A nondestructive detection method and system for rice SPAD value based on hyperspectral imaging technology

Abstract

The application discloses a kind of rice SPAD value nondestructive testing method and system based on hyperspectral imaging technology, it is related to rice nitrogen content detection technical field, by obtaining controllable light under source domain hyperspectral image and non-controlled field low-altitude remote sensing target domain image extraction standardization one-dimensional spectral feature sequence, constructs three parallel deep feature fusion regression model, connects and non-linear interaction regression to multi-view deep semantic feature, outputs SPAD initial prediction value, on this basis, introduce based on second-order statistic distribution alignment semi-supervised transfer learning framework, through dynamic weight balance mechanism realizes the cross-domain self-adaptive generalization of model from laboratory to complex field environment, input into the model after migration by real-time low-altitude remote sensing data, on-line inversion and generate rice SPAD spatial distribution atlas and nitrogen diagnosis prescription chart, the application has high detection precision and strong cross-scene generalization ability, suitable for fine nutrient management of intelligent agriculture.

Description

Technical Field

[0001] This invention relates to the field of rice nitrogen content detection technology, specifically to a non-destructive detection method and system for rice SPAD value based on hyperspectral imaging technology. Background Technology

[0002] Low accuracy of low-altitude remote sensing prediction in open fields and large interference from environmental background noise: When existing UAV hyperspectral imaging is used in low-altitude operations in open fields, it is easily affected by environmental noise such as uneven natural lighting, aircraft vibration, atmospheric disturbance, and wind speed, as well as complex field backgrounds (such as soil, weeds, shadows, water bodies, etc.). This causes high-frequency spikes, baseline shifts and skews in the original hyperspectral spectral curves, resulting in huge errors in the directly derived rice SPAD values, which lacks feasibility for field application.

[0003] Redundancy of multi-source hyperspectral features and insufficient expressive power of traditional single models: Hyperspectral images typically contain hundreds of bands, with extremely high data dimensionality and strong collinearity. Traditional machine learning models such as Partial Least Squares Regression (PLSR) and Support Vector Regression (SVR) cannot effectively handle high-dimensional nonlinear relationships. Existing single-architecture deep learning networks (such as pure CNN or pure LSTM networks) can only extract spectral features from a single scale or dimension, and cannot simultaneously and collaboratively model local physical absorption peaks of spectral curves, long-term and short-term dependencies of wavelength sequences, and global spectral context self-attention associations. This results in severe redundancy and loss of key physical features during deep transmission within the network, severely limiting the prediction accuracy of SPAD values.

[0004] Cross-device / cross-environment applications face severe "domain shift" and the high annotation costs of remodeling: Ground-based laboratory hyperspectral imaging systems, due to their stable measurement environment, generate highly accurate predictive models; however, these models cannot be directly applied to complex low-altitude UAV platforms because significant differences in spectral distribution exist between the two in terms of sensor performance, shooting altitude, geometric distortion, and light source conditions. Adapting the model to field environments would require large-scale destructive sampling and chlorophyll chemical calibration in the field for each new UAV or field scenario to rebuild the model, consuming enormous human and time costs and defeating the purpose of non-destructive remote sensing monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a non-destructive testing method and system for the SPAD value of rice based on hyperspectral imaging technology, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology, comprising the following steps:

[0007] Step S1, Multi-source hyperspectral data acquisition and standard preprocessing: Acquire source domain hyperspectral images of the rice sample under a fixed controllable light source, and target domain hyperspectral images under uncontrolled field low-altitude remote sensing conditions; Perform background radiation elimination correction, edge noise band clipping, and baseline drift denoising processing on the two sets of hyperspectral images in sequence based on standard radiometric calibration materials, and extract source domain one-dimensional spectral feature sequences and target domain one-dimensional spectral feature sequences with uniform spatial dimension and physical dimensions;

[0008] Step S2, Multimodal Feature Collaborative Modeling and Deep Representation Extraction: Construct a deep feature fusion regression model that includes a parallel feature extraction stage and a deep interactive regression stage; in the parallel feature extraction stage, the one-dimensional spectral feature sequence is synchronously input into multiple deep feature learning networks with different structures and complementary characteristics. Local multi-scale fine-grained waveform representations are extracted through the local waveform feature perception unit, contextual short-term dependence and gradual trend representations between adjacent bands are extracted through the sequence temporal dependency modeling unit, and global attention association representations across the entire band are extracted through the global context association perception unit. The representations output by each unit are aligned and mapped into multiple sets of deep semantic feature vectors with completely unified feature dimensions.

[0009] Step S3, High-dimensional deep feature fusion and task-oriented pre-regression: In the deep interactive regression stage, the multiple sets of deep semantic feature vectors are connected along the feature channel direction to form a high-dimensional multi-view fusion feature vector; the high-dimensional multi-view fusion feature vector is fed into a deep nonlinear interactive regression network, and through layer-by-layer feature dimensionality reduction and deep cross-channel interactive calculation, feature redundancy is eliminated and the initial SPAD prediction value of the rice leaf to be tested is output.

[0010] Step S4, Semi-supervised cross-domain adaptation and model generalization transfer: Construct a semi-supervised transfer learning framework based on the alignment of feature space statistical distributions, using the model pre-trained with source domain data in steps S2 and S3 as the base network; based on labeled samples in the target domain, simultaneously combine the source domain supervised task adaptation loss, the target domain supervised real distribution pull loss, and the second-order statistical distribution difference alignment loss of deep activation representations in the source and target domains in the regression optimization loss; perform end-to-end joint gradient optimization of the multiple loss terms through a dynamic weight balancing mechanism to complete the adaptive generalization of the model from the controllable source domain to the complex field target domain;

[0011] Step S5, Non-destructive testing and crop nutrient characterization output in the field: Obtain real-time low-altitude remote sensing hyperspectral data of rice to be tested in the field under unknown operating conditions, perform standardized preprocessing, and input it into the deep feature fusion regression model completed by the migration adaptation in step S4. Online inversion and output of SPAD quantitative prediction results of rice in the test area.

[0012] Preferably, the method for extracting the one-dimensional spectral feature sequence specifically includes the following steps:

[0013] Dark field background data and standard diffuse reflectance reference plate data of the original hyperspectral intensity of the rice under test in the source domain and target domain were collected respectively. Through linear correction calculation based on the dark-white dual reference ratio, the digital quantization value of the original image intensity was converted into absolute physical reflectance data to eliminate the interference of dark current of imaging sensor and uneven illumination of heterogeneous ambient light sources.

[0014] Based on the signal-to-noise ratio response curves of the spectral sensor at different characteristic wavelengths, the edge bands located at both ends of the light wave, with a signal-to-noise ratio below the preset threshold and severely polluted by environmental noise are adaptively cut off, and only the effective characteristic spectral ranges that are highly correlated with crop nitrogen and chlorophyll absorption are retained.

[0015] For the spectral data after removing the edge bands, a sliding window smoothing filter algorithm is first used to eliminate high-frequency random instrument noise and spikes. Then, a standard normal variable transformation algorithm is used to perform full-band centering and scaling standard processing on the spectral sequences of each sample to eliminate baseline shift and tilt caused by multi-angle surface scattering from the low-altitude remote sensing blades in the field. The output is a standard one-dimensional spectral feature sequence with homogeneous feature dimensions and unified physical dimensions.

[0016] Preferably, the multiple sets of deep semantic feature vectors are connected along the feature channel direction to generate a high-dimensional multi-view fusion feature vector with a second feature dimension that is three times the number of the first feature dimension.

[0017] The deep nonlinear interactive regression network is specifically a non-fully connected convolutional regression head composed of multiple layers of one-dimensional convolutional blocks that are progressively reduced in dimensionality. It includes: a feature reshaping layer for converting the high-dimensional multi-view fused feature vector into a sequence mapping tensor; at least two layers of one-dimensional convolutional and average pooling components that progressively reduce the dimensionality of feature channels and perform spatial aggregation; and an output projection layer connected to the end of the network for mapping interactive semantics into scalar SPAD regression values.

[0018] Preferably, the specific method for outputting the initial SPAD prediction value of the rice leaf to be tested in step S3 includes the following steps:

[0019] The high-dimensional multi-view fusion feature vector generated by splicing is transformed from a one-dimensional physical vector structure into a two-dimensional sequence feature tensor suitable for convolution kernel sliding through a feature reshaping operator, so that the feature channels extracted by heterogeneous units have spatial and sequence-related topology.

[0020] The two-dimensional sequence feature tensor is continuously input into at least two cascaded one-dimensional convolutional and pooling layers. By utilizing the local sliding mapping of the one-dimensional convolutional kernel on the feature channel dimension, the dimension of the feature channel is reduced step by step, and the deep nonlinear cross-correlation between the representations of different perceptual units is captured.

[0021] An adaptive average pooling operation is performed on the interaction feature tensor after layer-by-layer compression to aggregate and compress the non-fixed-length sequence dimensions into a globally compressed feature vector with a unique length, which greatly eliminates redundant background noise generated by multi-source feature splicing.

[0022] The global compressed feature vector is progressively characterized and refined through a multi-layer fully connected network with layer normalization and nonlinear activation functions. Finally, the scalar regression prediction results of the SPAD values ​​of the rice leaves and canopy are obtained by inversion through the output projection layer at the end of the network.

[0023] Preferably, the adaptive generalization method from the controllable source domain to the complex field target domain in step S4 specifically includes the following steps:

[0024] Labeled hyperspectral samples are extracted from the laboratory source domain dataset and input into the network. The mean square error between the initial SPAD predictions of the network and the actual physical sampling labels is calculated, and a source domain supervised task adaptation loss is constructed. The hyperspectral to crop physiological parameter mapping relationship that dominates the foundation of network learning;

[0025] A very small number of labeled samples are extracted from the low-altitude remote sensing target domain dataset of Datian and input into the network. The mean square error between the corresponding predicted values ​​and the field measured labels is calculated, and a supervised real distribution pull loss for the target domain is constructed. This provides real physical space anchor point constraints for cross-domain adaptation in high-dimensional feature spaces;

[0026] The source domain labeled samples and the target domain labeled samples are mixed and input into the network, and the activation feature matrices of the two types of samples at the end of the deep feature projection layer are extracted simultaneously. The empirical covariance matrices of the source domain deep feature matrix and the target domain deep feature matrix are calculated respectively, and the second-order statistical distribution difference alignment loss is constructed by calculating the nonlinear norm statistical distance between the two empirical covariance matrices. ;

[0027] In the backpropagation optimization process, based on real-time computation and The numerical unit ratio is used to generate adaptive weight balance coefficients in real time. Combination An end-to-end joint gradient descent is performed on the overall network weights to collaboratively achieve cross-domain statistical distribution alignment and refined regression adaptation in complex field environments.

[0028] Preferably, the multiple deep feature learning networks with different structures in step S2 specifically include:

[0029] Local waveform feature sensing unit: a one-dimensional convolutional neural network module with multi-scale receptive fields, which contains at least two sets of cascaded deep convolution operators, batch normalization operators and nonlinear activation operators, used to focus on capturing local subtle features of chlorophyll at characteristic wavelengths.

[0030] Sequence temporal dependency modeling unit: This is a recurrent gated network module with sequence context memory function. It contains a multi-layer stacked long short-term memory network structure, which is used to learn the dependencies between bands point by point according to wavelength growth order.

[0031] Global Context Association Aware Unit: This is a graph or sequence modeling module based on position encoding and attention mechanisms. It contains a multi-head self-attention encoding structure, which is used to directly calculate the cross-self-attention weights between any two feature wavelength positions in the input sequence.

[0032] Each of the three network units has a feature projection layer attached to its end, which maps the output representation uniformly to a vector space with a preset first feature dimension.

[0033] Preferably, after outputting the attention representation sequence, the global context association perception unit further compresses the multi-band features into a fixed-length feature representation through an adaptive sequence aggregation operator, and then generates a semantically refined deep semantic feature vector of the first feature dimension through a fully connected network with distribution normalization processing.

[0034] Preferably, step S4 includes multi-stage progressive optimization logic for training control that enables adaptive generalization of the model from a controllable source domain to a complex field target domain, specifically including:

[0035] Task warm-up phase: Masking feature distribution alignment loss There is supervised loss with respect to the target domain. The depth feature fusion regression model is optimized using only laboratory source domain hyperspectral samples to establish a robust initial state mapping from physical spectra to crop physiological parameters.

[0036] Joint Adaptive Phase: The three functional loss terms are enabled simultaneously, and the alignment strength coefficient is dynamically amplified with each training iteration cycle, driving the deep network to maintain the ability to regress and discriminate crop physiological parameters while eliminating the spectral domain distribution shift caused by the low-altitude remote sensing background.

[0037] Distribution fine-tuning stage: Freeze the parameters of each deep feature learning network in the previous stage, and use only a small number of real physical sampling labels in the target domain to fine-tune the deep nonlinear interactive regression network at the end with a low learning rate to eliminate the residual system error caused by the mixed pixels of the UAV.

[0038] Preferably, the second-order statistical distribution difference alignment loss specifically includes deep correlation alignment loss, which constructs the empirical covariance matrix of the source domain activation feature matrix and the empirical covariance matrix of the target domain activation feature matrix, and uses nonlinear matrix norm to calculate the statistical discrete distribution difference between the two in the same high-dimensional mapping space, thereby achieving forced alignment of cross-environment spectral feature expressions.

[0039] Preferably, in step S1, the background radiation elimination correction based on the standard radiation calibration object specifically involves using a dark noise reference quantity and a diffuse reflection standard reference quantity to eliminate the influence of dark current and uneven illumination of the device through a linear ratio calculation. The specific algorithm is expressed as follows:

[0040] ;

[0041] Where I represents the spectral intensity value of the original hyperspectral image, D represents the spectral intensity value of the dark field background, W represents the spectral intensity value of the standard diffuse reflectance reference calibrator, and R represents the corrected true reflectance data; the edge noise band clipping is to cut off the front and back end distortion bands according to the sensor signal-to-noise ratio threshold; the baseline drift denoising process is implemented by using a sliding window-based moving smoothing filter algorithm combined with a standard normal variable transformation algorithm.

[0042] Preferably, in step S4, the total optimization loss function used in the optimization training process is composed of three functional loss terms adaptively balanced and weighted, and its mathematical logical form is expressed as follows:

[0043] ;

[0044] in, This represents the supervised task adaptation loss determined based on the prediction bias of labeled samples in the source domain. This indicates the supervised true distribution pull loss for the target domain, determined based on the prediction bias of a small number of labeled samples. The covariance alignment loss represents the difference between the deep feature distribution matrices of the source domain and the deep feature distribution matrices of the target domain on a second-order statistical index. The effective weight ratio coefficient is controlled by a dynamic adaptive mechanism.

[0045] Preferably, to prevent feature alignment loss caused by multidimensional hyperspectral data The effective weight ratio coefficient is too large in numerical dimensions, thus suppressing the gradient direction of the main regression task. Based on a dynamic weight balancing mechanism that updates in real time, its mathematical expression is:

[0046] ;

[0047] ;

[0048] in, As a dynamic adaptive balance factor, To prevent division by zero anomalies in extremely small positive real constants, The hyperparameter for setting the baseline alignment strength is used; through the mechanism described, the gradient update magnitude of the migration alignment loss and the end-to-end task adaptation loss during backpropagation is forced to remain equal and stable.

[0049] Furthermore, this invention also provides a non-destructive testing system for the SPAD value of rice based on hyperspectral imaging technology, comprising:

[0050] Multi-source hyperspectral sensing and data link module: It includes a ground-based stationary hyperspectral imaging acquisition end and an airborne low-altitude hyperspectral remote sensing end for field, which are used to acquire the original spectral images of the rice to be tested and transmit them to the cloud or local computing core.

[0051] Standard preprocessing and feature extraction module: Coupled downstream of the data link module, it is used to perform adaptive background radiation compensation, automatic noise band removal and baseline smoothing on the input spectral image online, and output a standardized feature sequence.

[0052] Multi-feature collaborative extraction and semi-supervised cross-domain processing module: It has a built-in high-performance computing processor and non-volatile storage medium. The storage medium has a deep feature fusion and transfer processing program embedded in it. When the processor executes the program, it calls the complementary multi-channel neural network in parallel to extract deep semantic features and connects them. At the same time, it uses a dynamic adaptive transfer algorithm to align and generalize the indoor pre-trained model to the complex field test environment in real time.

[0053] The task representation output and intelligent decision-making module is used to receive the quantitative prediction values ​​generated by the multi-feature collaborative extraction and semi-supervised cross-domain processing module, dynamically generate the electronic map of the spatial distribution of rice SPAD values ​​and the prescription map for nitrogen nutrient deficiency diagnosis in the corresponding spatial coordinate system, and instruct or assist in the intelligent variable fertilization operation in the field.

[0054] In summary, the beneficial effects of this invention are:

[0055] This invention abandons the traditional single-network structure. In the parallel feature extraction stage, it creatively mines the deep semantics of spectral curves from three completely independent perspectives through structurally distinct "local waveform perception network, long-term and short-term dependency modeling network, and global self-attention encoding network." This architecture can focus on the subtle local photophysical absorption characteristics of chlorophyll molecules at specific characteristic wavelengths, capture the temporal variation trends of adjacent high-width and low-frequency spectral sequences, and capture global long-distance dependencies across wavelengths across the entire spectrum. The multi-perspective fused feature vector formed by the fusion of these three technologies has extremely high information entropy and strong semantic complementarity, completely breaking through the technical bottleneck of traditional models in the extraction of complex hyperspectral features.

[0056] To address the issues of dimensionality curse and information redundancy that easily arise from multi-source feature concatenation, this invention replaces the traditional direct fully connected (MLP) regression structure with a one-dimensional convolutional regression head. Before regression, the concatenated features are transformed into a reshaped tensor with a sliding one-dimensional convolutional kernel through feature topology reconstruction. Convolutional operations are then used to perform cross-channel deep nonlinear interactions and adaptive compressed pooling along the feature channel directions. This approach significantly reduces model computation and feature redundancy, while the shared weight constraint enables the deep model to exhibit extremely strong generalization stability and regression prediction accuracy in small-sample scenarios such as rice cultivation.

[0057] This invention, through an improved semi-supervised DeepCoral transfer method, perfectly solves the industry challenge of directly applying high-precision laboratory models to field UAV platforms. During the transfer alignment process, it not only utilizes second-order statistical alignment constraints between the source and label-target domain data, but also innovatively introduces a very small number of labeled samples from the target domain as real physical space anchor points. This design allows us to rapidly eliminate spectral domain distribution shifts caused by sensor heterogeneity, changes in ambient light, and field background interference using only a very low proportion of contact-based manual testing in the field. This enables the perfect and low-cost application of high-precision laboratory ground models to low-altitude field scenarios. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the overall process framework of a non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to the present invention.

[0060] Figure 2This is a schematic diagram of a laboratory-taken hyperspectral image of a sample in a non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to the present invention.

[0061] Figure 3 This is a schematic diagram of the laboratory spectral curve of rice variety C in the non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology of the present invention.

[0062] Figure 4 This is a schematic diagram of the spectral curve of rice from a drone for variety C in the non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology of the present invention.

[0063] Figure 5 This is a scatter plot of the actual and predicted values ​​of the SPAD value of rice for variety C in a laboratory CNN model for a non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to the present invention.

[0064] Figure 6 This is a scatter plot of the actual and predicted values ​​of the laboratory CNN model for all varieties in the non-destructive testing method of rice SPAD value based on hyperspectral imaging technology of the present invention.

[0065] Figure 7 This is a scatter plot of the actual and predicted values ​​of the deep feature fusion regression model for rice variety C in the non-destructive testing method for rice SPAD value based on hyperspectral imaging technology of the present invention.

[0066] Figure 8 This is a scatter plot of the actual and predicted values ​​of the deep feature fusion regression model for all varieties in the non-destructive testing method of rice SPAD value based on hyperspectral imaging technology of the present invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. These drawings are simplified schematic diagrams, which are only used to illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0068] To facilitate understanding of the present invention, a more complete description of the invention will be given below with reference to the accompanying drawings, which illustrate several embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the invention will be more thorough and complete.

[0069] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0070] Any feature disclosed in this specification (including any appended claims, abstract, and drawings) may be replaced by other equivalent or similar features for a similar purpose, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0071] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; they can refer to the internal communication of at least two elements or the interaction relationship of at least two elements, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0072] Please see Figures 1-8 This invention provides an embodiment of a non-destructive method for detecting SPAD values ​​in rice based on hyperspectral imaging technology. The SPAD (Soil and Plant Analyzer Development) value is used to directly characterize and indicate the nitrogen nutrient levels in rice leaves and canopy. Experimental data were collected from experimental fields at the Wuxing District Agricultural Science Research Institute in Huzhou City, Zhejiang Province. The research subjects included three rice varieties, denoted by codes A, B, and C in this embodiment. Figure 2 The hyperspectral images captured in the image correspond to a, b, and c in the figure, respectively. The specific method steps are as follows:

[0073] Step 1: Acquisition and Standardization Preprocessing of Hyperspectral Data of Rice

[0074] 1.1 Data Acquisition System Configuration

[0075] The hardware architecture used in this embodiment includes a ground-based laboratory data acquisition system and a low-altitude unmanned aerial vehicle (UAV) remote sensing platform:

[0076] Ground-based: The hyperspectral imager used was an FX10 spectrophotometer from SPECIM, Finland, with a spectral acquisition range of 397–1004 nm. The test sample was placed on a motorized stage, with the camera maintaining a constant focal length of 30 cm relative to the leaf. The stage moved at a constant speed of 24 mm / s, and data acquisition was controlled using LUMOScanner software. Immediately after acquiring the hyperspectral images, the true SPAD value of the rice leaves was determined using a SPAD-502 chlorophyll meter. Measurements were taken uniformly at three points on each leaf: the tip, middle, and base, and the average value was used as the physical label for the corresponding sample. Figure 3 Spectral curve of rice variety C in the text.

[0077] Low-altitude UAV terminal: The flight platform uses a large UAV equipped with an Isuzu 9010L airborne hyperspectral sensor. The sensor detector is CMOS, with a spectral range of 400~1000nm and a spectral resolution better than 2.8nm. The UAV flight route is selected during cloudless, stable lighting conditions from 9:00 AM to 11:00 AM, with a flight altitude set at 25m. Figure 4 The spectral curve of rice variety C is shown.

[0078] 1.2 Absolute Radiative Reflectivity Calibration

[0079] Due to physical limitations such as lighting, environmental variations, and sensor dark current noise, the digital quantization values ​​(DN values) captured by the original hyperspectral camera cannot be directly used for depth feature extraction. This invention implements absolute radiometric reflectance calibration: during ground-based imaging, a completely dark reference image is captured before each shot by covering the lens cap, and a white reference image is captured using a standard white Teflon plate (with a reflectance close to 100%). During UAV flight in Daejeon, a standard diffuse reflectance calibration plate is laid flat on the ground, and the UAV simultaneously acquires hyperspectral images of this calibration plate during takeoff and landing.

[0080] The formula for calculating the corrected reflectance R is as follows:

[0081] ;

[0082] Where I represents the original spectral intensity DN value of the rice sample; D represents the DN value of the completely dark background (reflectivity approximately 0%); and W represents the DN value of the white reference calibration plate (reflectivity approximately 100%). This ratio calculation completely eliminates the physical interference of uneven lens illumination, lens transmittance variations with wavelength, and internal camera dark current noise on the spectral reflectance curve.

[0083] 1.3 Adaptive Truncation of Edge Distortion Bands

[0084] Hyperspectral imaging sensors exhibit low optical response and severely degraded signal-to-noise ratio (SNR) at the extremes of the spectral range (ultraviolet and near-infrared edges), often accompanied by intense electrical noise. To prevent unwanted edge noise from interfering with the learning of deep networks:

[0085] For ground-based laboratory cameras with a large number of original bands, the first 61 bands (mainly ultraviolet bands below 400nm) and the last 103 bands (mainly near-infrared edge attenuation bands above 950nm) were adaptively removed, ultimately retaining 284 core bands.

[0086] For UAV-borne hyperspectral sensors, the first 40 bands at the front end and the last 75 bands at the back end are adaptively cut off, ultimately retaining 185 core bands. The spectral range after the cut-off retains the strong absorption characteristics (red edge effect) of rice chlorophyll in the red and blue bands, as well as the high reflectance characteristics of the near-infrared platform.

[0087] 1.4 Baseline translation cancellation and high-frequency denoising

[0088] Although the cropped spectral signal removes distortion at both ends, the spectral curve in the field remote sensing image still contains random noise and overall baseline shift due to variations in leaf tilt angle, multi-angle scattering, and minor sensor jitter. This invention employs a moving smoothing filter combined with the Standard Normal Variable Transform (SNV) algorithm for denoising.

[0089] Moving smoothing filter: Ground spectral data uses 5-point moving smoothing filter; UAV hyperspectral data, due to greater noise, uses cascaded two-stage 7-point moving smoothing filter, which effectively eliminates high-frequency instrument glitches while preserving the gentle slope and characteristic absorption peak regions of the spectral waveform.

[0090] Standard Normal Variable Transform (SNV): This transforms each smoothed spectral sample sequence to eliminate baseline drift caused by scattering from solid particle surfaces, uneven microstructure of blades, or changes in the optical path. For a spectral sample containing N bands... Its transformation formula is:

[0091] ;

[0092] in, This represents the mean spectral reflectance of the sample. This represents the standard deviation of the spectral reflectance of the sample. The SNV transformation converts each spectral sample into a standardized sequence with a mean of 0 and a standard deviation of 1, ensuring complete alignment of the physical dimensions between the two heterogeneous hyperspectral datasets from the ground and the UAV.

[0093] Step 2: Joint and collaborative deep extraction of multiple features and nonlinear interactive regression inversion

[0094] By utilizing the constructed deep feature fusion regression model, multi-view parallel feature extraction and deep channel interactive regression are performed on the preprocessed source / target domain spectral feature sequences.

[0095] 2.1 Three-way parallel feature extraction architecture

[0096] The processed (Ground D=284, UAV D=185) The one-dimensional sequence of spectral reflectance is synchronously input into three parallel heterogeneous sub-networks to extract features from three completely independent perspectives: local receptive field, temporal waveform sequence evolution, and global correlation self-attention.

[0097] 2.1.1 Local Waveform Feature Sensing Unit

[0098] Composed of three cascaded one-dimensional convolutional blocks, it is used to extract local multi-scale physical absorption features of a one-dimensional spectral reflectance sequence:

[0099] First convolutional block: Input is The sequence uses a one-dimensional convolution operator Conv1d (kernel size of 7, stride of 3) with 64 output channels; followed by a one-dimensional batch normalization layer BatchNorm1d, a ReLU activation function, and a max pooling layer MaxPool1d (kernel size of 2, stride of 2).

[0100] The second convolutional block uses the one-dimensional convolution operator Conv1d (kernel size of 5, stride of 2) with 128 output channels; followed by BatchNorm1d, ReLU activation function and MaxPool1d (kernel size of 2, stride of 2).

[0101] The third convolutional block contains two cascaded one-dimensional convolutions, namely input channel 128 and output channel 256 (kernel size 3, stride 1) and input channel 256 and output channel 512 (kernel size 3, stride 1); finally, the sequence dimension is compressed by the AdaptiveAvgPool1d(1) adaptive average pooling layer.

[0102] At the end, a multilayer perceptron containing three fully connected layers, Linear(512,1024)->Linear(1024,512)->Linear(512,256), uniformly maps the local waveform representation to a 256-dimensional one-dimensional convolutional feature vector. .

[0103] 2.1.2 Sequence Temporal Dependency Modeling Unit

[0104] Spectral reflectance, with continuous increase in wavelength, essentially exhibits a "time-series variation curve" along the wavelength dimension. This module employs a three-layer stacked Long Short-Term Memory (LSTM) network to sequentially read the spectral sequence and adaptively model the gradient dependence and gradual change trend of the wavelength reflectance signal before and after wavelength changes.

[0105] Network structure: It consists of an LSTM with 128 hidden units in the first layer, an LSTM with 64 hidden units in the second layer, and an LSTM with 32 hidden units in the third layer. Each layer runs in channel-first mode (batch_first=True).

[0106] Feature transformation: The state features output by the third-layer LSTM network at the last spectral point are taken and passed through a linear projection layer consisting of Linear(32,512), a one-dimensional batch normalization layer, a ReLU activation function, and Linear(512,256). At the end, a LayerNorm(256) normalization layer is introduced to generate a 256-dimensional sequence time feature vector. .

[0107] 2.1.3 Global Context Awareness Unit

[0108] In spectral absorption characteristics, strong biochemical coupling exists between specific spectral regions that are geographically distant (such as the red absorption valley and the near-infrared reflectance plateau). This module utilizes a self-attention mechanism to model this global long-range band correlation:

[0109] Input projection and position encoding: First, the original spectral reflectance is projected into a 64-dimensional low-dimensional space through a one-dimensional linear projection layer Linear(D,64), and an additional position encoding vector is added to preserve the wavelength order topology of the spectral curve.

[0110] Transformer encoder: It contains two stacked Transformer encoders, each with 4 multi-head self-attention heads (Heads=4) and a 128-dimensional feedforward network, which adaptively learns the correlation weights between the reflectance of any two bands in the full spectrum.

[0111] Aggregate projection: Sequence adaptive average pooling is performed on the encoder output, followed by a Linear(256,128) projection transformation layer with LayerNorm(256) constraints to output a 256-dimensional global self-attention feature vector. .

[0112] 2.2 Deep Nonlinear Interactive Regression Network

[0113] After obtaining the aforementioned three-view deep feature vectors, the conventional approach is to directly concatenate and stitch them together with a fully connected layer for regression. However, this is prone to overfitting in large-scale, low-sample scenarios due to excessive parameters and isolated information. This invention proposes a CNN regression head that includes topological reconstruction, convolutional interaction, spatial aggregation, and projection mapping.

[0114] 2.2.1 Multi-source high-dimensional feature topology reconstruction: ... and Three 256-dimensional vectors are concatenated along the feature dimensions to form a 768-dimensional multi-view fusion feature vector. This is then processed by a feature reshaping operator. A one-dimensional vector is converted to a vector of size . Alternatively, it can be reshaped into a two-dimensional topological tensor with feature channels (reshaped tensor), simulating a spatial feature map with spatial depth topology, and endowing the features of each heterogeneous unit with local sequence adjacency in the convolution sliding.

[0115] 2.2.2 Layer-by-layer cross-channel nonlinear interaction and feature compression: The reshaped feature tensor is continuously fed into four cascaded one-dimensional convolutional layers, and the one-dimensional convolutional kernel slides and maps along the reconstructed feature channel direction:

[0116] The first layer of one-dimensional convolution Conv1d(1,256,kernel_size=7,stride=2,padding=3) combines BatchNorm1d and ReLU activation;

[0117] The second layer of one-dimensional convolution Conv1d(256,128,kernel_size=5,stride=2,padding=2) combines BatchNorm1d and ReLU activation;

[0118] The third layer of one-dimensional convolution Conv1d(128,64,kernel_size=3,stride=2,padding=2) combines BatchNorm1d and ReLU activation;

[0119] The fourth layer, a one-dimensional convolution Conv1d(64,32,kernel_size=3,stride=2,padding=1), combines BatchNorm1d and ReLU activation. Due to the weight sharing characteristic of the one-dimensional convolution kernel, the four-layer convolution operation performs multiple non-linear cross-mappings in the feature depth dimension, forcing heterogeneous features to interact and fuse in the deep layer, which greatly eliminates the information redundancy introduced by multi-source splicing.

[0120] 2.2.3 Adaptive global spatial aggregation: The output of the fourth convolution layer is automatically aggregated in space through a one-dimensional adaptive average pooling layer AdaptiveAvgPool1d(1), which merges redundant and non-fixed-length feature dimensions into a 32-dimensional global compressed feature vector with a unique length, thus filtering out irrelevant environmental and model background noise from a statistical perspective.

[0121] 2.2.4 Feature Regularization and Scalar Inversion Output: The 32-dimensional vector is input into a cascaded fully connected network with layer normalization layers and nonlinear activation functions: Linear(32,256)->LayerNorm(256)->ReLU->Linear(256,128)->LayerNorm(128)->ReLU->Linear(128,64)->ReLU->Linear(64,1). Layer normalization (LayerNorm) forces the stable distribution range of activation features during training. Finally, the scalar SPAD prediction values ​​of the rice leaf surface and canopy are output through the output projection layer.

[0122] Step 3: Semi-supervised domain adaptive and progressive transfer training

[0123] Hyperspectral images from UAVs in Datian are easily affected by lighting, shadows, and soil background, resulting in extremely low accuracy in direct modeling and prediction. In contrast, laboratory imaging environments are stable, leading to high model accuracy, but significant domain shift occurs due to differences in equipment and environment. This embodiment details how to seamlessly transfer ground models to the Datian remote sensing platform through semi-supervised transfer learning.

[0124] In this invention, a ground-based laboratory hyperspectral dataset and real SPAD labels are used as the source domain, and a Datian UAV hyperspectral remote sensing dataset is used as the target domain. The target domain contains a large amount of partially labeled remote sensing image data, and a very small amount of samples with real SPAD labels obtained through Datian fixed-point quadrat testing.

[0125] The objective function for model transfer training consists of three specific function loss terms:

[0126] ;

[0127] 3.1.1 Source domain has supervised task adaptation loss

[0128] Using all labeled samples from the ground laboratory Forward propagation outputs predicted values Its actual SPAD tag value Calculate the mean squared error loss (MSE):

[0129] ;

[0130] in, This represents the number of samples trained in each batch of the source domain. This loss method guides the model network to establish an accurate, high-resolution, noise-free spectral mapping to crop SPAD values ​​as an initial state regression mechanism.

[0131] 3.1.2. Supervised real distribution pull loss in the target domain

[0132] A very small number of labeled samples are extracted from the target domain of the field, for example, only a very small number of measured samples from the fourth cell in five batches of experimental data in the field, and their forward prediction values ​​are calculated. SPAD physical tags measured in Daejeon Mean squared error loss between:

[0133] ;

[0134] in, This represents a very small number of labeled training samples for the target domain. This semi-supervised loss term acts as a strong anchor constraint, driving the deep feature space to align with the target domain while maintaining sensitivity to optical and physical degradation caused by mixed pixels or multi-angle scattering in the field scene being tested.

[0135] 3.1.3 Alignment Loss of Second-Order Covariance Statistical Distribution

[0136] To ensure that the network outputs unbiased (neighborhood-invariant) feature representations of source and target domain samples in the feature space, this invention simultaneously feeds batches of labeled samples from the source domain and batches of unlabeled samples from the target domain into the network. The activation feature representations of the two types of data are extracted from the projection output layer (corresponding to a 256-dimensional space) of the deep feature fusion model, and the source domain covariance matrix is ​​constructed accordingly. Covariance matrix of target domain .

[0137] The statistical distance of the second-order statistical covariance matrix is ​​calculated using the Frobenius norm, and an alignment loss is constructed. :

[0138] ;

[0139] Where d is the dimension of the activation feature (here d=256). By minimizing The network was forced to adjust the weights of the shallow one-dimensional convolution and gating network to narrow the second-order correlation distribution distance between the ground spectrum and the UAV field remote sensing spectrum in deep space.

[0140] 3.1.4 Adaptive Dynamic Weight Scaling Mechanism

[0141] Alignment loss due to the feature covariance calculated from high-dimensional hyperspectral data The absolute value of the gradient is usually very large, which can easily generate an excessively large backward gradient in the early stages of training, suppressing the regression loss of the main task. and This can lead to parameter collapse or non-convergence in the model during gradient backpropagation. This invention introduces a dynamic adaptive balance factor. :

[0142] ;

[0143] ;

[0144] in, It is a very small constant used to prevent the denominator from being zero and to ensure the stability of the calculated numerical value; The base transfer alignment constant is set, preferably 1.0. After each iteration of forward computation, the model dynamically adjusts the effective alignment weights. This forces the backward gradient magnitudes of the transfer alignment loss and supervised regression loss to be controlled to be of the same order of magnitude, thereby ensuring extremely high convergence speed of joint optimization and robustness of training.

[0145] 3.2 Three-stage incremental optimization logic

[0146] To ensure stable generalization of complex heterogeneous networks with multiple loss terms, this invention implements a three-stage progressive training control strategy:

[0147] Task warm-up phase: The Coral alignment loss and target domain loss are masked, and only samples with SPAD labels from the ground laboratory are input into the model for basic regression training. The training is run for 200 epochs, so that the network substructure learns a stable basic discriminative power for mapping spectral features to SPAD.

[0148] Joint Adaptive Phase: Enable all three loss terms , and Utilizing dynamic scaling weights Constraints are applied at this stage, where the model is forcibly aligned between the low-altitude high-dimensional remote sensing image of Datian and the ground spectrum, and the background is removed due to optical nonlinear noise shifts caused by uneven illumination and surface shading. The model is run for 500 epochs.

[0149] Distribution fine-tuning stage: Freeze the weights of the front-end CNN, LSTM, and Transformer modules (reduce their learning rate to 1 / 10 of the initial value), activate only the tail interactive convolutional regression head network, use a very small number of labeled field samples in the target domain to fine-tune the model globally with a low learning rate, run for 100 epochs, so that the final output of the model thoroughly fits the actual test conditions of the field drone.

[0150] Furthermore, this invention, combined with a system functional module architecture, provides a specific apparatus and system for hardware deployment of the algorithm of this invention:

[0151] Multi-source hyperspectral sensing and data link module: The ground end includes a SPECIMFX10 laboratory camera, and the field remote sensing end includes an Isuzu9010L hyperspectral camera carried by a drone. During field operations, the drone scans the paddy field's reflectance DN value images in the air and transmits the raw SPE data back to the field edge computing terminal or local processing server in real time via an onboard 5G network or wireless downlink.

[0152] Standard preprocessing and feature extraction module: The edge computing unit calls the MAXSPEX image processing software tool online in the background: to perform absolute diffuse whiteboard correction, automatically remove front edge noise wavelengths, and use 7-point smoothing denoising and standard normal variable transformation (SNV) to achieve background elimination and output a dimensionally standardized one-dimensional reflectance spectral feature sequence.

[0153] Multi-feature collaborative extraction and semi-supervised cross-domain processing module: It has a built-in high-performance GPU computing card, and the storage medium has the aforementioned trained deep feature fusion model embedded in it. The processor receives the processed one-dimensional spectral feature sequence, calls the parallel processing units of CNN, LSTM and Transformer, and performs regression calculations through a one-dimensional convolutional interactive regression head to directly obtain the quantitative SPAD prediction scalar for each pixel region / cell online.

[0154] The operational representation output and intelligent decision-making module: Predictive data is calibrated using GIS geospatial coordinates to dynamically generate an online 3D spatial distribution visualization map of SPAD / nitrogen levels across the entire paddy field. The system incorporates diagnostic logic for the critical nitrogen nutrient index (NNI) during the rice growth period, automatically generating "variable fertilization prescription electronic drawings" for corresponding precise field plots. It also wirelessly connects to intelligent unmanned variable fertilization machines in the field, precisely adjusting the opening of urea or compound fertilizer dispensing valves to achieve intelligent precision agricultural management.

[0155] Furthermore, the deep feature fusion regression model proposed in this invention was compared with classic machine learning and single deep network architecture on various hyperspectral datasets of laboratory and UAV types. Pearson correlation coefficient (r) and root mean square error (RMSE) were used as quantitative evaluation indicators on the test set.

[0156] Table 1: Comparison of SPAD inversion test performance of various models on laboratory ground datasets for rice.

[0157] (A) r 0.888 0.898 0.852 0.903 0.882 0.876 0.904 RMSE 1.168 1.114 1.464 1.069 1.237 1.397 1.152 (B) r 0.887 0.906 0.911 0.919 0.909 0.887 0.910 RMSE 0.832 0.768 0.915 1.157 0.895 0.972 0.863 (C) r 0.903 0.904 0.909 0.915 0.915 0.871 0.923 RMSE 0.863 0.808 0.952 1.086 0.719 1.711 0.704 All mixed datasets r 0.857 0.876 0.911 0.886 0.846 0.833 0.899 RMSE 1.059 0.979 0.915 0.884 1.177 1.597 0.933

[0158] As shown in Table 1, under controlled datasets in a ground-based laboratory, the deep feature fusion model constructed in this invention significantly outperforms classic machine learning and single network structures on most subsets of datasets. For example, on the Ning 81(C) dataset, the correlation coefficient of this invention on the test set reaches 0.923, and the RMSE decreases to 0.704, greatly improving the model's fitting accuracy and stability against background interference.

[0159] Table 2: Comparison of SPAD inversion test performance of various models on the Datian UAV hyperspectral remote sensing dataset.

[0160] (A) r 0.653 0.398 0.644 0.613 0.559 0.547 0.605 RMSE 1.392 1.664 1.683 1.389 1.490 2.079 1.823 (B) r 0.731 0.723 0.750 0.760 0.693 0.705 0.725 RMSE 1.138 1.153 1.128 1.037 1.219 1.416 1.179 (C) r 0.554 0.558 0.526 0.532 0.542 0.551 0.569 RMSE 1.384 1.353 1.379 1.346 1.330 1.274 1.321 All mixed datasets r 0.554 0.602 0.665 0.625 0.619 0.573 0.679 RMSE 1.576 1.523 1.489 1.618 1.955 1.664 1.441

[0161] As shown in Table 2, in the complex low-altitude UAV remote sensing environment of Datian, the performance of all conventional methods for direct modeling was significantly limited due to various weather and soil background interferences. The correlation coefficient of variety C on the test set dropped to around 0.52 to 0.55. In contrast, the deep feature fusion model of this invention exhibited strong feature robustness, achieving a correlation coefficient of 0.569 on variety C on the test set, comprehensively surpassing all comparative benchmarks. After mixing datasets from all varieties, the correlation coefficient reached 0.679, demonstrating extremely strong cross-variety generalization performance.

[0162] The above description is merely a specific embodiment of the invention, but the scope of protection of the invention is not limited thereto. Any variations or substitutions conceived without inventive effort should be included within the scope of protection of the invention. Therefore, the scope of protection of the invention should be determined by the scope defined in the claims.

Claims

1. A non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology, characterized in that: Includes the following steps: Step S1: Acquire the source domain hyperspectral image of the rice sample under a fixed controllable light source, and the target domain hyperspectral image under uncontrolled field low-altitude remote sensing conditions. One-dimensional spectral feature sequences of the source domain and the target domain with uniform spatial dimension and physical dimensions were extracted. Step S2: Construct a deep feature fusion regression model that includes a parallel feature extraction stage and a deep interactive regression stage; in the parallel feature extraction stage, the one-dimensional spectral feature sequence is synchronously input into multiple deep feature learning networks with different structures and complementary characteristics, and outputs multiple sets of deep semantic feature vectors with completely unified feature dimensions. Step S3: In the deep interactive regression stage, the multiple sets of deep semantic feature vectors are connected along the feature channel direction to form a high-dimensional multi-view fusion feature vector; the high-dimensional multi-view fusion feature vector is fed into the deep nonlinear interactive regression network to eliminate feature redundancy and output the initial SPAD prediction value of the rice leaf to be tested. Step S4: Construct a semi-supervised transfer learning framework based on the alignment of feature space statistics distribution, using a model pre-trained on source domain data as the base network; utilize labeled samples in the target domain, and perform gradient optimization on multiple loss terms through a dynamic weight balancing mechanism to complete the adaptive generalization of the model from the controllable source domain to the complex field target domain. Step S5: Obtain real-time low-altitude remote sensing hyperspectral data of rice in the field to be tested, and output the SPAD quantitative prediction results of rice in the field to be tested through an adaptive deep feature fusion regression model.

2. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 1, characterized in that: The method for extracting the one-dimensional spectral feature sequence specifically includes the following steps: Dark field background data and standard diffuse reflectance reference plate data of the original hyperspectral intensity of the rice under test in the source domain and target domain were collected respectively. Through linear correction calculation based on the dark-white dual reference ratio, the digital quantization value of the original image intensity was converted into absolute physical reflectance data to eliminate the interference of dark current of imaging sensor and uneven illumination of heterogeneous ambient light sources. Based on the signal-to-noise ratio response curves of the spectral sensor at different characteristic wavelengths, the edge bands located at both ends of the light wave, with a signal-to-noise ratio below the preset threshold and severely polluted by environmental noise are adaptively cut off, and only the effective characteristic spectral ranges that are highly correlated with crop nitrogen and chlorophyll absorption are retained. For the spectral data after removing the edge bands, a sliding window smoothing filter algorithm is first used to eliminate high-frequency random instrument noise and spikes. Then, a standard normal variable transformation algorithm is used to perform full-band centering and scaling standard processing on the spectral sequences of each sample to eliminate baseline shift and tilt caused by multi-angle surface scattering from the low-altitude remote sensing blades in the field. The output is a standard one-dimensional spectral feature sequence with homogeneous feature dimensions and unified physical dimensions.

3. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 2, characterized in that: After the multiple sets of deep semantic feature vectors are connected along the feature channel direction, a high-dimensional multi-view fusion feature vector with a second feature dimension that is three times the number of the first feature dimension is generated. The deep nonlinear interactive regression network is specifically a non-fully connected convolutional regression head composed of multiple layers of one-dimensional convolutional blocks that are progressively reduced in dimensionality. It includes: a feature reshaping layer for converting the high-dimensional multi-view fused feature vector into a sequence mapping tensor; at least two layers of one-dimensional convolutional and average pooling components that progressively reduce the dimensionality of feature channels and perform spatial aggregation; and an output projection layer connected to the end of the network for mapping interactive semantics into scalar SPAD regression values.

4. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 3, characterized in that: The specific method for outputting the initial SPAD prediction value of the rice leaf to be tested in step S3 includes the following steps: The high-dimensional multi-view fusion feature vector generated by splicing is transformed from a one-dimensional physical vector structure into a two-dimensional sequence feature tensor suitable for convolution kernel sliding through a feature reshaping operator, so that the feature channels extracted by heterogeneous units have spatial and sequence-related topology. The two-dimensional sequence feature tensor is continuously input into at least two cascaded one-dimensional convolutional and pooling layers. By utilizing the local sliding mapping of the one-dimensional convolutional kernel on the feature channel dimension, the dimension of the feature channel is reduced step by step, and the deep nonlinear cross-correlation between the representations of different perceptual units is captured. An adaptive average pooling operation is performed on the interaction feature tensor after layer-by-layer compression to aggregate and compress the non-fixed length sequence dimension into a globally compressed feature vector with a unique length, which greatly eliminates redundant background noise generated by multi-source feature splicing. The global compressed feature vector is progressively characterized and refined through a multi-layer fully connected network with layer normalization and nonlinear activation functions. Finally, the scalar regression prediction results of the SPAD values ​​of the rice leaves and canopy are obtained by inversion through the output projection layer at the end of the network.

5. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 4, characterized in that: The adaptive generalization method from the controllable source domain to the complex field target domain in step S4 specifically includes the following steps: Labeled hyperspectral samples are extracted from laboratory source domain datasets and input into the network. The mean square error between the initial SPAD prediction value of the network and the real physical sampling label is calculated. The source domain supervised task adaptation loss is constructed to dominate the hyperspectral to crop physiological parameter mapping relationship that forms the basis of network learning. A very small number of labeled samples are extracted from the low-altitude remote sensing target domain dataset of Datian and input into the network. The mean square error between the corresponding predicted value and the field measured label is calculated. The supervised real distribution pull loss of the target domain is constructed to provide real physical space anchor point constraints for cross-domain adaptation of high-dimensional feature space. The source domain labeled samples and the target domain labeled samples are mixed and input into the network, and the activation feature matrices of the two types of samples at the end of the deep feature projection layer are extracted simultaneously. The empirical covariance matrices of the source domain deep feature matrix and the target domain deep feature matrix are calculated respectively, and the second-order statistical distribution difference alignment loss is constructed by calculating the nonlinear norm statistical distance between the two empirical covariance matrices. During the backpropagation optimization process, the adaptive weight balancing coefficient is generated in real time based on the ratio of the numerical dimensions of the supervised task adaptation loss and the distribution difference alignment loss calculated in real time. Combined with the supervised task adaptation loss and the real distribution pull loss, end-to-end joint gradient descent is performed on the overall network weights to collaboratively achieve cross-domain statistical distribution alignment and refined regression adaptation in the complex background of Datian.

6. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 5, characterized in that: The multiple deep feature learning networks with distinct structures in step S2 specifically include: Local waveform feature sensing unit: a one-dimensional convolutional neural network module with multi-scale receptive fields, which contains at least two sets of cascaded deep convolution operators, batch normalization operators and nonlinear activation operators, used to focus on capturing local subtle features of chlorophyll at characteristic wavelengths. Sequence temporal dependency modeling unit: This is a recurrent gated network module with sequence context memory function. It contains a multi-layer stacked long short-term memory network structure, which is used to learn the dependencies between bands point by point according to wavelength growth order. Global Context Association Aware Unit: This is a graph or sequence modeling module based on position encoding and attention mechanisms. It contains a multi-head self-attention encoding structure, which is used to directly calculate the cross-self-attention weights between any two feature wavelength positions in the input sequence. Each of the three network units has a feature projection layer attached to its end, which maps the output representation uniformly to a vector space with a preset first feature dimension.

7. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 6, characterized in that: After outputting the attention representation sequence, the global context association perception unit further compresses the multi-band features into a fixed-length feature representation through an adaptive sequence aggregation operator, and then generates a semantically refined deep semantic feature vector of the first feature dimension through a fully connected network with distribution normalization.

8. The non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 7, characterized in that: In step S4, the training control for completing the adaptive generalization of the model from the controllable source domain to the complex field target domain includes multi-stage progressive optimization logic, specifically including: Task warm-up phase: Mask the feature distribution alignment loss and the supervised loss of the target domain, and optimize the deep feature fusion regression model only using laboratory source domain hyperspectral samples to establish a robust initial state mapping from physical spectra to crop physiological parameters; Joint Adaptive Phase: The three functional loss terms are enabled simultaneously, and the alignment strength coefficient is dynamically amplified with each training iteration cycle, driving the deep network to maintain the ability to regress and discriminate crop physiological parameters while eliminating the spectral domain distribution shift caused by the low-altitude remote sensing background. Distribution fine-tuning stage: Freeze the parameters of each deep feature learning network in the previous stage, and use only a small number of real physical sampling labels in the target domain to fine-tune the deep nonlinear interactive regression network at the end with a low learning rate to eliminate the residual system error caused by the mixed pixels of the UAV.

9. A non-destructive testing method for the SPAD value of rice based on hyperspectral imaging technology according to claim 8, characterized in that: The second-order statistical distribution difference alignment loss specifically includes deep correlation alignment loss, which constructs the empirical covariance matrix of the source domain activation feature matrix and the empirical covariance matrix of the target domain activation feature matrix, and uses nonlinear matrix norm to calculate the statistical discrete distribution difference between the two in the same high-dimensional mapping space, thereby achieving forced alignment of cross-environment spectral feature expressions.

10. A non-destructive testing system for rice SPAD value based on hyperspectral imaging technology, implementing the non-destructive testing method for rice SPAD value based on hyperspectral imaging technology as described in any one of claims 1 to 9, characterized in that: include: Multi-source hyperspectral sensing and data link module: It includes a ground-based stationary hyperspectral imaging acquisition end and an airborne low-altitude hyperspectral remote sensing end for field, which are used to acquire the original spectral images of the rice to be tested and transmit them to the cloud or local computing core. Standard preprocessing and feature extraction module: Coupled downstream of the data link module, it is used to perform adaptive background radiation compensation, automatic noise band removal and baseline smoothing on the input spectral image online, and output a standardized feature sequence. Multi-feature collaborative extraction and semi-supervised cross-domain processing module: It has a built-in high-performance computing processor and non-volatile storage medium. The storage medium has a deep feature fusion and transfer processing program embedded in it. When the processor executes the program, it calls the complementary multi-channel neural network in parallel to extract deep semantic features and connects them. At the same time, it uses a dynamic adaptive transfer algorithm to align and generalize the indoor pre-trained model to the complex field test environment in real time. The task representation output and intelligent decision-making module is used to receive the quantitative prediction values ​​generated by the multi-feature collaborative extraction and semi-supervised cross-domain processing module, dynamically generate the electronic map of the spatial distribution of rice SPAD values ​​and the prescription map for nitrogen nutrient deficiency diagnosis in the corresponding spatial coordinate system, and instruct or assist in the intelligent variable fertilization operation in the field.

Images