Sugarbeet nutrition detection system based on multi-spectral data of unmanned aerial vehicle

Abstract

The invention discloses a beet nutrition detection system based on unmanned aerial vehicle multispectral data, and relates to the technical field of agricultural remote sensing and precision agriculture, and the beet nutrition detection system comprises an unmanned aerial vehicle multispectral acquisition module used for obtaining multispectral reflectivity data of beet canopies; according to the beet nutrition detection system based on the unmanned aerial vehicle multi-spectral data, the precision and the real-time performance of beet nutrition detection are remarkably improved through dynamic spectrum-physiological coupling modeling and multi-dimensional feature decoupling technologies; based on a differentiable PROSAIL radiation transport sub-network, the system fuses a physical mechanism and deep learning, and dynamically corrects spectrum deviation caused by canopy structure change, so that the root-mean-square error of nitrogen content detection is reduced, and the precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural remote sensing and precision agriculture, and in particular to a sugar beet nutrition detection system based on multispectral data of unmanned aerial vehicle (UAV). Background Art

[0002] In precision agricultural management, rapid, nondestructive monitoring of sugar beet nutritional status is crucial for improving yield and quality. In recent years, drone-mounted multispectral remote sensing technology has become an increasingly important tool for sugar beet nutritional monitoring due to its high efficiency and wide-area monitoring capabilities. Existing technologies primarily use drone-mounted multispectral sensors to acquire crop canopy reflectance data. These sensors, combined with traditional vegetation indices or statistical models, establish linear mappings between spectral characteristics and sugar beet nutritional indicators such as nitrogen, phosphorus, and potassium. However, a core flaw of these methods is their reliance on static, empirical spectral response models, which fail to effectively address the complex dynamic coupling between physiological state and spectral characteristics during sugar beet growth. Sugar beet canopy structure, leaf biochemical composition, and environmental background exhibit significant dynamic changes from seedling to maturity. For example, low canopy cover in the early stages results in significant soil reflectance interference, while increased leaf thickness in the mid-to-late stages leads to shifts in spectral absorption characteristics. Furthermore, physiological responses such as chlorophyll degradation triggered by nutrient stress exhibit a time lag between the spectral signals observed by drones.

[0003] Existing models lack the ability to dynamically decouple growth stages, environmental noise, and physiological delay effects, resulting in a significant decrease in detection accuracy as the growth period progresses. In particular, it is difficult to capture weak spectral changes in the early stages of nutrient deficiency, resulting in delayed diagnosis. In addition, traditional methods usually train models based on data from specific regions or varieties. They lack generalization capabilities when faced with different soil types, climatic conditions, or beet genotypes, limiting the large-scale application of the technology. The essence of this problem is the adaptability contradiction between static spectral models and dynamic agricultural scenarios. There is an urgent need to build a multi-dimensional dynamic coupling analysis system that can adapt to beet growth characteristics and environmental interference. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides a sugar beet nutrient detection system based on drone multispectral data, which solves the problem of how to achieve high-precision coupling modeling of sugar beet multispectral data and dynamic physiological state to overcome the influence of growth stage, environmental interference and physiological hysteresis on the accuracy of nutrient detection.

[0006] (2) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A sugar beet nutrition detection system based on drone multispectral data, comprising:

[0008] The drone multispectral acquisition module is used to obtain multispectral reflectance data of the sugar beet canopy. The drone is equipped with a multispectral sensor and flies along a preset route to obtain multispectral reflectance data of the sugar beet canopy from visible light to infrared wavelengths.

[0009] A dynamic spectral-physiological coupling analysis module is communicatively connected to the UAV multispectral acquisition module and includes a differentiable radiation transmission subnetwork, a time-space-spectral joint dynamic decoupling unit, and a physiological lag compensation unit;

[0010] The differentiable radiation transfer subnetwork is constructed based on the PROSAIL model, which converts the biochemical parameters of sugar beet leaves, the geometric structure of the canopy, and the optical properties of the soil background into physical constraint equations and embeds them into a deep neural network. The time-space-spectrum joint dynamic decoupling unit dynamically screens features through time window division, asymmetric convolution kernels, and spectral attention weights. The physiological lag compensation unit generates a real-time-prediction joint feature vector through a ground sensor network and an LSTM prediction module.

[0011] A verification feedback module, connected to the dynamic spectral-physiological coupling analysis module, for dual-path inversion verification and model iteration;

[0012] The edge computing gateway is deployed between the drone and the ground base station, using the MQTT protocol to compress and transmit data and synchronize and calibrate timestamps.

[0013] Preferably, the differentiable radiation transfer subnetwork includes:

[0014] The physical constraint equation generation layer dynamically adjusts the leaf optical parameters and canopy structure parameters according to the sugar beet growth period;

[0015] The virtual leaf optical fingerprint generation layer corrects the mechanism model error through back propagation and outputs the spectral characteristics that match the current growth period;

[0016] The soil noise suppression layer constructs an orthogonal projection matrix to separate soil reflectance and canopy spectrum when the canopy coverage is lower than a preset threshold.

[0017] Among them, the construction and parameter adjustment of the physical constraint equation generation layer:

[0018] PROSAIL model parameterization: The biochemical parameters of sugar beet leaves and canopy structure parameters are used as input variables. The biochemical parameters of sugar beet leaves include leaf chlorophyll content C ab , blade equivalent water thickness C w The canopy structure parameters include the canopy leaf area index LAI, the canopy average leaf inclination angle θ l , the theoretical spectral reflectance of the sugar beet canopy at wavelength λ is recorded as R(λ), and the radiation transfer equation is constructed:

[0019] R(λ)=f(Cab , C w , LAI, θ l , ρ soil (λ)), where R(λ) is the theoretical spectral reflectance of the sugar beet canopy at wavelength λ; C ab : Leaf chlorophyll content; C w : leaf equivalent water thickness; LAI: canopy leaf area index; θ l : average leaf inclination angle of the canopy; ρ soil (λ): reflectance of soil background at wavelength λ;

[0020] Back propagation correction: the drone measured spectrum R obs (λ) and model output R model The mean square error (MSE) of (λ) is used as the loss function, and the parameters such as Cab and LAI are updated by the chain rule:

[0021] in, is the mean square error, which is used to measure the UAV measured spectrum R obs (λ) and model output R model The difference between (λ), R obs (λ) is the spectral reflectance measured by the UAV, R model (λ) is the spectral reflectance output by the model, C ab is the chlorophyll content of the beet leaf, which is the parameter for partial derivative; λ represents the wavelength of the spectrum, which is used to describe the spectral reflectance characteristics of the beet canopy or soil background in different bands;

[0022] Soil noise suppression: When the canopy coverage is less than 40%, principal component analysis (PCA) is used to extract the first three principal components of the soil spectrum (V1, V2, and V3), and the orthogonal projection matrix P is constructed. The formula is P = IV·V T , where V T It is the matrix obtained by swapping the rows and columns of V. It is the transpose of the principal component matrix and is used to construct the orthogonal projection matrix. I is the identity matrix, and V is the matrix composed of the first three principal components of the soil spectrum extracted by principal component analysis (PCA). T represents the transpose operation of the matrix and is used to construct the orthogonal projection matrix to separate noise.

[0023] Filter the original spectrum R: R filtered =P·R, where R filtered is the filtered spectrum, P is the orthogonal projection matrix used to separate the soil reflection noise, and R is the original spectrum.

[0024] Preferably, the time-space-spectrum joint dynamic decoupling unit includes:

[0025] The time dimension decoupling module extracts the spectral response patterns of different growth stages based on the gated recurrent unit and outputs the identification signal of the sensitive period of nutritional stress;

[0026] The spatial dimension decoupling module uses an asymmetric convolution kernel to extract canopy texture. The size of the convolution kernel is dynamically adjusted according to the flight altitude of the UAV, and the convolution kernel size is increased every time the flight altitude increases by a set distance.

[0027] The spectral dimension decoupling module dynamically allocates the contribution of the red edge band and the infrared band through attention weights.

[0028] The GRU network training of the time dimension decoupling module includes:

[0029] Input data: Spectral sequence of time windows divided by growth period, i.e., each window is 10 days, the step length is 5 days, and it is normalized to the interval [0,1].

[0030] GRU structure: hidden layer neurons 64, Dropout rate 0.2, output layer is the nutrient stress probability P stress A dropout rate of 0.2 means that during the training of the GRU network, each neuron has a 20% probability of being randomly dropped in each training step. In other words, in each round of training, approximately 20% of the neurons will be temporarily removed from the network and will not participate in the current calculation.

[0031] Warning trigger: When P stress >0.8 and the red edge band slope S RE =(R 750 -R 700 ) / 50nm drops by more than 15%, the warning signal is activated; among them, S RE is the red edge band slope, R 750 and R 700 are the spectral reflectances at wavelengths of 750nm and 700nm respectively.

[0032] The asymmetric convolution kernel operation process of the spatial dimension decoupling module is as follows:

[0033] Dynamic adjustment of convolution kernel: The mapping rule between flight height H (unit: meter) and convolution kernel size is:

[0034] Long axis size = 3 × 1.5 [H / 10] , minor axis size = 1×1.2 [H / 10] , where H is the flight altitude of the UAV;

[0035] Texture feature extraction: The canopy RGB image is grayscaled, the local binary pattern (LBP) feature is calculated, and the feature is concatenated with the spectral feature and input into the fully connected layer.

[0036] The attention weight calculation of the spectral dimension decoupling module includes:

[0037] Band importance score: Generate the weight w of each band through the fully connected layer i :w i =Softmax(W·R(λi)+b); where w i is the weight of each band, W is the weight matrix, the initial weight of the red edge band (720-750nm) in the seedling stage is set to 0.6, and the initial weight of the infrared band (1550-1750nm) in the mature stage is set to 0.7. i ) is the wavelength λ i The spectral reflectance at the red edge band is b; b is the bias term. The initial weight of W in the seedling stage is set to 0.6 in the red edge band of 720-750nm, and it is set to 0.7 in the infrared band of 1550-1750nm in the mature stage. i is the index number of the spectral band, which is used to traverse the band data collected by all sensors. Its core function is to use the dynamic weight w i Screening key spectral information to support high-precision detection of sugar beet nutritional status;

[0038] Weighted spectrum reconstruction: output feature F = ∑ i w i R(λ i ), where F is the output feature after weighted spectrum reconstruction, w i is the weight of each band, R(λ i ) is the wavelength λ i The spectral reflectance at the location; i is the index number of the spectral band, which is used to traverse the band data collected by all sensors. Its core function is to use the dynamic weight w i Filter key spectral information to support high-precision detection of the nutritional status of sugar beets.

[0039] Preferably, the physiological hysteresis compensation unit comprises:

[0040] A ground sensor network is deployed in the monitoring area to collect real-time temperature and humidity of the root layer and stomatal conductance of leaves;

[0041] A spatiotemporal alignment module maps the timestamps and spatial coordinates of drone observation data and ground sensor data to the same benchmark;

[0042] The LSTM prediction module inputs historical ground data and current spectral characteristics, and outputs the lagged physiological state prediction value within the future set period, that is, the chlorophyll content prediction value for the next 24 hours

[0043] The hardware configuration of the ground sensor network is as follows: the temperature and humidity sensor uses the SHT35 chip, the humidity measurement accuracy is within plus or minus 2% of the relative humidity, the temperature measurement accuracy is within plus or minus 0.3°C of the relative temperature, and the sampling frequency is 5 minutes / time; the stomatal conductance sensor is based on leaf clamp impedance measurement, with a range of 0-10mmol / (m 2 ·s), resolution 0.1mmol / (m 2 ·s).

[0044] The implementation details of the LSTM prediction module include:

[0045] Input sequence: 48 hours of historical ground sensor data, including temperature, humidity, stomatal conductance, and the red edge and infrared reflectance of the drone's spectral characteristics, with a time step of 1 hour.

[0046] The network structure is a 2-layer LSTM with 128 neurons in each layer, which outputs the predicted chlorophyll content for the next 24 hours. and nitrogen concentration N pred ;

[0047] The chlorophyll content predicted by LSTM for the next 24 hours Chlorophyll content detected by drone in real time Fusion by weight is used to obtain the final value of chlorophyll content after fusion The formula is: in, is the final value of chlorophyll content after fusion; It is the predicted value of chlorophyll content in the next 24 hours predicted by the LSTM prediction submodule; It is the chlorophyll content detection value detected by the drone in real time.

[0048] Among them, when data is missing on cloudy days, the weight is adjusted to 0.9:0.1, and 0.9:0.1 respectively represent the weight ratio of the chlorophyll content prediction value in the next 24 hours and the chlorophyll content detection value detected by the drone in real time. That is, the system adjusts the fusion ratio of the chlorophyll content prediction value in the next 24 hours predicted by LSTM and the chlorophyll content detection value detected by the drone in real time, that is, the final chlorophyll content value after fusion is The predicted chlorophyll content in the next 24 hours And the chlorophyll content detection value detected by drone in real time The fusion is done as follows:

[0049] in, is the final value of chlorophyll content after fusion; The chlorophyll content in the next 24 hours is predicted by the LSTM prediction submodule; It is the chlorophyll content detection value detected by the drone in real time.

[0050] Preferably, in the spatial dimension decoupling module, the size adjustment rule of the asymmetric convolution kernel is: when the flight altitude of the drone increases by 10 meters, the long axis size of the convolution kernel is expanded by 1.5 times, and the short axis size is expanded by 1.2 times.

[0051] Preferably, the output of the LSTM prediction module includes the predicted value of chlorophyll content and nitrogen concentration in the next 24 hours, which are superimposed with the real-time spectral characteristics of the drone and then input into the dynamic spectral-physiological coupling analysis module.

[0052] Preferably, the verification feedback module includes:

[0053] The mechanism pathway verification submodule infers the nutritional indicators to the differentiable radiation transfer subnetwork to generate a theoretical spectrum and calculates the residual with the actual observed spectrum;

[0054] The biochemical pathway verification submodule performs deviation analysis between the measured values ​​of leaf near-infrared spectra in random sampling areas and the predicted values ​​of chlorophyll content in the next 24 hours;

[0055] The model iteration trigger unit starts parameter fine-tuning when the residual or deviation exceeds the limit continuously.

[0056] Among them, the residual calculation process of the mechanism path verification submodule is as follows:

[0057] Theoretical spectrum generation: The inverted C ab , LAI input PROSAIL model to generate the theoretical spectral reflectance R of 400-2500nm theory (λ);

[0058] Residual evaluation: Calculate the residual sum of squares in the red edge band 700-750nm:

[0059] Among them, RSS is the residual sum of squares of the red edge band (700-750nm), R theory (λ) is the inverted C ab , LAI input PROSAIL model to generate theoretical spectral reflectance, R obs (λ) is the spectral reflectance measured by the UAV, where λ represents the wavelength of the spectrum and is used to describe the spectral reflectance characteristics of the sugar beet canopy or soil background in different bands;

[0060] If RSS is >5% at the seedling stage or >3% at the mature stage, the online update of PROSAIL parameters is triggered.

[0061] The sampling and verification process of the biochemical pathway verification submodule is as follows:

[0062] Leaf sampling specifications: Randomly select 5% of the monitoring area, sample the top third fully expanded leaf in each area, and use ASD FieldSpec 4 near-infrared spectrometer to measure chlorophyll content;

[0063] Deviation calculation: When the chlorophyll content forecast value for the next 24 hours is repeated for three times Chlorophyll content detected by drone in real time satisfy , freeze the fully connected layers and reinitialize the classifier weights.

[0064] Preferably, the residual threshold of the mechanism path verification submodule is dynamically set according to the growth period of sugar beet: the threshold in the seedling stage is 5%, and the threshold in the mature stage is 3%.

[0065] Preferably, it also includes a cross-region generalization adaptation module, which performs domain adaptive optimization on the pre-training model through a transfer learning algorithm and uses the maximum mean difference loss function to align the data distribution of different soil types and genotypes.

[0066] The training process of the cross-region generalization adaptation module is as follows:

[0067] Pre-training stage: The initial model is trained using 1000 sets of samples in the black soil area (source domain), with a learning rate of 1e-4 and a batch size of 32;

[0068] Domain adaptation stage: 50 groups of samples were collected in the sandy loam area (target domain), the PROSAIL mechanism layer was frozen, and only the fully connected layer was fine-tuned. The MMD loss function was calculated as:

[0069] Among them, N s and N t are the number of samples in the source domain and the target domain respectively; and is a sample in the source domain; and is a sample in the target domain; k(·) is a Gaussian kernel function with bandwidth σ = 1.0; L MMD : Distribution alignment loss for cross-region transfer learning; x s : Source domain data samples, such as black soil area data samples; x t : Target domain data samples, such as sandy loam area data samples; and are the squares of the number of samples in the source domain and the target domain, respectively, used to normalize the internal similarity between the source domain and the target domain; m and n are index variables that traverse the sample pairs, covering all possible combinations; k(,) is the kernel function used to calculate the similarity of samples in the high-dimensional feature space.

[0070] Preferably, the UAV multispectral acquisition module, dynamic spectral-physiological coupling analysis module, and verification feedback module implement data exchange through an edge computing gateway. The gateway is deployed between the UAV and the ground base station to compress and transmit data and synchronize and calibrate timestamps. The hardware deployment process of the edge computing gateway is as follows:

[0071] Drone: Equipped with the NVIDIA Jetson Xavier NX module, it runs a lightweight spectral preprocessing model and TensorRT acceleration, with a power consumption of less than 15W.

[0072] Ground base station: uses Raspberry Pi 4B to receive sensor data and uploads it to the cloud via a 4G module;

[0073] Clock synchronization: Based on the NTP protocol, the clock is calibrated every 30 seconds, and the timestamp error is controlled within ±5ms.

[0074] (3) Beneficial effects

[0075] The present invention provides a sugar beet nutrient detection system based on drone multispectral data. It has the following beneficial effects:

[0076] (1) The sugar beet nutrient detection system based on UAV multispectral data has significantly improved the accuracy and real-time performance of sugar beet nutrient detection through dynamic spectral-physiological coupling modeling and multi-dimensional feature decoupling technology; based on the differentiable PROSAIL radiation transfer subnetwork, the system integrates physical mechanisms with deep learning, dynamically corrects the spectral offset caused by canopy structure changes, reduces the root mean square error of nitrogen content detection, and improves the accuracy; the time-space-spectrum joint decoupling mechanism divides the growth period window through the GRU network, captures the red edge band slope anomaly in the early stage of nutrient stress, that is, the asymptomatic stage, and realizes early warning; the adaptive asymmetric convolution kernel dynamically adjusts its size according to the flight altitude to eliminate the leaf overlapping aliasing effect, and the accuracy of spatial feature extraction is improved; the physiological lag compensation module combines LSTM prediction with real-time data fusion to ensure the continuity of detection in severe weather and reduce errors.

[0077] (2) This sugar beet nutrient detection system based on drone multispectral data, through dual-path verification and cross-regional generalization design, balances robustness and potential for large-scale application. The mechanistic-biochemical dual-path verification mechanism triggers dynamic model iteration, resulting in minimal fluctuation in accuracy over long-term operation and avoiding the overfitting risk associated with a traditional single-path approach. Cross-regional transfer learning requires only 50 sets of samples from the target area, reducing deployment costs. The edge computing gateway utilizes low-power hardware and lossless compression to achieve efficient data transmission and spatiotemporal synchronization of multi-source data. In actual applications, the system has a low false alarm rate in growing areas, improving efficiency compared to manual inspections while reducing excessive fertilizer use. This system combines economic benefits with ecological sustainability, providing a highly reliable solution for precision agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a schematic diagram of the overall framework of the present invention;

[0079] Figure 2 This is a control logic timing diagram of the present invention. DETAILED DESCRIPTION

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0081] See also Figure 1 and Figure 2 The present invention provides a technical solution: a sugar beet nutrition detection system based on drone multispectral data, comprising:

[0082] The drone multispectral acquisition module, equipped with a multispectral sensor and flying along a preset route, acquires multispectral reflectance data of the sugar beet canopy in the 400-2500nm band;

[0083] The dynamic spectral-physiological coupling analysis module is connected to the UAV multispectral acquisition module and includes a differentiable radiation transmission subnetwork, a time-space-spectral joint dynamic decoupling unit, and a physiological lag compensation unit.

[0084] The differentiable radiative transfer subnetwork is constructed based on the PROSAIL model, with the chlorophyll content C as input. ab , equivalent water thickness C w , leaf area index LAI, average leaf inclination angle θ l and soil reflectivity ρ soil (λ), correcting the model parameter errors through back-propagation;

[0085] The time-space-spectrum joint dynamic decoupling unit achieves feature decoupling through time window division, dynamic adjustment of asymmetric convolution kernels and spectral attention weight allocation;

[0086] The physiological lag compensation unit generates a real-time-prediction joint feature vector through the ground sensor network and the LSTM prediction module;

[0087] Verification feedback module, connected to the dynamic spectral-physiological coupling analysis module, for dual-path inversion verification and model iteration;

[0088] The edge computing gateway is deployed between the drone and the ground base station, using the MQTT protocol to compress and transmit data and synchronize the clock error to ±5ms through the NTP protocol.

[0089] It should be further explained that, in the specific implementation process, the differentiable radiation transfer subnetwork includes:

[0090] Physical constraint equation generation layer, constructing the radiation transfer equation R(λ)=f(C ab , C w , LAI, θ l , ρ soil (λ)), dynamically update soil reflectance parameters; where R(λ) is the theoretical spectral reflectance of the sugar beet canopy at wavelength λ; C ab : Leaf chlorophyll content; C w : leaf equivalent water thickness; LAI: canopy leaf area index; θ l : average leaf inclination angle of the canopy; ρ soil (λ): reflectance of soil background at wavelength λ;

[0091] The virtual blade optical fingerprint generation layer is modified by back propagation of the mean square error loss function C ab , LAI parameters;

[0092] The soil noise suppression layer, when the canopy coverage is less than 40%, uses principal component analysis to extract the first three principal components of the soil spectrum and constructs the orthogonal projection matrix P = IV·V T Filtering is performed; where V T It is the matrix obtained by swapping the rows and columns of V. It is the transpose of the principal component matrix and is used to construct the orthogonal projection matrix. I: the identity matrix, V: the matrix composed of the first three principal components of the soil spectrum extracted by principal component analysis (PCA), and T represents the transpose operation of the matrix, which is used to construct the orthogonal projection matrix to separate noise.

[0093] It should be further explained that, in the specific implementation process, the time-space-spectrum joint dynamic decoupling unit includes:

[0094] The time dimension decoupling module inputs the time window spectrum sequence divided by the growth period, uses the GRU network with 64 neurons in the hidden layer, and outputs the nutritional stress probability P stress , when P stress >0.8 and the red edge slope drops by more than 15%, triggering an alarm;

[0095] The spatial dimension decoupling module uses an asymmetric convolution kernel, and its size is 3×1.5 as the flight height H increases. (H / 10) Long axis and 1×1.2 (H / 10) Dynamic adjustment of short axis;

[0096] Spectral dimension decoupling module generates band weights w through the fully connected layer i =Softmax(W·R(λi ), where w i is the weight of each band, W is the weight matrix, the initial weight of the red edge band (720-750nm) in the seedling stage is set to 0.6, and the initial weight of the infrared band (1550-1750nm) in the mature stage is set to 0.7. i ) is the wavelength λ i The spectral reflectance at the red edge band of 720-750nm is set to 0.6 in the seedling stage, and to 0.7 in the infrared band of 1550-1750nm in the mature stage; i is the index number of the spectral band, which is used to traverse the band data collected by all sensors. Its core function is to use the dynamic weight w i Filter key spectral information to support high-precision detection of the nutritional status of sugar beets.

[0097] It should be further explained that, in the specific implementation process, the ground sensor network of the physiological hysteresis compensation unit is deployed in the monitoring area. The SHT35 chip is used at every 10-meter node. The humidity measurement accuracy is within plus or minus 2% of the relative humidity, and the temperature measurement accuracy is within plus or minus 0.3℃ of the relative temperature. The soil data at a depth of 20cm in the root layer is collected. The leaf clamped stomatal conductance sensor has a range of 0-10mmol / (m 2 ·s), sampling every 5 minutes;

[0098] The LSTM prediction module inputs the historical 48-hour ground data and the drone spectral characteristics, and outputs the predicted chlorophyll content value for the next 24 hours through two layers of LSTM, with 128 neurons in each layer. And the predicted value of nitrogen concentration N pred and with real-time data press Fusion, where is the final value of chlorophyll content after fusion; The chlorophyll content in the next 24 hours is predicted by the LSTM prediction submodule; It is the chlorophyll content detection value detected by the drone in real time.

[0099] Among them, when data is missing on cloudy days, the weights are adjusted to 0.9:0.1, and 0.9:0.1 represent the predicted chlorophyll content values ​​for the next 24 hours. And the chlorophyll content detection value detected by drone in real time The weight ratio of the system is to adjust the fusion ratio of the chlorophyll content prediction value of the next 24 hours predicted by LSTM and the chlorophyll content detection value of the drone real-time detection value, that is, the final value of the chlorophyll content after fusion The predicted chlorophyll content in the next 24 hours and the predicted chlorophyll content for the next 24 hours The fusion is done as follows:

[0100] in, is the final value of chlorophyll content after fusion; It is the predicted value of chlorophyll content in the next 24 hours predicted by the LSTM prediction submodule; It is the chlorophyll content detection value detected by the drone in real time.

[0101] It should be further explained that, in the specific implementation process, the spatial dimension decoupling module grayscales the canopy RGB image, calculates the local binary pattern (LBP) texture features, and inputs them into the fully connected layer after splicing with the spectral features.

[0102] It should be further explained that in the specific implementation process, the time step of the LSTM prediction module is 1 hour, and the correlation coefficient R between the predicted chlorophyll content value in the next 24 hours and the chlorophyll content detection value detected by the drone in real time is 2 =0.87.

[0103] It should be further explained that, in the specific implementation process, the verification feedback module includes:

[0104] Mechanism path verification submodule, calculates the residual sum of squares between the theoretical spectral reflectance in the red edge band 700-750nm and the actual observation The threshold value is 5% in the seedling stage and 3% in the mature stage. RSS is the residual sum of squares of the red edge band (700-750nm), R theory (λ) is the inverted C ab , LAI input PROSAIL model to generate theoretical spectral reflectance, R obs (λ) is the spectral reflectance measured by the UAV, where λ represents the wavelength of the spectrum and is used to describe the spectral reflectance characteristics of the sugar beet canopy or soil background in different bands;

[0105] In the biochemical pathway verification submodule, chlorophyll content is measured by an ASD FieldSpec 4 near-infrared spectrometer. When the prediction deviation exceeds 10% for three consecutive times, the fully connected layer is frozen and the classifier weights are reinitialized.

[0106] It should be further explained that, in the specific implementation process, the mechanism path verification submodule generates the 400-2500nm theoretical spectral reflectance R theory (λ), and the residual is kept below the threshold by updating the PROSAIL parameters online.

[0107] It should be further explained that, in the specific implementation process, the sugar beet nutrition detection system based on UAV multispectral data also includes a cross-regional generalization adaptation module, and the maximum mean difference loss function L is used in the sandy loam area. MMDThe data distribution of the source domain and the target domain are aligned, the Gaussian kernel bandwidth σ = 1.0, the PROSAIL mechanism layer is frozen and the fully connected layer is fine-tuned. The source domain is the black soil area and the target domain is the sandy loam area.

[0108] The hardware of the edge computing gateway includes:

[0109] The NVIDIA Jetson Xavier NX module on the drone side runs a lightweight model accelerated by TensorRT with a power consumption of less than 15W;

[0110] The ground base station Raspberry Pi 4B uploads data via the 4G module;

[0111] Data compression uses JPEG2000 lossless compression with a compression ratio of no less than 5:1, and the clock error is synchronized to ±5ms through the NTP protocol.

[0112] It should be further explained that during the implementation, a drone equipped with a multispectral acquisition module, equipped with a multispectral sensor, flew over a sugar beet planting area along a pre-set route. Its flight altitude was dynamically adjusted based on canopy cover, acquiring real-time multispectral reflectance data from the sugar beet canopy in the 400-2500nm band. This dynamic adjustment of altitude based on canopy cover included adjusting the flight altitude to 20 meters for the seedling stage and 50 meters for the mature stage. The sensor collected data at a rate of 10 frames per second and transmitted it to a ground base station via the MQTT protocol via an edge computing gateway. The data packets contained geographic coordinates, a timestamp, and a reflectance matrix. The edge computing gateway used the JPEG2000 lossless compression algorithm, achieving a 5:1 compression ratio to ensure efficient transmission. The drone and ground sensor clocks were synchronized using the NTP protocol, maintaining timestamp error within ±5ms. For example, during a flight mission during the sugar beet bulking period, canopy images captured by the drone at an altitude of 30 meters were compressed, reducing the size of a single image from 50MB to 10MB, with a transmission latency of less than 200ms.

[0113] After receiving the multispectral data, the dynamic spectral-physiological coupling analysis module first performs physical mechanism modeling by the differentiable radiation transfer subnetwork. This subnetwork is built based on the PROSAIL model and inputs the biochemical parameters of sugar beet leaves, canopy structure parameters, and soil reflectance ρ. soil (λ), generating the theoretical spectral reflectance equation R(λ)=f(C ab , C w , LAI, θ l , ρ soil (λ)), where the biochemical parameters of sugar beet leaves include chlorophyll content C ab , equivalent water thickness C w , canopy structure parameters include leaf area index LAI, average leaf inclination angle θ lWhere, R(λ): theoretical spectral reflectance of sugar beet canopy at wavelength λ; C ab : Leaf chlorophyll content; C w : leaf equivalent water thickness; LAI: canopy leaf area index; θ l : average leaf inclination angle of the canopy; ρ soil (λ): reflectance of soil background at wavelength λ.

[0114] In actual operation, for the case where the canopy coverage is less than 40% during the seedling stage, the soil noise suppression layer is activated: the first three principal components V1, V2, and V3 of the soil spectrum are extracted through principal component analysis, and the orthogonal projection matrix P = IV·V is constructed. T , filter the original spectrum R to obtain R filtered＝ P·R, eliminate soil background interference; where R filtered is the filtered spectrum, P is the orthogonal projection matrix used to separate the soil reflection noise, and R is the original spectrum.

[0115] Among them, V T It is the matrix obtained by swapping the rows and columns of V. It is the transpose of the principal component matrix and is used to construct the orthogonal projection matrix. I: the identity matrix, V: the matrix composed of the first three principal components of the soil spectrum extracted by principal component analysis (PCA), and T represents the transpose operation of the matrix, which is used to construct the orthogonal projection matrix to separate noise.

[0116] The process of multi-dimensional feature decoupling of filtered spectral data by the time-space-spectrum joint dynamic decoupling unit is as follows:

[0117] Time dimension decoupling: The time window is divided according to the growth period of sugar beets, i.e., 0-30 days for seedling stage, 30-70 days for expansion stage, and 70-120 days for maturity stage. The gated recurrent unit GRU is used to process the time series data. The GRU network hidden layer has 64 neurons, inputs the normalized spectral sequence, each window is 10 days, the step length is 5 days, and the output is the nutritional stress probability P stress When P stress >0.8 and the red edge slope S RE =(R 750 -R 700 ) / 50nm drops by more than 15%, triggering an early warning signal; among them, S RE is the red edge band slope, R 750 and R 700 are the spectral reflectance at wavelengths of 750nm and 700nm respectively. RE The system automatically marked the area as a high-risk area for nitrogen deficiency.

[0118] Spatial Dimension Decoupling: An asymmetric convolution kernel with a major axis of 3×1 and a minor axis of 1×3 is used to extract inter-row texture features of the canopy. The convolution kernel size is dynamically adjusted based on the flight altitude: for every 10-meter increase in altitude, the major axis is expanded by a factor of 1.5, and the minor axis is expanded by a factor of 1.2. For example, at a flight altitude of 30 meters, the convolution kernel size is adjusted to a major axis of 6.75×1 and a minor axis of 1.44×3 to match the spatial resolution of the canopy texture. Simultaneously, the canopy RGB image is grayscaled, and local binary pattern (LBP) features are calculated. These features are then concatenated with the spectral features and fed into the fully connected layer.

[0119] Spectral dimension decoupling: Dynamically assign band weights through the self-attention mechanism. The red edge band (720-750nm) is given an initial weight of 0.6 during the seedling stage, and the infrared band (1550-1750nm) is enhanced to a weight of 0.7 during the mature stage. The weighted spectral feature F = ∑ i w i R(λ i ) is input into the subsequent analysis module, where F is the output feature after weighted spectrum reconstruction, w i is the weight of each band, R(λ i ) is the wavelength λ i The spectral reflectance at the location; i is the index number of the spectral band, which is used to traverse the band data collected by all sensors. Its core function is to use the dynamic weight w i Filter key spectral information to support high-precision detection of the nutritional status of sugar beets.

[0120] The physiological hysteresis compensation unit works in conjunction with the LSTM prediction module through the ground sensor network:

[0121] Ground data collection: A sensor node is deployed every 10 meters in the monitoring area. The SHT35 chip is used to collect the temperature and humidity of the root layer soil at a depth of 20 cm. The humidity measurement accuracy is within plus or minus 2% of the relative humidity, and the temperature measurement accuracy is within plus or minus 0.3°C of the relative temperature. The leaf clamped stomatal conductance sensor measures stomatal conductance with a range of 0-10mmol / (m 2 ·s), resolution 0.1mmol / (m 2 ·s), and the data sampling interval is 5 minutes.

[0122] Spatiotemporal alignment and prediction: The spatiotemporal alignment module maps drone observation data and ground sensor data to a 1m×1m grid cell and aligns timestamps through linear interpolation. The LSTM prediction module inputs 48 hours of historical ground data, including temperature, humidity, stomatal conductance, and current spectral characteristics, namely red edge and infrared band reflectance. After passing through a two-layer LSTM network with 128 neurons per layer, it outputs the predicted chlorophyll content value for the next 24 hours. And the predicted value of nitrogen concentration N pred .

[0123] The predicted chlorophyll content value in the next 24 hours and the chlorophyll content detected by the drone in real time are compared. Fusion, generating a joint feature vector, where is the final value of chlorophyll content after fusion; It is the predicted value of chlorophyll content in the next 24 hours predicted by the LSTM prediction submodule; It is the chlorophyll content detection value detected by the drone in real time.

[0124] Among them, in the LSTM prediction module, N pred The generation of is achieved through the following logical deduction:

[0125] Input data: historical ground sensor data (temperature, humidity, and stomatal conductance) and drone spectral characteristics, such as red-edge and infrared reflectance.

[0126] LSTM network structure: 2 layers of LSTM, 128 neurons per layer, a time step of 1 hour, and an input sequence length of 48 hours.

[0127] Output formula: N pred =f LSTM (historical temperature and humidity data, historical stomatal conductance data, current spectral characteristics); where N pred : predicted nitrogen concentration; f LSTM : The nonlinear mapping function of the LSTM network represents the predictive relationship learned by the model through time series data; historical temperature and humidity data: the root layer soil temperature collected by ground sensors in the past 48 hours; historical stomatal conductance data: the stomatal conductance of leaves in the past 48 hours, reflecting the intensity of plant transpiration; current spectral characteristics: canopy reflectance data collected in real time by UAV multispectral sensors, such as the red edge band (735nm) and infrared band (1550nm).

[0128] When continuous cloudy days result in missing drone data, the weight of the predicted chlorophyll content in the next 24 hours is increased to 0.9 to ensure detection continuity.

[0129] The verification feedback module execution mechanism and biochemical dual-path verification process are as follows:

[0130] Mechanism path verification: The nutritional indicators output by the dynamic spectrum-physiological coupling analysis module are used as the ab = 2.8 mg / g, which is then inferred from the PROSAIL model to generate the theoretical spectral reflectance R at 400-2500 nm. theory (λ), calculate the residual sum of squares in the red edge band 700-750nm:

[0131] Among them, RSS is the residual sum of squares of the red edge band (700-750nm), R theory (λ) is the inverted C ab , LAI input PROSAIL model to generate theoretical spectral reflectance, R obs (λ) is the spectral reflectance measured by the UAV, and λ represents the wavelength of the spectrum, which is used to describe the spectral reflectance characteristics of the sugar beet canopy or soil background in different bands.

[0132] If the RSS at the seedling stage exceeds 5% or at the mature stage exceeds 3%, the PROSAIL parameter is triggered to be updated online. For example, if the RSS at the mature stage reaches 4.2% during a test, the system automatically adjusts the LAI parameter from 3.5 to 3.2, reducing the residual error to 2.7%.

[0133] Biochemical pathway verification: 5% of the monitoring area was randomly selected for leaf sampling, and the chlorophyll content was measured using an ASD FieldSpec4 near-infrared spectrometer with a wavelength range of 900-1700nm. Chlorophyll content detected by drone in real time When the deviation exceeds 10% for three consecutive times, freeze the fully connected layer and reinitialize the classifier weights. For example, the chlorophyll content detection value of the regional drone in real time The predicted chlorophyll content in the next 24 hours is 2.5 mg / g, with a deviation of 19.4%, triggering model iteration.

[0134] The cross-region generalization adaptation module improves model adaptability through transfer learning as follows:

[0135] Pre-training: In the black soil area (the source domain), the initial model is trained using 1000 sets of samples, with a batch size of 32, a learning rate of 1e-4, and 200 iterations. The learning rate of 1e-4 is scientific notation, representing 0.0001, that is, the learning rate value is 1×10 -4 ;

[0136] Domain Adaptation: 50 sets of samples were collected in the sandy loam area (the target domain). The maximum mean difference (MMD) loss function was used to align the data distribution, with a Gaussian kernel bandwidth of σ = 1.0. The PROSAIL mechanism layer was frozen and the fully connected layer was fine-tuned. The transferred model's detection accuracy in the sandy loam area only decreased by 8.7%, while the traditional model decreased by 32.1%. Regarding hardware, the drone is equipped with an NVIDIA Jetson Xavier NX module to run the lightweight model, accelerated by TensorRT, and consumes less than 15W. The ground base station uses a Raspberry Pi 4B to upload data via a 4G module, ensuring low-latency remote monitoring.

[0137] Through dynamic spectral-physiological coupling modeling and multi-dimensional feature decoupling technology, the accuracy and real-time performance of sugar beet nutrition detection have been significantly improved; based on the differentiable PROSAIL radiation transfer subnetwork, the system integrates physical mechanisms with deep learning, dynamically corrects the spectral offset caused by canopy structure changes, reduces the root mean square error of nitrogen content detection, and improves accuracy; the time-space-spectrum joint decoupling mechanism divides the growth period window through the GRU network, captures the abnormal slope of the red edge band in the early stage of nutritional stress, that is, the asymptomatic stage, and realizes early warning; the adaptive asymmetric convolution kernel dynamically adjusts its size according to the flight altitude to eliminate the overlapping and aliasing effect of leaves, and the accuracy of spatial feature extraction is improved; the physiological lag compensation module combines LSTM prediction with real-time data fusion to ensure the continuity of detection in severe weather and reduce errors.

[0138] Through dual-path verification and cross-regional generalization design, the system balances robustness with scalable application potential. The mechanistic-biochemical dual-path verification mechanism triggers dynamic model iteration, resulting in minimal fluctuation in accuracy over long periods of time and avoiding the overfitting risk associated with a traditional single-path approach. Cross-regional transfer learning requires only 50 sets of samples from the target region, reducing deployment costs. The edge computing gateway utilizes low-power hardware and lossless compression to achieve efficient data transmission and spatiotemporal synchronization of multi-source data. In practical applications, the system has a low false alarm rate in growing areas, improving efficiency compared to manual inspections while also reducing excessive fertilizer use. This delivers both economic benefits and ecological sustainability, providing a highly reliable solution for precision agriculture.

[0139] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0140] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A sugar beet nutrition detection system based on drone multispectral data, characterized in that: include: A drone multispectral acquisition module is used to obtain multispectral reflectance data of the sugar beet canopy; A dynamic spectral-physiological coupling analysis module is communicatively connected to the UAV multispectral acquisition module and includes a differentiable radiation transfer subnetwork, a time-space-spectral joint dynamic decoupling unit, and a physiological lag compensation unit. The differentiable radiation transfer subnetwork is constructed based on the PROSAIL model, which converts the biochemical parameters of sugar beet leaves, the canopy geometry, and the soil background optical properties into physical constraint equations and embeds them into a deep neural network. The time-space-spectral joint dynamic decoupling unit dynamically screens features through time window partitioning, asymmetric convolution kernels, and spectral attention weights. The physiological lag compensation unit generates a real-time-prediction joint feature vector through a ground sensor network and an LSTM prediction module. The verification feedback module is connected to the dynamic spectrum-physiological coupling analysis module and is used for dual-path inversion verification and model iteration.

2. The beet nutrition detection system based on drone multispectral data according to claim 1, characterized in that: The differentiable radiative transfer subnetwork includes: The physical constraint equation generation layer dynamically adjusts the leaf optical parameters and canopy structure parameters according to the sugar beet growth period; The virtual leaf optical fingerprint generation layer corrects the mechanism model error through back propagation and outputs the spectral characteristics that match the current growth period; The soil noise suppression layer constructs an orthogonal projection matrix to separate soil reflectance and canopy spectrum when the canopy coverage is lower than a preset threshold.

3. The beet nutrition detection system based on drone multispectral data according to claim 1, characterized in that: The time-space-spectrum joint dynamic decoupling unit includes: The time dimension decoupling module extracts the spectral response patterns of different growth stages based on the gated recurrent unit and outputs the identification signal of the sensitive period of nutritional stress; The spatial dimension decoupling module uses an asymmetric convolution kernel to extract canopy texture. The size of the convolution kernel is dynamically adjusted according to the flight altitude of the UAV, and the convolution kernel size is increased every time the flight altitude increases by a set distance. The spectral dimension decoupling module dynamically allocates the contribution of the red edge band and the infrared band through attention weights.

4. The beet nutrition detection system based on drone multispectral data according to claim 1, characterized in that: The physiological hysteresis compensation unit comprises: A ground sensor network is deployed in the monitoring area to collect real-time temperature and humidity of the root layer and stomatal conductance of leaves; A spatiotemporal alignment module maps the timestamps and spatial coordinates of drone observation data and ground sensor data to the same benchmark; The LSTM prediction module inputs historical ground data and current spectral features and outputs the lagged physiological state prediction value within a set period in the future.

5. The beet nutrition detection system based on drone multispectral data according to claim 3, characterized in that: In the spatial dimension decoupling module, the size adjustment rule of the asymmetric convolution kernel is: when the flight altitude of the drone increases by 10 meters, the long axis size of the convolution kernel is expanded by 1.5 times, and the short axis size is expanded by 1.2 times.

6. The beet nutrition detection system based on drone multispectral data according to claim 4, characterized in that: The output of the LSTM prediction module is the predicted value of chlorophyll content and nitrogen concentration for the next 24 hours, which is superimposed with the real-time spectral features of the UAV and then input into the dynamic spectral-physiological coupling analysis module.

7. The beet nutrition detection system based on drone multispectral data according to claim 1, characterized in that: The verification feedback module includes: The mechanism pathway verification submodule infers the nutritional indicators to the differentiable radiation transfer subnetwork to generate a theoretical spectrum and calculates the residual with the actual observed spectrum; The biochemical pathway verification submodule performs deviation analysis between the measured and predicted values ​​of leaf near-infrared spectra in random sampling areas; The model iteration trigger unit starts parameter fine-tuning when the residual or deviation exceeds the limit continuously.

8. The beet nutrition detection system based on drone multispectral data according to claim 7, characterized in that: The residual threshold of the mechanism path verification submodule is dynamically set according to the growth period of sugar beet: the threshold in the seedling stage is 5%, and the threshold in the mature stage is 3%.

9. The beet nutrition detection system based on drone multispectral data according to claim 1, characterized in that: It also includes a cross-region generalization adaptation module, which performs domain adaptive optimization on the pre-trained model through a transfer learning algorithm and uses a maximum mean difference loss function to align the data distribution of different soil types and genotypes.

10. The beet nutrition detection system based on drone multispectral data according to claim 1, characterized in that: The UAV multispectral acquisition module, dynamic spectral-physiological coupling analysis module and verification feedback module realize data interaction through the edge computing gateway. The gateway is deployed between the UAV and the ground base station to compress the transmission data and synchronize the calibration timestamp.

Images