A method for unmanned aerial vehicle hyperspectral online calibration measurement for agricultural remote sensing

Abstract

The present application relates to the technical field of agricultural remote sensing measurement, and discloses a kind of unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing, comprising: obtaining the hyperspectral data of target crop and same scene reference vegetation by unmanned aerial vehicle in real time, extracting characteristic band reflectivity difference ΔR, based on the comparison result of the time series change rate of ΔR and the preset threshold, dynamically distinguishing crop physiological change and environmental disturbance and realizing online calibration, the present application constructs self-consistent calibration mechanism using the time series stability of the spectral characteristics of crop itself, without relying on external reference material, can eliminate the influence of light mutation, simultaneously through the synchronous analysis of gravity vector and spectral drift effectively identifies lodging scene, so that the system can still keep accurate tracking to the real physiological state of crop under extreme working condition.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing, belong to agricultural remote sensing measurement technical field. BACKGROUND

[0002] In the field of agricultural remote sensing measurement, unmanned aerial vehicle hyperspectral technology realizes physiological state monitoring by capturing crop reflectance spectrum, and the existing calibration method mainly relies on two ways: physical reference calibration (such as ground whiteboard calibration) and environmental model compensation (such as radiation transfer model). The former needs to interrupt the flight to obtain reference data, and the latter requires real-time atmospheric parameter input, aerosol optical depth, solar zenith angle, etc. Although these two methods are widely used, they have fundamental limitations. Dynamic continuous spectrum monitoring is simplified to static spatial calibration, and the time evolution characteristics of illumination conditions and crop physiological state in agricultural scenes are ignored.

[0003] When the above method is applied to actual farmland operation, its systematic defects are exposed in typical scenarios: whiteboard calibration requires unmanned aerial vehicle hovering or landing, resulting in flight line interruption. In the farmland area where the cloud layer moves rapidly, the illumination condition changes 3-5 times per hour. Frequent interruptions make continuous monitoring impossible, and the timeliness of the data is lost. Radiation model relies on real-time data from weather stations, but the coverage rate of weather stations in farmland areas is less than 15%. When model parameters are missing or lagging, atmospheric scattering compensation error is transferred to reflectance calculation, ultimately distorting stress diagnosis results. In order to compensate for calibration errors, industry solutions stack multispectral sensors and edge AI chips. This approach may improve accuracy, but it violates the core trend of agricultural unmanned aerial vehicle democratization and lightweight, making it difficult to deploy the technology on a large scale.

[0004] The industry tries to improve by fusing laser radar point cloud or adding sensor frequency bands, but these solutions still face new contradictions of external dependence reinforcement and increasing computational complexity: laser radar modeling requires high-precision positioning RTK, significantly increasing hardware costs; multi-band fusion algorithms rely on GPU operations, which are difficult for edge devices to handle; existing technologies consider calibration as a physical process independent of monitoring objects, but ignore the optical stability of self-grown vegetation in farmland scenes, which can serve as a zero-cost dynamic reference. This cognitive blind spot has led the industry to rely on high-cost external equipment for a long time, making it impossible to build a low-cost calibration mechanism that is endogenous to agricultural ecological laws. Therefore, how to build a lightweight hyperspectral online calibration method that does not rely on external references, can adapt to continuous flight conditions, and avoids environmental parameter coupling, has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a kind of unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing, which mainly aims to solve the problems of dynamic continuity loss, environmental strong coupling and hardware redundancy caused by the dependence of existing spectral calibration technology on external equipment and complex model.

[0006] To achieve the above object, the present application provides a method for online calibration of hyperspectral data of unmanned aerial vehicle for agricultural remote sensing, comprising the following steps:

[0007] Step a, obtaining continuous hyperspectral data of crop area, the continuous hyperspectral data containing spectral reflectance information of multiple wave bands;

[0008] Step b, extracting multiple spectral feature wavelengths determined according to crop type and growth stage from the continuous hyperspectral data in real time, the spectral feature wavelengths being directly related to crop physiological state, and constructing a spectral feature centroid representing current hyperspectral morphology based on the multiple spectral feature wavelengths;

[0009] Step c, calculating drift amount of the spectral feature centroid in unit time in real time;

[0010] Step d, comparing the drift amount of the spectral feature centroid with a physiological change threshold determined based on statistical analysis and modeling of historical data, and performing spectral centroid drift attribution determination according to the comparison result;

[0011] Step e, when the drift amount of the spectral feature centroid does not exceed the physiological change threshold, determining that the drift of the spectral feature centroid is caused by change of crop physiological state, updating the current spectral feature centroid as a new calibration reference, and recording the drift amount as an indication parameter of change of crop health state;

[0012] Step f, when the drift amount of the spectral feature centroid exceeds the physiological change threshold, determining that the drift of the spectral feature centroid is caused by external light or change of unmanned aerial vehicle posture, not updating the calibration reference, and generating a correction factor to calibrate the continuous hyperspectral data obtained subsequently in real time.

[0013] Preferably, the multiple spectral feature wavelengths for constructing the spectral feature centroid in step b include red edge inflection point wavelength, green light reflection peak wavelength and near-infrared platform starting point wavelength.

[0014] Preferably, in step d, the physiological change threshold is determined based on statistical analysis and modeling of historical spectral data of the target crop type at different growth stages.

[0015] Preferably, the method realizes online calibration of hyperspectral data of unmanned aerial vehicle without relying on any ground physical reference whiteboard or real-time atmospheric parameter input.

[0016] Preferably, in step e, when it is determined that the drift of the spectral feature centroid is caused by change of crop physiological state, the method further comprises: taking the amplitude and direction of the centroid drift amount as an indication of degree and type of crop physiological stress.

[0017] Preferably, in step f, when the drift of the spectral feature centroid exceeds the physiological change threshold, the method further comprises the following sub-steps: step f1, obtaining the current flight attitude data from the UAV on-board inertial measurement unit; step f2, performing crop lodging observation scene determination based on the flight attitude data; step f3, calling a pre-established lodging attitude-spectral morphology response model, and correcting the current spectral feature centroid according to the model to adapt to the specific spectral response characteristics of the lodging crop.

[0018] Preferably, in step f2, the crop lodging observation scene determination is performed by comparing the instantaneous change rate of the flight attitude angle with a pre-determined maneuvering flight angle threshold, and if the change rate exceeds the maneuvering flight angle threshold and is synchronized in time with the drift of the spectral feature centroid, it is determined to be in the crop lodging observation scene.

[0019] Preferably, the lodging attitude-spectral morphology response model is a calibrated and stored lookup table or low-order polynomial function, which is established by data collection and model training on the spectral response of simulated lodging crops at different inclination angles.

[0020] Preferably, the method further comprises performing time series micro-oscillation energy analysis on the spectral feature centroid for early warning of crop physiological stress, the analysis comprising the following steps: step g1, buffering the position data of the spectral feature centroid within a continuous time period at a high sampling frequency to form a centroid time series signal; step g2, processing the centroid time series signal to calculate a total energy index of micro-oscillation, which is obtained by calculating the variance of the centroid time series signal; step g3, comparing the total energy index of micro-oscillation with a physiological resting baseline in real time, the physiological resting baseline representing the normal energy level of crop micro-oscillation, and outputting early stress warning information when the index continuously exceeds the physiological resting baseline.

[0021] Preferably, in step g2, the variance of the centroid time series signal is calculated as follows: wherein, represents the spectral feature centroid collected at the i-th time point, represents the total number of data points of the buffered centroid time series signal, represents the average value of the buffered centroid time series signal. Compared with the prior art, the present application has the following advantages:

[0022]

[0023] ​​1. By constructing the time series drift analysis mechanism of the spectral feature centroid of crops, the decoupling of environmental interference and physiological changes is converted into the recognition of the essential difference in drift rate. When the centroid change rate is lower than the physiological change threshold, the system absorbs it as a new calibration reference. When the rate abnormally jumps, it triggers instantaneous correction. This mechanism makes the spectral calibration process free from dependence on external references or atmospheric parameters, forming a stable measurement closed loop in a dynamic farmland environment.

[0024] 2. When the unmanned aerial vehicle experiences a dramatic change in attitude due to observing the lodging crops, the system actively identifies the physical form mutation scene through the strong spatio-temporal synchronization of flight attitude and spectral drift, calls the pre-set lodging attitude-spectral response model, converts the gravity vector into a physical scale for correcting the spectral centroid, and makes the physiological state of crops in the lodging area still be continuously tracked. This makes the traditional measurement blind area into a post-disaster diagnosis window that can be quantitatively evaluated. By analyzing the degree of energy deviation of the micro-oscillation of the high-frequency calculated spectral centroid time series data stream from the healthy baseline, the signs of physiological regulation disorder caused by water stress are identified when the macro-physical form of crops has not changed. This interpretation of the spectral heartbeat rhythm provides an early warning signal in the intervention golden window period before the conventional diagnosis process is started.

[0025] 3. The basic calibration mechanism, the lodging adaptation mechanism and the early warning mechanism are not functionally superimposed, but form three layers of cognitive transition: the first layer filters out environmental noise and locks the essence of physiological changes; the second layer penetrates through physical form changes and maintains diagnostic continuity; the third layer perceives hidden signs of instability, expands the time boundary of monitoring, and forms a synergistic enhancement network with low resource consumption through sharing of centroid calculation data stream; the output data evolves from traditional reflectance spectrum to multi-dimensional diagnosis set integrating centroid drift trajectory, lodging correction parameters and micro-oscillation energy index, which not only avoids the interference of redundant environmental information on decision-making, but also converts complex spectral data into biological language representing the spatio-temporal evolution logic of crops, reducing the interpretation threshold of agronomists. BRIEF DESCRIPTION OF DRAWINGS

[0026] Fig. 1 The timing diagram of the crop hyperspectral online calibration measurement method of the present application;

[0027] Fig. 2 The spectral reflectance change graph of the crop growth stage of the present application;

[0028] Fig. 3 The flow chart of crop hyperspectral data processing and analysis of the present application.

[0029] The purpose of the present application, the functional characteristics and the advantages will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0030] In order to make the purposes, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in combination with specific examples.

[0031] The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing disclosed in the embodiments of the present application aims to construct a closed-loop adaptive spectral data processing framework, which fundamentally uses the inherent time sequence stability of the spectral characteristics of crops as a dynamic reference to accurately distinguish and real-time calibrate data fluctuations caused by external environmental factors (such as light mutation), and deeply mines spectral time sequence information to realize accurate tracking and early stress warning of the physiological state of crops.

[0032] The method framework, in specific implementation, mainly follows the core stages of real-time hyperspectral data acquisition and feature extraction, attribution analysis of spectral feature centroid temporal drift, and adaptive calibration and physiological state diagnosis based on attribution results, and its complete operation process can be explained in a typical field inspection scene; assuming that a UAV equipped with a hyperspectral imager is conducting normal monitoring operations on a rice field in the jointing stage, the specific implementation steps of the method are as follows: first, the UAV flies above the crop canopy along the preset flight route, and the hyperspectral imager on board continuously collects image data of the rice canopy below at a predetermined frequency, and generates a continuous hyperspectral data cube covering the visible light to near-infrared range and containing hundreds of narrow-band spectral reflectance information in real time, which constitutes the original input for all subsequent analyses; then, the on-board or ground processing system analyzes the data stream in real time, i.e., extracts the spectral feature wavelengths most closely related to the current growth stage and physiological state of the rice from the continuous hyperspectral data; according to the prior knowledge of agronomy, for rice in the jointing stage, such feature wavelengths are accurately determined as the red edge inflection point wavelength, the green light reflection peak wavelength, and the near-infrared platform starting point wavelength, which respectively represent the physiological status of the crop from three dimensions of chlorophyll content, cell structure health, and canopy water content; instead of using these three discrete wavelength values in isolation, the system performs weighted fusion of them in the spectral dimension to construct a multi-dimensional spectral feature centroid, which can be conceptually regarded as a condensation point of the current comprehensive physiological state of the crop in the spectral space, and the dynamic change of its position can macroscopically and sensitively represent the evolution of the overall spectral morphology of the crop; then, the system enters a continuous monitoring cycle, continuously locates the spectral feature centroid in a very short time scale (e.g., seconds or sub-seconds), and calculates its drift in unit time, which is a vector that accurately quantifies the speed and specific trend of the spectral morphology at the moment; at this time, the system starts the core decision logic, compares the real-time calculated spectral feature centroid drift with a preset physiological change threshold, and determines the root cause of the drift; the physiological change threshold is established by deep statistical analysis and modeling of a large amount of historical spectral data of the target crop at a specific growth stage according to the above principles, for example, the model learns the range of natural rhythmic fluctuations of the spectrum caused by physiological activities such as normal photosynthesis and water transpiration of rice in the jointing stage, and gives a reasonable upper limit of the drift rate, which thus scientifically defines the maximum change rate of the spectral morphology that can be caused by the physiological evolution of the crop under stable external environment.

[0033] Based on the comparison result, the system will automatically be shunted to two different processing paths: one, if the drift amount of spectral feature centroid does not exceed the preset physiological change threshold, the system determines that the drift is caused by the slow and real physiological state evolution of crops, such as the gradual decrease of leaf water content due to daytime continuous transpiration; in this case, the system updates the current spectral feature centroid position which has experienced a small drift as the new calibration reference, which means that the zero point of the measurement system can actively and slowly adapt to the real growth rhythm of crops, thereby ensuring the long-term accuracy of the measurement; at the same time, as the deepening application defined in the invention, the system also records and analyzes the amplitude and direction of the centroid drift, a continuous small drift trajectory towards a specific spectral region may be accurately interpreted as an early signal of a certain physiological stress (such as mild nitrogen deficiency), and output as an indication parameter of crop health status change, the second, if the drift amount of spectral feature centroid significantly exceeds the physiological change threshold, for example, due to the rapid movement of clouds in the sky, the light intensity decreases sharply within a few seconds, and then causes the reflectance of all wavebands to jump sharply and synchronously, which is a non-physiological change; at this time, the system will lock the calibration reference without updating, and immediately calculate a dynamic correction factor according to the amplitude and direction of the drift, which will be applied to all subsequent collected hyperspectral data until the external disturbance disappears, thereby realizing online real-time calibration of environmental light mutation without relying on any ground physical reference whiteboard or real-time atmospheric parameter input, ensuring the continuity and internal consistency of the monitoring data stream; further, in a specific scene, the system will also activate a more detailed sub-process to deal with the defined crop lodging and other special observation conditions: when the system monitors a threshold-exceeding sharp spectral drift, it will obtain the current flight attitude data from the inertial measurement unit (IMU) on board the unmanned aerial vehicle; then the system will analyze the instantaneous change rate of the flight attitude angle, if the rate exceeds the preset flight angle threshold representing sharp maneuvering, and the attitude change is strictly synchronous with the sharp drift of the spectral feature centroid in timestamp, the system is highly confident that it is currently in the observation scene of crop lodging; once the determination is established, the system calls a pre-established lodging attitude-spectral morphology response model, which is a lookup table or low-order polynomial function calibrated and stored after a large amount of data collection and model training on the spectral response of simulated lodging crops under different inclination angles, which accurately maps the internal law of spectral morphology change (such as increased shadow ratio, multiple scattering path change) caused by the physical inclination of crop canopy;Based on the current UAV attitude angle, the system instantly queries or calculates the corresponding spectral correction value from the model and uses it to accurately correct the current spectral feature centroid. This effectively eliminates structural artifacts introduced by observing the lodging attitude, enabling the system to penetrate appearances and accurately track the true physiological health status of crops even under extreme conditions of drastic physical changes.

[0034] Finally, the method of this invention also integrates a high-level analysis module for early warning of physiological stress. This module achieves prospective diagnosis by analyzing the temporal micro-oscillation energy of the spectral feature centroid: First, the system caches high-density, fine-grained positional data of the spectral feature centroid over continuous time periods at a sampling frequency far higher than conventional monitoring, forming a centroid temporal signal that reflects its subtle dynamics. Subsequently, the system processes this temporal signal to calculate the total energy index of its micro-oscillations. Specifically, this index is calculated by measuring the variance of the centroid temporal signal. To quantify it, its calculation method strictly follows the formula. In progress, among which Indicates the first The centroids of spectral features collected at each time point This represents the total number of data points for buffered timing signals, while This is the average value of the buffered timing signal; ultimately, the system will calculate the total energy index of the micro-oscillations in real time. Continuous comparisons are made against a pre-defined physiological resting baseline, representing the normal energy level of healthy crops in a resting state, characterized by small, regularly fluctuating spectral centroids. Once the indicators are monitored... If the stress level remains consistently and significantly above the baseline, even if no stress symptoms are yet apparent in the macroscopic morphology, the system will immediately output early stress warning information, as it indicates that the physiological regulatory functions within the crop may have become unstable or disordered.

[0035] Before entering the online operation phase, all the core parameters upon which this method depends are deterministically generated through an offline calibration procedure. This procedure first targets the spectral feature centroids. Calculation formula Weighting coefficients in Calibration is performed by applying specific stresses to the target crop under controlled conditions and simultaneously collecting spectral and physiological indicators, identifying the characteristic wavelengths that provide the most sensitive response. Initial weights are assigned, and the remaining weights are equally distributed. Then, using maximizing the correlation coefficient as the objective function, the weights are iteratively adjusted until convergence, thus obtaining the weight combination most correlated with a specific stress. Secondly, physiological change thresholds are determined. The determination of the drift threshold is achieved by applying a slow physiological stress to healthy crop samples, collecting their time series of centroid positions, calculating the statistical distribution of the drift rates, and taking the upper limit of the distribution as the threshold. The threshold of the maneuvering flight angle for the crop lodging scenario is obtained by taking a certain multiple of the 95% quantile value of the historical data of the rate of change of the attitude angle of the UAV in standard flight, and the lodging attitude-spectral morphological response model is obtained by simulating the series of inclination angles of the crops from vertical to horizontal on a mechanical platform , collecting spectral data at each angle to calculate the centroid position, and finally fitting the angle-centroid position data pair into a low-order polynomial function; and the physiological resting baseline for early warning is directly calculated from the variance of the time series signal of the centroid of healthy non-stressed crops continuously monitored for 24 hours in a stable environment; during online operation, when the spectral feature centroid drift exceeds the threshold and the UAV attitude is stable, the system starts the light mutation correction process, which generates a correction factor based on the spectral data of the two frames before and after the light mutation, by calculating the overall band average reflectance of the stable spectral curve before the mutation and the overall band average reflectance of the spectral curve after the mutation , and taking their ratio , the factor is applied as a multiplicative scalar to all subsequent spectral data until the drift is within the threshold; if the spectral drift exceeds the threshold and the rate of change of the attitude angle simultaneously exceeds the threshold of the maneuvering flight angle, it is determined that the observed crop is lodging, and the system immediately calls the response model calibrated offline , taking the current attitude angle provided by the onboard inertial measurement unit as input, calculating the spectral centroid drift artifact value caused by physical inclination , and subtracting the artifact value from the currently measured centroid position to obtain the corrected centroid that excludes physical morphological interference

[0036] In another implementation, the weight coefficients of the spectral feature centroids are not preset fixed values, but are deterministically generated through an offline calibration procedure. This procedure first applies a pre-trained semantic segmentation model to the raw data cube acquired by the airborne hyperspectral camera. This model is a deep learning network whose function is to classify each pixel into a predefined category, thereby accurately dividing the image into the target crop canopy region and background regions such as soil and shadows. Subsequently, using only the pure pixel spectral data extracted from the target crop canopy region, combined with true physiological indicators of the crop ground collected synchronously in a controlled experiment, such as leaf water potential measured by the leaf pressure chamber, an optimization problem is constructed. The objective function is to maximize the value obtained by the formula... Calculated spectral characteristic centroid Pearson correlation coefficient between the time series and the corresponding physiological indicator time series This coefficient is a statistic ranging from -1 to 1, used to measure the degree of linear correlation between two variables, and is solved using numerical optimization algorithms under constraints. Furthermore, with all weight components being non-negative, the objective function... Weight combination that reaches the maximum value .

[0037] Furthermore, to ensure accurate judgment of the synchronization between the UAV's attitude changes and spectral drift, the system integrates a data stream alignment mechanism based on hardware timestamps. The pulse-per-second (PPS) signal output by the UAV's GPS module, a high-precision time reference provided by the GPS satellite system, is used as a unified clock source to simultaneously trigger data sampling by the inertial measurement unit (IMU) and image exposure by the hyperspectral imager, adding nanosecond-level precision synchronization timestamps to the data packets of both. During online processing, the system uses a fixed-size time window, such as 50 milliseconds, to match and associate the IMU data points with the closest timestamps with the spectral feature centroid data points, thus forming synchronization event pairs. The maneuvering flight angle threshold used to determine the collapse scenario is a variable dynamically adjusted based on real-time wind speed. Its data comes from an offline-built lookup table, which is established by analyzing historical attitude data of the UAV flying along a standard flight path at different wind speed levels. This table maps the real-time wind speed input to a specific angle change rate threshold output, which is set to the 99th percentile of the statistical distribution of historical attitude angle change rates under the corresponding wind speed conditions. These are all extended implementation methods known to those skilled in the art.

[0038] Embodiment 1: In a large-scale, automated irrigation corn plantation base with high sensitivity to weather and crop growth, the technical solution of the present application is deployed to support precision fertilization and disaster assessment based on growth, the climate characteristics of this area are that strong convective weather caused by local short-time strong wind and rapid change of light in the afternoon of summer, which poses a severe test to any continuous remote sensing monitoring; when the unmanned aerial vehicle equipped with the measurement method of the present application performs routine monitoring tasks over the 10,000 mu of corn land in the base, the inherent technical architecture advantage is revealed through the response to a series of continuous events. At the initial stage of flight, a rapidly moving cumulus cloud causes a sharp fluctuation in ground light intensity in a short period of time. Traditional methods that rely on external reference or fixed time interval calibration will cause serious distortion in reflectance calculation due to the inability to respond synchronously, or sacrifice the continuity of monitoring operations due to frequent interruptions of the flight path for physical calibration. This is the long-standing internal contradiction between data accuracy and monitoring continuity in the field of dynamic remote sensing. The present application provides a fundamental solution by comparing the drift rate of the spectral feature centroid with the physiological change threshold. At this moment, the system determines that the centroid drift rate far exceeds the upper limit of the normal physiological rhythm of corn, and determines it as an external light mutation, and instantaneously generates a correction factor to compensate for the subsequent data stream. The entire process is automatically closed in flight and does not affect the continuity of monitoring, thereby avoiding the aforementioned internal contradiction in a single architecture. The essence of the operation of this mechanism is to redefine the calibration problem from how to accurately measure and compensate for an external physical variable to whether the current spectral change rate conforms to biological laws. It no longer seeks external dependence on complex environmental models, but turns to internal examination of the endogenous biological rhythm of the system, thereby making the calibration process simple, self-consistent and robust.

[0039] Subsequently, a region in front of the flight route of the unmanned aerial vehicle encountered a downburst caused by the aforementioned strong convective weather, resulting in lodging of part of the corn, and when the unmanned aerial vehicle flew over the region, the data of the inertial measurement unit carried by the unmanned aerial vehicle and the spectral data simultaneously appeared a sharp instantaneous jump, at this key node, the synergistic effect between different technical features in the scheme is triggered, on the one hand, the sharp drift of the spectral feature centroid exceeds the physiological change threshold, triggering abnormal diagnosis, but on the other hand, the instantaneous change rate of the flight attitude angle is strictly synchronized with the spectral drift in time, so that the system determines the event to be attributed to the lodging observation scene according to the preset rule, rather than catastrophic biological stress, the direct result of this attribution is to prevent the huge spectral artifact caused by physical lodging from being incorrectly included in the time sequence micro-oscillation energy analysis model for early stress warning, thereby ensuring the long-term stability and reliability of the diagnosis baseline of the model, at the same time, the system calls the lodging attitude-spectral form response model to start correcting the spectrum of the lodging area, so that post-disaster physiological state assessment of the lodging corn becomes possible, therefore, the accurate physical identification of the lodging scene and the continuous monitoring of the physiological state form a logical positive feedback at this point, the former provides uncontaminated data input for the latter, and the latter gives the former practical value in post-disaster diagnosis based on this input, the system as a whole shows complete analysis capability for complex compound disaster scenes, the measurement method embodied in the entire application example is the fundamental principle of its system architecture: its reliability and accuracy are not based on the reliance on external high-precision sensors or complex physical models, but are rooted in the deep use of the internal life rhythm of the monitoring object, the method switches the measurement reference from the external uncertain physical environment to the internal relatively stable biological time sequence logic, thereby constructing a measurement closed loop that can self-calibrate and recognize the real physiological state under extremely low hardware redundancy, and realizing high adaptability to the real agricultural ecosystem.

[0040] Example 2: To quantitatively verify the effectiveness of the proposed online calibration method in distinguishing crop physiological evolution from sudden light change, this example was designed and implemented. The purpose of the test was to prove that by analyzing the time series drift rate of spectral feature centroid, the method could accurately attribute and isolate the apparent spectral distortion caused by sudden environmental light change without relying on any external reference, while maintaining accurate tracking of the true physiological state of the crop. To this end, we built a highly controllable indoor simulation test platform consisting of a quadcopter unmanned aerial vehicle equipped with a Cubert S185 hyperspectral imager, a programmable full-spectrum LED array lighting system covering an area of 10 square meters, and an observation area containing 20 pots of rice samples in the jointing stage with uniform growth. The unmanned aerial vehicle was fixed on a gantry that could move accurately along a pre-set trajectory in three-dimensional space, running at a speed of 1 meter per second at a height of 2 meters from the ground. This avoided interference from changes in flight attitude on spectral acquisition, ensuring the purity of test variables. The test parameters were set in accordance with a careful engineering trade-off logic. The hyperspectral data acquisition period was set to 100 milliseconds, which was optimized to balance the instantaneous capture capability of light change events and the data processing bandwidth of the on-board processor. The setting rule was that the reciprocal of the sampling frequency, 10 Hz, must be significantly higher than the Nyquist frequency of the fastest environmental change in the observed scene to ensure that the complete form of the light change process could be reconstructed without distortion, providing high-density data support for accurate calculation of the centroid drift rate. The time window used for drift calculation of the spectral feature centroid was 1 second, corresponding to 10 consecutive sampling points. This parameter aimed to balance the smoothing ability of the sensor's inherent noise and the response delay of the target event. A too short window would amplify noise artifacts, while a too long window would blunt the real instantaneous mutation. The value was determined based on the analysis of the sensor noise power spectrum and the target light change duration, achieving an optimal signal-to-noise ratio engineering compromise. The physiological change threshold was based on prior knowledge modeling of crop physiological change rate, which was the core scale for distinguishing normal physiological evolution from external abnormal disturbances. The threshold was based on statistical analysis of the historical big data of the maximum natural drift of the key spectral features of Jimai 22 during the jointing stage due to water deficit, taking the upper limit of the 99.7% confidence interval of the normal distribution to establish it. As a non-limiting example, the value was set to an equivalent centroid drift rate of 0.8 nanometers per minute, which represented the statistical limit of the centroid drift rate caused by the crop's own physiological changes without external severe disturbances.

[0041] The test procedure is performed in a continuous 3-hour monitoring period, the first stage from 0 to 2 hours, the LED array continuously irradiates at constant power, simulating sunny and cloudless weather, at the same time, the irrigation of the wheat sample is stopped to induce a natural, slow drought stress process, the system continuously collects hyperspectral data and calculates the position of the spectral feature centroid and its drift rate in real time, the observed phenomenon is that the centroid position presents a slow but continuous directional drift towards the short-wave direction, the second stage from 2 to 3 hours, while maintaining drought stress, at the time of 2 hours and 30 minutes, the light intensity of the LED array is programmed to drop by 70% in 500 milliseconds, simulating thick cloud rapid shielding, and lasts for 30 seconds, and then restores to the original intensity in 500 milliseconds, this process aims to introduce a severe external light disturbance unrelated to physiological changes, the data records of the key time nodes are shown in the following table, Table 1: Spectral feature centroid drift rate and system decision action time node comparison table.

[0042]

[0043] The above test data directly verifies the inherent calibration mechanism of the present application, in the first two hours of the test, the observed centroid drift rate is always below the preset physiological change threshold, the system correctly attributes this slow drift to the real evolution of the crop's internal physiological state caused by water deficit, and continues to absorb the current centroid position as the new dynamic calibration reference, however, at the moment of the 150-minute light mutation, the centroid drift rate instantaneously jumps to a magnitude far exceeding the threshold, this dramatic change is accurately captured by the system's inherent comparison logic and immediately triggers the decision branch, the system therefore refuses to update the calibration reference, and generates a transient correction factor to offset the impact of the disturbance on subsequent spectral data, thereby ensuring the continuity and authenticity of the measurement results.

[0044] Embodiment 4: This embodiment combines Figs. 1 to 3 , a kind of unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing is implemented, as shown in Fig. 1As shown, at 8:00 AM, under normal light conditions, the crops (taking rice as an example) were in a stable physiological state. The hyperspectral imager began collecting the first set of spectral data. Based on this, the processing system calculated the current spectral characteristic centroid C0 of the crop and set it as the initial calibration benchmark. As time progressed for 30 minutes, the moisture content of the crop leaves decreased slightly by about 2%, and the chlorophyll concentration increased slightly, indicating that the crop was in a normal transpiration and photosynthesis process. At time T1, the system collected the second set of spectral data and calculated the centroid C1. The centroid drift rate at this time was 0.45 nm / min, which was lower than the preset physiological change threshold of 0.8 nm / min, and was determined to be caused by physiological changes in the crop itself. The system updates the calibration benchmark to C1 and records the current centroid change trajectory, initially identifying the signal trend of water stress. After another 30 minutes, the crop leaf water content continues to decrease, and the canopy structure undergoes slight adjustments. At time T2, the system collects the third set of spectral data, and the centroid is updated to C2, with a corresponding drift rate of 0.68 nm / min, still within the physiological threshold range. The processing system continues to update the calibration benchmark and performs trend analysis on the data from multiple time points before and after. After receiving continuous centroid change data, the diagnostic module, combined with changes in microscopic drift direction and energy indicators, identifies that the crop is under continuous water stress, and finally outputs a diagnostic conclusion, recommending the initiation of irrigation intervention measures.

[0045] like Fig. 2 As shown in the figure, three main growth stages are marked: the technical stage, the heading stage, and the maturity stage, and represented by different line types. The horizontal axis represents wavelength (nm), and the vertical axis represents reflectance (%). The reflectance change during the technical stage shows a relatively gradual trend, especially when the wavelength is greater than 600nm, the reflectance begins to gradually increase. During the heading stage, the spectral reflectance shows a significant increase, especially near the green light reflectance peak, indicating a characteristic band of crop physiological changes. In the maturity stage, the increase in reflectance is further amplified, especially near the red edge inflection point, and the reflectance of the near-infrared plateau also shows a continuous increasing trend. The positions of the green light reflectance peak, the red edge inflection point, and the near-infrared plateau are also marked in the figure to show the spectral characteristics of the crop at each stage. These bands are of great significance for monitoring the physiological state and health assessment of crops and can effectively support the application of hyperspectral remote sensing technology in agricultural remote sensing.

[0046] like Fig. 3As shown, first, in the starting stage, the system acquires continuous hyperspectral data of the crop area through step a, and the acquired continuous spectral data are taken as the basis for subsequent processing, next, step b is for extracting spectral features and calibrating feature centroids, by extracting characteristic wavelengths such as red edge inflection point, green light reflection peak, near-infrared platform, etc., the spectral feature centroid of the crop is calibrated, in step c, the drift amount of the spectral feature centroid in unit time is calculated, the dynamic change of the centroid is calculated, and the stability of the spectral data is monitored in real time, the process enters the decision-making part: step d judges whether the drift amount exceeds the physiological threshold, if it does not exceed the physiological change value (determined as no), it is considered that the drift amount is caused by the normal physiological change of the crop, and step e is performed to judge as physiological change, the latest calibration reference is updated, the change data of the spectral centroid is recorded, and the new health status parameter is updated, if the drift amount exceeds the physiological threshold (determined as yes), step f is entered to judge as external disturbance caused, such as light change or posture change, the system generates a correction factor to correct the data, and continues to perform subsequent analysis operations, finally, in the early stress diagnosis module, steps g1 to g3 are combined, the micro-oscillation energy is calculated through high-frequency sampling to identify possible crop stress, through dynamic monitoring and analysis of the spectral data, the whole process realizes efficient, real-time crop physiological state monitoring and early warning.

[0047] In this embodiment, the off-line calibration of the whole measurement system and the whole cycle of on-line operation are integrated. Before deploying the unmanned aerial vehicle for field operation, a rigorous off-line modeling and parameter calibration stage is completed. The output of this stage provides a quantitative basis for subsequent on-line attribution and calibration. The calibration process takes rice as the target crop and is carried out at the jointing stage. First, for the physiological change threshold as the core judgment basis, its calibration follows a standardized statistical analysis process. A healthy sample group of the same variety as the target crop and at the same growth stage is selected. The sample group is continuously monitored by hyperspectral imaging for 72 hours without interruption in a growth chamber with controlled environmental parameters. During this period, a slow physiological change process driven by drought stress is induced by gradually reducing water supply. The processing system collects hyperspectral data at a frequency of once per minute, and calculates the spectral feature centroid according to the red edge inflection point wavelength, green light reflection peak wavelength and near-infrared platform starting point wavelength defined in the present application. The centroid here is precisely defined as a weighted fused equivalent wavelength, and its calculation follows the formula wherein , respectively represent the real-time positions of the above three characteristic wavelengths, and the weight coefficient satisfies ​constraints, and determined according to prior agronomic knowledge to reflect the sensitivity of each feature to the change of integrated physiological state, for drought stress, the near-infrared plateau starting point wavelength which is most sensitive to the change of canopy water content is given the highest weight; by calculating the centroid drift amount of all continuous one-minute time intervals within the entire monitoring period After calculation and statistics, a drift amount distribution containing thousands of sample points is obtained, which is approximately a normal distribution centered at zero, and the standard deviation thereof represents the natural fluctuation amplitude of spectral centroid caused by pure physiological changes of crops.

[0048] Secondly, for the recognition and calibration of the special physical scene of crop lodging, the construction of the lodging posture-spectral morphology response model relied thereon is also completed in the offline stage, using a mechanical platform that can accurately control the tilt angle, under the same controlled light environment as described above, a single healthy crop sample is simulated from zero degrees of verticality to ninety degrees of complete lying down at five-degree steps, at each tilt angle, the unmanned aerial vehicle hyperspectral imager is stably hovering from directly above to record the corresponding spectral feature centroid position, by fitting the centroid position data at all angles, a low-order polynomial function accurately mapping the deterministic relationship between the physical tilt angle of the crop canopy and the spectral feature centroid drift can be established, this model is stored in the on-board processing system, its role is to calculate the spectral artifact introduced by physical morphology changes and remove it in reverse once lodging is determined online, at the same time, the physiological resting baseline for early stress warning is established by analyzing the spectral feature centroid time series signal of a completely healthy, non-stressed crop sample continuously monitored for twenty-four hours under constant environment, using the variance calculation formula , the variance value of the healthy sample centroid time series signal is calculated, which is defined as the micro-vibration energy baseline representing physiological homeostasis.

[0049] When the unmanned aerial vehicle carries the threshold, model and baseline established by the above offline calibration into the actual field operation scene, the entire online calibration and diagnosis process can be efficiently and deterministically operated, the unmanned aerial vehicle flies along the route, continuously acquires hyperspectral data cubes, the on-board processing system calculates the position of the spectral feature centroid and its drift amount per unit time , the system compares the absolute value of with the calibrated physiological change threshold in real time, if , the system determines that the change is caused by the intrinsic physiological activity of the crop, at this time, the calibration reference is updated to the current centroid position to ensure that the measurement system can flexibly follow the natural growth rhythm of the crop, at the same time, the system caches the centroid position at high frequency, and calculates the time series micro-vibration energy And compared with the physiological resting baseline, once the energy value is continuously higher than the baseline, the early stress warning is triggered; otherwise, if If the system determines that the current drift is caused by external interference, it immediately queries the rate of change of the attitude angle of the onboard inertial measurement unit. If the rate of change of the attitude angle also appears a sharp fluctuation exceeding the threshold value, the system determines that the event is caused by the observed lodging crop, and immediately calls the pre-set lodging attitude-spectrum morphological response model to correct the spectral feature centroid according to the current attitude angle. Thus, under the interference of physical morphological changes, the real physiological state of the crop can still be calculated. If the spectral drift exceeds the threshold value while the attitude is stable, it is determined to be a sudden change in light. At this time, the calibration reference remains unchanged, and a temporary correction factor is generated according to the size and direction of the drift to compensate for the subsequent data stream in the reverse direction until the light returns to stability. This off-line calibration and on-line decision-making process together constitute a measurement closed loop that is not dependent on external reference and can adapt to complex farmland environments.

[0050] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0051] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application.

Claims

1. A method for unmanned aerial vehicle hyperspectral online calibration measurement for agricultural remote sensing, characterized in that, The method comprises the following steps: Step a, acquiring continuous hyperspectral data of crop area; Step b, extracting a plurality of spectral feature wavelengths determined according to crop types and growth stages from the continuous hyperspectral data in real time, and constructing a spectral feature centroid representing the current hyperspectral morphology based on the plurality of spectral feature wavelengths; Step c, calculating the drift amount of the spectral feature centroid in unit time in real time; Step d, comparing the drift amount of the spectral feature centroid with a physiological change threshold value determined based on statistical analysis and modeling of historical data, and performing spectral centroid drift attribution determination according to the comparison result; Step e, when the drift amount of the spectral feature centroid does not exceed the physiological change threshold value, it is judged that the spectral feature centroid drift is caused by the change of the physiological state of the crop, and the current spectral feature centroid is updated as a new calibration reference; Step f, when the drift amount of the spectral feature centroid exceeds the physiological change threshold value, it is judged that the spectral feature centroid drift is caused by external light or change of the attitude of the unmanned aerial vehicle, at this time the calibration reference is not updated, and a correction factor is generated to calibrate the continuously acquired hyperspectral data in real time; And when the drift amount of the spectral feature centroid exceeds the physiological change threshold value, the method further comprises the following sub-steps: step f1, acquiring current flight attitude data from the inertial measurement unit on board the unmanned aerial vehicle; step f2, performing crop lodging observation scene determination based on the flight attitude data, wherein the crop lodging observation scene determination is performed by comparing the instantaneous change rate of the flight attitude angle with a predetermined maneuvering flight angle threshold value, if the change rate exceeds the maneuvering flight angle threshold value, and is synchronized in time with the drift of the spectral feature centroid, it is determined that it is in the crop lodging observation scene; step f3, calling a pre-established lodging attitude-spectral morphology response model, and correcting the current spectral feature centroid according to the model. 2.The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing according to claim 1, characterized in that, The plurality of spectral feature wavelengths for constructing the spectral feature centroid in step b include a red edge inflection point wavelength, a green light reflection peak wavelength and a near-infrared platform starting point wavelength. 3.The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing according to claim 1, characterized in that, In step d, the physiological change threshold value is determined based on statistical analysis and modeling of historical spectral data of the target crop type at different growth stages. 4.The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing of claim 1, wherein, In step e, when it is judged that the spectral feature centroid drift is caused by the change of the physiological state of the crop, the method further comprises: taking the amplitude and direction of the centroid drift as an indication of the degree and type of physiological stress of the crop.

5. The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing according to claim 1, characterized in that, The lodging attitude-spectral morphology response model is a calibrated and stored lookup table or low-order polynomial function, which is established by data collection and model training on the spectral response of simulated lodging crops at different inclination angles. 6.The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing of claim 1, wherein, The method further comprises a time series micro-vibration energy analysis of the spectral feature centroid for early physiological stress warning of the crop, and the analysis comprises the following steps: g1, buffering position data of the spectral feature centroid in a continuous time period at a high sampling frequency to form a centroid time series signal; g2, processing the centroid time series signal to calculate a total energy index of micro-vibration, which is obtained by calculating a variance of the centroid time series signal; g3, comparing the total energy index of micro-vibration with a physiological resting baseline in real time, the physiological resting baseline representing a normal energy level of micro-vibration of the crop, and outputting early stress warning information when the index continuously exceeds the physiological resting baseline.

7. The unmanned aerial vehicle hyperspectral online calibration measurement method for agricultural remote sensing according to claim 6, characterized in that, The variance of the centroid time series signal in step g2 is calculated as: wherein denotes the spectral feature centroid acquired at the time point, denotes the total number of data points of the buffered centroid time series signal, denotes the mean value of the buffered centroid time series signal.

Images