracting remote sensing spatial feature vectors is as follows: For hyperspectral reflectance data R(λ), the K-means clustering algorithm is used to extract the central reflectance a1 and area proportion a2 of each cluster, forming a spatial clustering feature Fl(λ) = {(a1(λ, 1), a2(λ, 1)), ..., (a1(λ, K), a2(λ, K))}, where a1(λ, K) represents the reflectance of the cluster center of the Kth aerosol type in the sampling area at wavelength λ, and a2(λ, K) represents the area proportion of the Kth aerosol type in the sampling area at wavelength λ. The partial derivatives of the height of each spatial point in the x and y directions are calculated, and the spatial gradient value is obtained by taking the square root of the sum of the squares, denoted as GH(x, y). The spatial clustering feature is integrated with the spatial gradient of the pixels to obtain the remote sensing spatial feature vector Rs = [ =GH(x,y), =Fl(λ)].

6. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 1, characterized in that, The specific method for extracting remote sensing temporal feature vectors is as follows: Calculate the difference in boundary layer height between two adjacent time points, and then divide it by the time interval to obtain the time rate of change of boundary layer height, rh. The cumulative change is obtained by taking the absolute value of the boundary layer height difference at all adjacent time points within a certain period and then summing them. The combined result of the rate of change over time and the cumulative change is denoted as the remote sensing temporal feature vector Rt=[ =rh(t), ], where t represents the t-th time period, and rh(t) represents the rate of change of time between the start and end times of the t-th time period.

7. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 1, characterized in that, The dynamic weights are determined using the analytic hierarchy process (AHP). The specific method is as follows: A three-level hierarchical structure is constructed: the target layer O, which is used to determine the importance weights of multimodal features; the criterion layer C, which, in combination with the core requirements of atmospheric sampling, sets two criteria, C1 and C2. C1 is used to measure the contribution of features to characterizing key physical quantities of the boundary layer, and C2 is used to measure the guiding value of features for sampling control strategies; and the indicator layer P, which clarifies the sensor features and remote sensing features to be compared. Sensor features include P1 temperature, P2 humidity, P3 air pressure, and P4 particulate matter concentration, while remote sensing features include P5 spatial gradient of boundary layer height, P6 spatial clustering features of aerosols, P7 temporal change rate of boundary layer height, and P8 cumulative change of boundary layer height. Based on the definition of the physical meaning of each feature in atmospheric physics, the 1-9 scale method is used to construct the judgment matrix of the index layer features for the criterion layers C1 and C2 respectively. The matrix elements represent the importance scale of the row features to the column features. Based on each judgment matrix, the weights of the index layer are calculated, ultimately forming the sensor feature subjective weight set and the remote sensing feature subjective weight set.

8. The multimodal fusion-based adaptive sampling and control method for the atmospheric boundary layer of unmanned aerial vehicles according to claim 7, characterized in that, The weights of the indicator layer are calculated based on each judgment matrix, using the following method: Based on the atmospheric sampling objectives, the importance of C1 and C2 is determined. For the judgment matrix that passes the consistency test, the weight of each feature under the corresponding criterion is calculated using the normalization method. The consistency index is calculated based on the largest eigenvalue, and then the consistency ratio CR is obtained. If CR < 0.1, it means that the consistency of the judgment matrix is ​​qualified and the weight is reliable. If CR ≥ 0.1, the scale of the pairwise comparison needs to be adjusted in combination with atmospheric physics theory until CR < 0.

1. The subjective weight set of sensor features and the subjective weight set of remote sensing features are ultimately formed by calculating the weight of the indicator layer as the weight of the criterion layer multiplied by the weight of the feature under the corresponding criterion.

Images