Multi-rotor unmanned aerial vehicle airborne flux observation platform method and system

Abstract

The invention relates to the technical field of unmanned aerial vehicle application and environment monitoring, and discloses a multi-rotor unmanned aerial vehicle airborne flux observation platform method and system. Comprising the following steps of S1, task planning and system initialization, S2, dynamic data acquisition, S3, data preprocessing and quality control, S4, flux calculation and multi-dimensional correction, S5, data transmission and distributed storage, and S6, multi-source data fusion and application analysis. As the airborne platform of the unmanned aerial vehicle has good stability, stable operation of the sensor can be maintained in the flight process, and the influence of factors such as flight attitude change on observation data is reduced, so that the stability and reliability of the data are improved, and the operation and maintenance cost of the unmanned aerial vehicle is relatively low; a large amount of manpower and material resource investment, such as tower footing construction and cable laying, is not needed, the long-term observation cost can be effectively reduced, and the observation sustainability is improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) applications and environmental monitoring technology, specifically to a method and system for an airborne flux observation platform for multi-rotor UAVs. Background Technology

[0002] UAV-borne flux observation technology integrates sensors onto a UAV platform, leveraging its maneuverability to dynamically observe environmental parameters such as atmospheric boundary layer fluxes and ecosystem carbon cycles. This technology overcomes the spatial limitations of traditional ground-based observations by utilizing the flight capabilities of UAVs, and combines high-precision sensors to collect regional-scale flux data, providing a new technological approach for fields such as ecological environment research and agricultural monitoring.

[0003] In existing technologies, unmanned aerial vehicle (UAV) flux observation systems typically consist of multi-rotor UAVs equipped with sensors such as three-dimensional ultrasonic anemometers and carbon dioxide water vapor analyzers. They achieve constant altitude and speed flight through flight control systems and use data acquisition devices to record sensor data in real time.

[0004] In existing technologies, the stability of UAV flux observation platforms is insufficient. Fluctuations in flight attitude can easily cause sensor measurement references to shift. Attitude changes in complex terrain or turbulent environments introduce additional measurement errors, affecting data reliability. They are difficult to meet the needs of long-term, large-scale observation and cannot adapt to the demand for high-precision flux data in large-scale applications such as scientific research monitoring and ecological assessment. In view of this, we propose a method and system for a multi-rotor UAV airborne flux observation platform. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for a multi-rotor UAV-borne flux observation platform, which solves the problem that existing UAV flux observation platforms cannot meet the demand for high-precision flux data in large-scale applications such as scientific research monitoring and ecological assessment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and system for a multi-rotor unmanned aerial vehicle (UAV) airborne flux observation platform, comprising the following steps:

[0007] S1 Task Planning and System Initialization Steps:

[0008] Based on terrain scanning data and underlying surface classification results, the multi-rotor UAV uses a deep learning model to generate an adaptive observation path containing height gradient layers, and simultaneously completes the spatiotemporal synchronous calibration of the three-dimensional ultrasonic anemometer, GPS, attitude sensor, and open-circuit carbon dioxide and water vapor analyzer to obtain the raw sensor data.

[0009] S2 dynamic data acquisition steps:

[0010] The multi-rotor UAV flies along a planned path, and high-frequency synchronous acquisition of three-dimensional wind speed, CO2 / H2O concentration, and six-degree-of-freedom attitude data is achieved through an FPGA hardware triggering system. The flight parameters of the multi-rotor UAV are dynamically adjusted according to the real-time turbulence intensity index.

[0011] S3 Data Preprocessing and Quality Control Steps:

[0012] A wavelet packet-EMD hybrid denoising algorithm combined with a density peak clustering anomaly removal model is used to perform spatiotemporal registration and quality labeling on the original sensor data in step S1, and to construct a multi-sensor data spatiotemporal cube.

[0013] S4 flux calculation and multidimensional correction steps:

[0014] Based on the improved eddy covariance model, combined with the UAV motion compensation model and atmospheric stability classification algorithm, the carbon / water / heat flux during the UAV operation in step S2 is calculated, and the flux footprint is corrected by the roughness spectrum model optimized by the genetic algorithm.

[0015] S5 data transmission and distributed storage steps:

[0016] An adaptive coding algorithm based on generative adversarial networks is adopted, and the data of the spatiotemporal cube of multi-sensor data in step S3 and the data after flux footprint correction in step S4 are transmitted in real time through a 5G / data radio dual-channel link to build a distributed database cluster of spatiotemporal index at the ground end.

[0017] S6 Multi-Source Data Fusion and Application Analysis Steps:

[0018] For the data from the distributed database cluster in step S5, a CNN-LSTM hybrid network is used to extract the time-frequency features of the throughput data, and the EnSRF multi-source assimilation algorithm is combined to fuse satellite remote sensing and ground observation data to output professional application products in the fields of ecology, atmosphere and agriculture.

[0019] Preferably, in step S1, the deep learning model uses a U-Net network architecture to perform semantic segmentation on the terrain LiDAR point cloud data, generating a surface feature map that includes vegetation height and water distribution; the observation path includes a spiral ascending trajectory for vertical profile measurement and a surface source scanning path based on fractal geometry.

[0020] Preferably, in step S2, the FPGA hardware triggering system achieves a synchronization accuracy of ±10μs for each sensor; the three-dimensional ultrasonic anemometer adopts a four-probe orthogonal array structure with a sampling frequency ≥30Hz; and the dynamic adjustment algorithm for the flight parameters of the multi-rotor UAV is based on the real-time calculation results of the Reynolds stress tensor.

[0021] Preferably, in step S3, spatiotemporal registration refers to constructing a six-dimensional transformation matrix that includes sensor position deviation, time delay, and coordinate rotation; quality labeling refers to using a random forest classifier to score the data quality at five levels.

[0022] Preferably, in step S4, the improved eddy covariance model introduces a three-dimensional correction factor for the flight velocity vector decomposition; the atmospheric stability classification algorithm uses a Richardson number classifier based on deep learning; and the roughness spectrum model optimizes the MO similarity theory parameters through a genetic algorithm.

[0023] Preferably, in the data transmission and distributed storage steps, the adaptive coding algorithm dynamically adjusts the LDPC code rate based on channel state information, and the distributed database adopts a spatiotemporal cube index structure, supporting real-time writing and multi-scale spatiotemporal query of high-frequency data ≥100Hz.

[0024] Preferably, in step S6, the CNN-LSTM hybrid network realizes the time-frequency feature decomposition of the flux data; the EnSRF multi-source assimilation algorithm constructs a state vector including boundary layer parameters, vegetation index, and soil moisture; and the professional application products include carbon cycle model-driven data, crop evapotranspiration estimation map, and atmospheric pollutant diffusion simulation results.

[0025] Preferably, step S1 further includes dynamically optimizing the observation path based on a reinforcement learning algorithm, adjusting the sampling strategy according to a real-time meteorological data prediction model, and constructing a sensor error propagation model for uncertainty analysis.

[0026] Preferably, step S3 further includes using a Bayesian fusion algorithm to perform spatiotemporal consistency verification on multi-sensor data, combining a Kalman filter algorithm to perform dynamic system state estimation, and constructing a data quality traceability chain to achieve full-process quality control.

[0027] A multi-rotor unmanned aerial vehicle (UAV) airborne flux observation platform system includes the following components:

[0028] The device includes a multi-rotor drone and an airborne flux mounting bracket and a data acquisition unit mounted on the multi-rotor drone. The top of the airborne flux mounting bracket is equipped with a three-dimensional ultrasonic anemometer, GPS, attitude sensor, and open-circuit carbon dioxide and water vapor analyzer. The data acquisition unit has a built-in airborne flux processing program and is electrically connected to the three-dimensional ultrasonic anemometer, GPS, attitude sensor, and open-circuit carbon dioxide and water vapor analyzer, respectively.

[0029] This invention provides a method and system for an airborne flux observation platform for multi-rotor unmanned aerial vehicles (UAVs). It has the following beneficial effects:

[0030] 1. In this invention, the UAV airborne platform has good stability, which can maintain the stable operation of the sensor during flight, reduce the impact of factors such as changes in flight attitude on the observation data, thereby improving the stability and reliability of the data. In addition, the operation and maintenance cost of the UAV is relatively low, and there is no need for a large amount of manpower and material resources, such as no need to build tower bases or lay cables, which can effectively reduce the cost of long-term observation and improve the sustainability of observation.

[0031] 2. When conducting observations in dangerous areas or harsh environments, this invention eliminates the need for personnel to enter the site, thus avoiding potential safety risks for personnel. For example, when conducting observations in disaster areas such as fires, floods, and pollution, it can ensure the safety of the observation personnel. Furthermore, the cost of the drone itself is relatively low, and even if it is accidentally damaged during flight, it will not cause excessive economic losses. Compared with large observation equipment, it reduces the risk of equipment loss.

[0032] 3. This invention combines a CNN-LSTM hybrid network with the EnSRF assimilation algorithm to achieve high-precision decomposition of spatiotemporal features of UAV throughput data and fusion of multi-source data, thereby improving the spatiotemporal resolution and accuracy of throughput observation, reducing errors, optimizing model-driven capabilities, and providing efficient and accurate data support and decision-making basis for fields such as ecological carbon cycle and precision agriculture.

[0033] 4. This invention achieves ±10μs synchronization accuracy of the sensor through an FPGA hardware triggering system, ensuring data time consistency; the four-probe orthogonal array structure combined with high-frequency sampling improves the accuracy of wind speed measurement; the flight parameter dynamic adjustment algorithm based on Reynolds stress tensor enables the UAV to adaptively adjust speed and altitude in different turbulent environments, ensuring observation stability, while also enabling high-frequency observation, and can capture dynamic processes of flux that change rapidly in a short time, such as the diurnal rhythm of plant photosynthesis and respiration, as well as the instantaneous impact of weather changes and sudden events on flux. Attached Figure Description

[0034] Figure 1 A three-dimensional view of the airborne flux observation platform system of this multi-rotor UAV;

[0035] Figure 2 Here is a flowchart of the method for the multi-rotor UAV airborne flux observation platform;

[0036] Figure 3 This is a schematic diagram of the FPGA hardware triggering system of the present invention. Detailed Implementation

[0037] The technical solutions in 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.

[0038] Example:

[0039] Please see the appendix Figure 1 - Appendix Figure 3 This invention provides a method for a multi-rotor unmanned aerial vehicle (UAV) airborne flux observation platform, comprising the following steps:

[0040] S1 Task Planning and System Initialization Steps:

[0041] Based on terrain scanning data and underlying surface classification results, the multi-rotor UAV uses a deep learning model to generate an adaptive observation path containing height gradient layers, and simultaneously completes the spatiotemporal synchronous calibration of the three-dimensional ultrasonic anemometer, GPS, attitude sensor, and open-circuit carbon dioxide and water vapor analyzer to obtain the raw sensor data.

[0042] S2 dynamic data acquisition steps:

[0043] The multi-rotor UAV flies along a planned path, and high-frequency synchronous acquisition of three-dimensional wind speed, CO2 / H2O concentration, and six-degree-of-freedom attitude data is achieved through an FPGA hardware triggering system. The flight parameters of the multi-rotor UAV are dynamically adjusted according to the real-time turbulence intensity index.

[0044] S3 Data Preprocessing and Quality Control Steps:

[0045] A wavelet packet-EMD hybrid denoising algorithm combined with a density peak clustering anomaly removal model is used to perform spatiotemporal registration and quality labeling on the original sensor data in step S1, and to construct a multi-sensor data spatiotemporal cube.

[0046] S4 flux calculation and multidimensional correction steps:

[0047] Based on the improved eddy covariance model, combined with the UAV motion compensation model and atmospheric stability classification algorithm, the carbon / water / heat flux during the UAV operation in step S2 is calculated, and the flux footprint is corrected by the roughness spectrum model optimized by the genetic algorithm.

[0048] S5 data transmission and distributed storage steps:

[0049] An adaptive coding algorithm based on generative adversarial networks is adopted, and the data of the spatiotemporal cube of multi-sensor data in step S3 and the data after flux footprint correction in step S4 are transmitted in real time through a 5G / data radio dual-channel link to build a distributed database cluster of spatiotemporal index at the ground end.

[0050] S6 Multi-Source Data Fusion and Application Analysis Steps:

[0051] For the data from the distributed database cluster in step S5, a CNN-LSTM hybrid network is used to extract the time-frequency features of the throughput data, and the EnSRF multi-source assimilation algorithm is combined to fuse satellite remote sensing and ground observation data to output professional application products in the fields of ecology, atmosphere and agriculture.

[0052] In step S1, the deep learning model uses the U-Net network architecture to perform semantic segmentation on the terrain LiDAR point cloud data, generating a surface feature map that includes vegetation height and water distribution; the observation path includes a spiral ascending trajectory for vertical profile measurement and a surface source scanning path based on fractal geometry.

[0053] In step S2, the FPGA hardware triggering system achieves a synchronization accuracy of ±10μs for each sensor; the three-dimensional ultrasonic anemometer adopts a four-probe orthogonal array structure with a sampling frequency ≥30Hz; the dynamic adjustment algorithm for the flight parameters of the multi-rotor UAV is based on the real-time calculation results of the Reynolds stress tensor, and includes the following algorithm steps:

[0054] 1. FPGA hardware-triggered synchronization mechanism initialization

[0055] The FPGA clock management unit is activated, configuring the global clock frequency to 100MHz (period 10ns). Multiple synchronous trigger pulses are generated using a hardware description language (such as Verilog). The trigger signals are distributed to each sensor via a low-latency buffer, ensuring the trigger timing t... trigger With sensor sampling time t sensor Satisfy |t sensor -t trigger |≤10μs. For example, for six sensors such as a three-dimensional ultrasonic anemometer and a CO2 analyzer, a time-division multiplexing trigger mode is adopted, with a trigger interval of 33.3ms (corresponding to a 30Hz sampling rate). The timestamps of each sensor are latched by a register, and the error is recorded via Δt=t sensor -t trigger Recorded and used for subsequent spatiotemporal registration, wherein:

[0056] t trigger : Trigger signal timing;

[0057] t sensor Sensor sampling time;

[0058] Δt: The time difference between the sensor sampling time and the triggering time, Δt = t sensor -t trigger ;

[0059] Δt sync Multi-sensor data synchronization error, Δt sync =t max -t min , where t max t represents the maximum sampling time for each sensor. min This represents the minimum sampling time for each sensor.

[0060] 2. Wind speed calculation using a four-probe orthogonal array

[0061] The four probes are arranged in an X-shaped orthogonal configuration (e.g., probes 1-3 along the X-axis, probes 2-4 along the Y-axis, and probes 5-6 along the Z-axis). The time difference Δt of the ultrasonic wave propagation in the air... i Measured by an FPGA counter (1 ns resolution). For the i-th probe pair, the sound path L i =0.15m, sound speed c = 340m / s, then the three-dimensional wind speed components are calculated as follows:

[0062]

[0063] During real-time calculations, each time difference data set (30Hz) is acquired, and the above calculations are performed through a pipelined multiplier. The results are stored in a dual-port RAM for subsequent processing. For example, when Δt1 = 441μs, Δt2 = 439μs, and Δt3 = 440μs, then u = 1.0m / s, v = 0.1m / s, and w = 0.5m / s, where:

[0064] Δt i : The ultrasonic wave propagation time difference of the i-th pair of probes;

[0065] L i : The acoustic path of the i-th probe pair (unit: m);

[0066] c: Speed ​​of sound (unit: m / s);

[0067] u, v, w: Three-dimensional wind speed components (unit: m / s), corresponding to the X, Y, and Z axes respectively.

[0068] 3. Real-time calculation of Reynolds stress tensor: A sliding window is opened in the FPGA to cache the most recent N = 1024 sets of three-dimensional wind speed data (u_k, v_k, w_k) (k is the data index, k = 1, 2, ..., N). Each time a new set of data is added, the Reynolds stress components are updated according to the following formula:

[0069] The formula for calculating the time mean (mean within a sliding window) of each component (u, v, w) of the three-dimensional wind speed is as follows:

[0070]

[0071] in:

[0072] u k v k w k The k-th group of three-dimensional instantaneous wind speed values ​​(unit: m / s), corresponding to the X, Y, and Z axis directions respectively;

[0073] Mean values ​​of wind speed components along the X, Y, and Z axes within the sliding window (unit: m / s);

[0074] N = 1024: The number of data sets buffered by the sliding window (fixed value)

[0075] Definition of wind speed component fluctuation value

[0076] The fluctuation value is the "deviation between the instantaneous value and the mean value," reflecting the random fluctuation characteristics of turbulence. The formula is:

[0077]

[0078] Where u′, v′, w′ are the fluctuating values ​​of the wind speed components along the X, Y, and Z axes (unit: m / s).

[0079] Reynolds stress component calculation

[0080] The Reynolds stress tensor comprises a normal component (diagonal element) and a tangential component (off-diagonal element). The core focus is on the normal component, which relates to turbulence intensity (reflecting the energy of turbulent fluctuations in all directions), as shown in the following formula:

[0081] Normal Reynolds stress (core component):

[0082]

[0083]

[0084] Tangential Reynolds stress (supplemented with the complete tensor):

[0085]

[0086] Where: the unit of Reynolds stress component is (m / s) 2 The normal component is always non-negative (energy term), while the tangential component can be positive or negative (reflecting the direction of pulsating coupling).

[0087] Turbulence Intensity Index (TI) Calculation

[0088] Average wind speed The three-dimensional average wind speed model (composite wind speed) reflects the overall velocity of the airflow:

[0089]

[0090] Turbulence Intensity Index (TI):

[0091]

[0092] in:

[0093] molecular The root mean square (RMS) of three-dimensional turbulent fluctuations (unit: m / s) represents the average intensity of the turbulent fluctuations.

[0094] denominator Three-dimensional average wind speed (unit: m / s);

[0095] TI is a dimensionless quantity. In engineering, the threshold is usually set to 0.15 (when TI≥0.15, the turbulence is severe and the flight parameters need to be adjusted).

[0096] Example:

[0097] Given a sliding window parameter N = 1024, input the normal Reynolds stress components:

[0098] Normal Reynolds stress:

[0099] Average wind speed:

[0100] Calculation steps:

[0101] Calculate the molecule (root mean square value of turbulent fluctuations):

[0102]

[0103] Calculate TI:

[0104]

[0105] (Threshold set to 0.15)

[0106] Parameter description:

[0107] u k v k w k : Instantaneous three-dimensional wind speed values ​​for the kth group (unit: m / s);

[0108] The mean values ​​(unit: m / s) of the wind speed components along the X, Y, and Z axes within the sliding window are calculated as follows: (The same applies to v and w);

[0109] u ′ v′ w ′ The fluctuation value of the wind speed component (unit: m / s) is calculated as follows:

[0110] Normal Reynolds stress components (unit: (m / s)) 2 ), the calculation method is (The same applies to other normal components);

[0111] Tangential Reynolds stress components (unit: (m / s)) 2 ), the calculation method is (The same applies to other tangential components);

[0112] TI: Turbulence Intensity Index (dimensionless), calculated as follows: Three-dimensional average wind speed (unit: m / s), calculated as follows:

[0113] 4. Dynamic adjustment of flight parameters

[0114] When the calculated TI ≥ 0.15, the flight parameter adjustment module is triggered.

[0115] The flight speed V is adjusted according to the exponential decay formula: V=V0·exp(-α·(TI-TI) thr ), where V0=10m / s, α=2, TI thr =0.15. For example, when TI = 0.2, V = 10·exp(-2×0.05)≈9.05m / s.

[0116] Flight altitude H is compensated using a PID algorithm: Where the error e = TI - TI thr proportionality coefficient k p =5m, integral coefficient k i =0.1m / s, differential coefficient k d =1m, where:

[0117] V0: Initial flight speed (unit: m / s);

[0118] V: Adjusted flight speed (unit: m / s);

[0119] α: Speed ​​adjustment coefficient;

[0120] H0: Initial flight altitude (unit: m);

[0121] H new Adjusted flight altitude (unit: m);

[0122] k p ,k i ,k d The proportional, integral, and derivative coefficients of a PID controller;

[0123] e: Error, e = TI - TI thr ;

[0124] 5. Multi-sensor data synchronization and caching

[0125] The 30Hz wind speed data, 10Hz gas concentration data (interpolated to 30Hz via FPGA), and 100Hz attitude data (downsampled to 30Hz) are aligned by timestamps and stored in a DDR3 cache. Synchronization error is measured via Δt. sync =t max -t min Monitoring, when Δt sync A resynchronization mechanism is triggered when the time exceeds 15μs, correcting the data phase through linear interpolation, where:

[0126] t: Target alignment timestamp (unit: s), which is the final unified sampling time. Wind speed, gas concentration, and attitude data must all be matched to this timestamp.

[0127] t1: The original sampling time of gas concentration data (unit: s), corresponding to the time point at the original sampling frequency of 10Hz;

[0128] t2: The next sampling time adjacent to the gas concentration data (unit: s), t2>t1;

[0129] c(t1): Gas concentration at time t1 (unit: e.g., ppm);

[0130] c(t2): Gas concentration at time t2 (unit: e.g., ppm);

[0131] c(t): The gas concentration at time t obtained through linear interpolation (unit: e.g., ppm), calculated using the following formula:

[0132] Δt sync Multi-sensor data synchronization error (unit: μs), which is the difference between the maximum and minimum timestamps of all sensor data, Δt. sync =t max -t min , where t max t represents the largest timestamp among the wind speed, gas concentration, and attitude data. min It is the smallest timestamp among the three.

[0133] 6. Data Packaging and Transmission

[0134] Each 30Hz data frame contains: 3D wind speed (3×4B), CO2 / H2O concentration (2×4B), UAV attitude (pitch / roll / yaw angle 3×4B), and GPS coordinates (2×8B), for a total of 112 bytes. It is transmitted to the data acquisition unit via LVDS differential signal lines at a transmission rate of 112B×30Hz = 33.6KB / s, using an 8B / 10B encoding protocol, with a bit error rate controlled within 10%. -7 The following applies: A 16-bit CRC checksum is added to the frame header, using the formula CRC-16=X. 16 +X 15 +X 2 +1 ensures reliable data transmission.

[0135] In step S3, spatiotemporal registration refers to constructing a six-dimensional transformation matrix that includes sensor position deviation, time delay, and coordinate rotation; quality labeling refers to using a random forest classifier to score the data quality at five levels.

[0136] In step S4, the improved eddy covariance model introduces a three-dimensional correction factor for the flight velocity vector decomposition; the atmospheric stability classification algorithm uses a Richardson number classifier based on deep learning; and the roughness spectrum model optimizes the MO similarity theory parameters through a genetic algorithm.

[0137] In the data transmission and distributed storage steps, the adaptive coding algorithm dynamically adjusts the LDPC code rate based on channel state information, and the distributed database adopts a spatiotemporal cube index structure, supporting real-time writing and multi-scale spatiotemporal query of high-frequency data ≥100Hz.

[0138] In step S6, the CNN-LSTM hybrid network decomposes the time-frequency features of the flux data; the EnSRF multi-source assimilation algorithm constructs a state vector containing boundary layer parameters, vegetation index, and soil moisture; professional application products include carbon cycle model-driven data, crop evapotranspiration estimation maps, and atmospheric pollutant diffusion simulation results, including the following steps:

[0139] I. Feature Decomposition Process of CNN-LSTM Hybrid Network

[0140] 1. Data Preprocessing and Input Construction

[0141] The flux time series (such as CO2 flux and water vapor flux) are reshaped into a three-dimensional tensor along the time-space dimension. Where T is the time step (e.g., 8 observations per day), W is the spatial grid dimension (e.g., a 10×10 underlying surface grid), and C is the number of feature channels (flux value + meteorological elements). For example, with an input dimension of 8×10×10×1, the data is standardized to conform to...

[0142] 2. CNN Spatial Feature Extraction

[0143] Two-dimensional convolutional layers are used to perform operations on spatial dimensions: Let the convolution kernel be... (e.g., k=3, F=64), then the output of the spatial position (w, w′) at time step t is:

[0144]

[0145] After applying the ReLU activation function ReLU(x) = max(0,x), spatial features such as surface roughness and vegetation cover are extracted. For example, after three layers of convolution, the spatial dimension is compressed to 5×5, and the number of channels increases to 64.

[0146] 3. Feature Dimension Transformation

[0147] The 3D features Y∈R output by the CNN T×W′×F′ (e.g., T = 8, W′ = 5, F′ = 64) Reconstructed into a two-dimensional sequence Z∈R T ×(W′×F′) That is, the input dimension of each time step is 5×5×64=1600, which serves as the time series input for LSTM.

[0148] 4. LSTM Time Series Modeling

[0149] For each time step t, the LSTM unit updates its state using the following formula:

[0150] Forget Gate: Determines the proportion of historical information discarded, f t =σ(W f ·[z t ,h t-1 ]+b f )

[0151] Input gate: confirms new information input, i t =σ(W i ·[z t ,h t-1 ]+b i ),

[0152] Cell state update: integrating historical and current information.

[0153] Output gate: generates hidden state, o t =σ(W o ·[z t ,h t-1 ]+b o ),h t =o t ⊙tanh(c t )

[0154] Where z t h is the current input. t-1 ,c t-1 Let W be the hidden state and cell state at the previous time step, σ be the Sigmoid function, and W be the hidden state and cell state at the previous time step. f W i W c W o and b f ,b i ,b c ,b o These are learnable parameters. A two-layer LSTM is used to capture the diurnal / seasonal fluctuations in flux.

[0155] 5. Feature Fusion and Output

[0156] The hidden state h of the last LSTM layer T Mapped to the target dimension via a fully connected layer, such as through FC(h) T )∈R D (D=10) The time-frequency characteristic decomposition results are obtained, which characterize the principal components of flux change (such as periodic changes and sudden fluctuations).

[0157] The algebraic meaning in this algorithm is as follows:

[0158] X: Input 3D tensor Where T is the time step, W is the spatial grid dimension, and C is the number of feature channels;

[0159] K: Convolution kernel, Where k is the kernel size and F is the number of output channels;

[0160] Y t,w,w′,f : The convolution output of the f-th channel at time step t, spatial location (w, w′);

[0161] b f The bias term for the f-th channel;

[0162] Z: The reshaped two-dimensional sequence Where W′ is the spatial dimension after convolution, and F′ is the number of channels after convolution;

[0163] f t The output of the forget gate at time :t;

[0164] i t Input gate output at time t;

[0165] The state of candidate cells at any given time;

[0166] c t Cell state at time t;

[0167] o t Output gate output at time t;

[0168] h t The state is hidden at time :t;

[0169] z t The input to the LSTM at time t;

[0170] h t-1 c t-1 The hidden state and cell state at time t-1;

[0171] W f W i W c W o Weight matrices for the forget gate, input gate, cell state, and output gate;

[0172] b f ,b i ,b c ,b o Bias terms for forget gate, input gate, cell state, and output gate;

[0173] FC(h T ): Fully connected layer output, Where D is the target dimension;

[0174] II. EnSRF Assimilation Algorithm Multi-Source Data Fusion Process

[0175] 1. State space definition and set initialization

[0176] Define a state vector x∈R n (Including boundary layer parameters, NDVI, soil moisture, etc.), observation vector (UAV flux observation + satellite inversion LST + ground tower data). Generate N = 100 set members. The initial mean and covariance are:

[0177]

[0178] The initial set is constructed by overlaying historical data from climate models. Anti-dynamic generation, Q0 is the initial error covariance.

[0179] 2. Model Prediction Update

[0180] Predicting set members using a physical model M (such as a boundary layer parameterized model):

[0181]

[0182] Calculate the predicted mean and covariance:

[0183]

[0184] 3. Observation projection and gain matrix calculation

[0185] The predicted state is projected onto the observation space through the observation operator H:

[0186]

[0187] Calculate the observed and predicted mean The Kalman gain is calculated using the following formula:

[0188]

[0189] Where R t For observation error covariance (such as the error matrix retrieved from satellite data)

[0190] 4. State assimilation update

[0191] Using actual observed values ​​y t Update the collection members: in This is the actual observed mean. The updated analysis set's mean. This is the fusion result. For example, when assimilating CO2 flux from drones with OCO-3 satellite data, this step is used to correct the parameters of the carbon cycle model.

[0192] 5. Application-driven and model output

[0193] The assimilated state vector Input professional application model:

[0194] Crop evapotranspiration estimation: based on the Penman-Monteith formula. Where Δ is the saturated vapor pressure slope, R n Where ρ is net radiation, G is soil heat flux, and ρC is the soil heat flux. p Where is the air heat capacity, D is the water vapor pressure difference, and r is the air heat capacity. a For aerodynamic drag, r s Soil surface resistance (corrected by NDVI assimilated by EnSRF).

[0195] Atmospheric pollutant diffusion simulation: The assimilated wind field and turbulence parameters are input into the WRF-chem model to update the initial conditions of the diffusion equation.

[0196]

[0197] Where C is the pollutant concentration, u, v, w are the wind speed components, and K...x K y K z K is the diffusion coefficient, S is the source-sink term, and K is optimized using boundary layer parameters assimilated by EnSRF. z Vertical distribution;

[0198] The algebraic meaning in this algorithm is as follows:

[0199] x: state vector n is the dimension of the state vector;

[0200] y: observation vector m is the dimension of the observation vector;

[0201] N: Number of members in the set;

[0202] The analysis value of the i-th set member at the initial moment;

[0203] The mean of the initial analysis values,

[0204] Analyze the error covariance matrix at the initial time.

[0205] ∈: perturbation term, ∈~N(0,Q0), where Q0 is the initial error covariance;

[0206] The predicted value of the i-th set member at time i;

[0207] M: Physical model;

[0208] Predict the disturbance term at any time. Q t Let be the prediction error covariance at time t;

[0209] The mean of the predicted values ​​at each time point.

[0210] Time-prediction error covariance matrix;

[0211] The predicted observation value of the i-th set member at time i;

[0212] H: Observation operator

[0213] The mean of the predicted values ​​observed at each time point.

[0214] K t Kalman gain at time t

[0215] R t :Observation error covariance matrix at time t

[0216] y t Actual observed value at time t

[0217] mean of actual observed values ​​at time 1

[0218] Analysis update value of the i-th set member at time i

[0219] Analyze the mean of the updated values ​​at any time.

[0220] III. Algorithm Linkage and Data Closed Loop

[0221] The time-frequency features extracted by CNN-LSTM (such as flux periodicity components) serve as prior constraints for EnSRF assimilation, while the high-resolution state data fused by EnSRF (such as refined soil moisture) feeds back into CNN-LSTM to improve the accuracy of feature decomposition, forming a closed-loop process of "observation - feature extraction - data assimilation - application", and finally outputting professional products such as carbon cycle driven data and evapotranspiration estimation maps.

[0222] Step S1 further includes dynamically optimizing the observation path based on reinforcement learning algorithms, adjusting the sampling strategy according to the real-time meteorological data prediction model, and constructing a sensor error propagation model for uncertainty analysis.

[0223] Step S3 also includes using a Bayesian fusion algorithm to perform spatiotemporal consistency checks on multi-sensor data, combining a Kalman filter algorithm to perform dynamic system state estimation, and constructing a data quality traceability chain to achieve full-process quality control.

[0224] A multi-rotor unmanned aerial vehicle (UAV) airborne flux observation platform system includes the following components:

[0225] The device includes a multi-rotor drone 7 and an airborne flux mounting bracket 5 and a data acquisition unit 6 mounted on the multi-rotor drone 7. The top of the airborne flux mounting bracket 5 is equipped with a three-dimensional ultrasonic anemometer 1, a GPS 2, an attitude sensor 3, and an open-circuit carbon dioxide and water vapor analyzer 4. The data acquisition unit 6 has a built-in airborne flux processing program and is electrically connected to the three-dimensional ultrasonic anemometer 1, GPS 2, attitude sensor 3, and open-circuit carbon dioxide and water vapor analyzer 4, respectively.

[0226] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for an airborne flux observation platform for a multi-rotor unmanned aerial vehicle (UAV), characterized in that, Includes the following steps: S1 Task Planning and System Initialization Steps: Based on terrain scanning data and underlying surface classification results, the multi-rotor UAV (7) uses a deep learning model to generate an adaptive observation path containing height gradient layers, and simultaneously completes the spatiotemporal synchronization calibration of the three-dimensional ultrasonic anemometer (1), GPS (2), attitude sensor (3), and open-circuit carbon dioxide water vapor analyzer (4) to obtain the original sensor data. S2 dynamic data acquisition steps: The multi-rotor UAV (7) flies along the planned path. The high-frequency synchronous acquisition of three-dimensional wind speed, CO2 / H2O concentration and six-degree-of-freedom attitude data is realized through the FPGA hardware triggering system. The flight parameters of the multi-rotor UAV (7) are dynamically adjusted according to the real-time turbulence intensity index. S3 Data Preprocessing and Quality Control Steps: A wavelet packet-EMD hybrid denoising algorithm combined with a density peak clustering anomaly removal model is used to perform spatiotemporal registration and quality labeling on the original sensor data in step S1, and to construct a multi-sensor data spatiotemporal cube. S4 flux calculation and multidimensional correction steps: Based on the improved eddy covariance model, combined with the UAV motion compensation model and atmospheric stability classification algorithm, the carbon / water / heat flux during the UAV operation in step S2 is calculated, and the flux footprint is corrected by the roughness spectrum model optimized by the genetic algorithm. S5 data transmission and distributed storage steps: An adaptive coding algorithm based on generative adversarial networks is adopted, and the data of the spatiotemporal cube of multi-sensor data in step S3 and the data after flux footprint correction in step S4 are transmitted in real time through a 5G / data radio dual-channel link to build a distributed database cluster of spatiotemporal index at the ground end. S6 Multi-Source Data Fusion and Application Analysis Steps: For the data from the distributed database cluster in step S5, a CNN-LSTM hybrid network is used to extract the time-frequency features of the throughput data, and the EnSRF multi-source assimilation algorithm is combined to fuse satellite remote sensing and ground observation data to output professional application products in the fields of ecology, atmosphere and agriculture.

2. The method for a multi-rotor UAV airborne flux observation platform according to claim 1, characterized in that, In step S1, the deep learning model uses the U-Net network architecture to perform semantic segmentation on the terrain LiDAR point cloud data, generating a surface feature map that includes vegetation height and water distribution; the observation path includes a spiral ascending trajectory for vertical profile measurement and a surface source scanning path based on fractal geometry.

3. The method for a multi-rotor UAV airborne flux observation platform according to claim 1, characterized in that, In step S2, the FPGA hardware triggering system achieves a synchronization accuracy of ±10μs for each sensor; the three-dimensional ultrasonic anemometer (1) adopts a four-probe orthogonal array structure with a sampling frequency ≥30Hz; the flight parameter dynamic adjustment algorithm of the multi-rotor UAV (7) is based on the real-time calculation results of the Reynolds stress tensor.

4. The method for a multi-rotor UAV airborne flux observation platform according to claim 1, characterized in that, In step S3, spatiotemporal registration refers to constructing a six-dimensional transformation matrix that includes sensor position deviation, time delay, and coordinate rotation; quality labeling refers to using a random forest classifier to score the data quality at five levels.

5. The method for a multi-rotor UAV-borne flux observation platform according to claim 1, characterized in that, In step S4, the improved eddy covariance model introduces a three-dimensional correction factor for the flight velocity vector decomposition; the atmospheric stability classification algorithm uses a Richardson number classifier based on deep learning; and the roughness spectrum model optimizes the MO similarity theory parameters through a genetic algorithm.

6. The method for a multi-rotor UAV airborne flux observation platform according to claim 1, characterized in that, In step S5, the adaptive coding algorithm dynamically adjusts the LDPC code rate based on channel state information; the distributed database adopts a spatiotemporal cube index structure, supporting real-time writing and multi-scale spatiotemporal query of high-frequency data ≥100Hz.

7. The method for a multi-rotor UAV-borne flux observation platform according to claim 1, characterized in that, In step S6, the CNN-LSTM hybrid network decomposes the time-frequency features of the flux data; the EnSRF multi-source assimilation algorithm constructs a state vector containing boundary layer parameters, vegetation index, and soil moisture; professional application products include carbon cycle model-driven data, crop evapotranspiration estimation maps, and atmospheric pollutant diffusion simulation results.

8. The method for a multi-rotor UAV airborne flux observation platform according to claim 2, characterized in that, Step S1 further includes dynamically optimizing the observation path based on reinforcement learning algorithms, adjusting the sampling strategy according to the real-time meteorological data prediction model, and constructing a sensor error propagation model for uncertainty analysis.

9. The method for a multi-rotor UAV airborne flux observation platform according to claim 4, characterized in that, Step S3 also includes using a Bayesian fusion algorithm to perform spatiotemporal consistency checks on multi-sensor data, combining a Kalman filter algorithm to perform dynamic system state estimation, and constructing a data quality traceability chain to achieve full-process quality control.

10. A multi-rotor UAV airborne flux observation platform system according to claim 1, characterized in that, The method for implementing any one of claims 1-9 includes a multi-rotor drone (7) and an airborne flux mounting bracket (5) and a data acquisition unit (6) mounted on the multi-rotor drone (7). The top of the airborne flux mounting bracket (5) is respectively provided with a three-dimensional ultrasonic anemometer (1), a GPS (2), an attitude sensor (3), and an open-circuit carbon dioxide water vapor analyzer (4). The data acquisition unit (6) has a built-in airborne flux processing program and is electrically connected to the three-dimensional ultrasonic anemometer (1), GPS (2), attitude sensor (3), and open-circuit carbon dioxide water vapor analyzer (4).

Images