Crop leaf area index retrieval method based on multi-angle gap fraction and physical constraint

Abstract

A crop leaf area index inversion method based on multi-angle gap fraction and physical constraint belongs to the field of agricultural intelligent remote sensing. The method projects the plot boundary to the overlapping original image of the standard vertical flight line of the unmanned aerial vehicle, extracts the multi-angle region of interest of the same plot, respectively aggregates two types of effective plant area indexes by the segmentation gap fraction and the spectral gap fraction proxy, and inputs the crown height and the vegetation index into the shared feature encoder; the real leaf area index and the structure conversion coefficient are output by two prediction heads, the effective plant area index is reconstructed through the physical reconstruction layer without training parameters; the model is trained by using the real leaf area index to supervise the anchor point loss and the physical reconstruction loss with the spectral proxy path effective plant area index as the anchor point, so that the sample without the real leaf area index label participates in the parameter update. The present application can reduce the inversion deviation caused by the optical saturation of the canopy, the leaf aggregation and the ear shielding, and reduce the destructive sampling label requirement.

Description

Technical Field

[0001] This invention belongs to the field of agricultural intelligent remote sensing technology, specifically involving a technical solution for retrieving the true leaf area index of crops using multi-angle observation by UAVs, gap rate inversion, effective plant area index, canopy height, physical consistency neural network, and three-dimensional radiative transfer simulation prior inversion. Background Technology

[0002] Leaf area index (LAI) is a crucial parameter characterizing crop canopy structure, growth status, canopy light utilization, and yield formation. Existing UAV remote sensing inversion methods primarily rely on empirical regression models based on single-angle orthophotos, spectral reflectance, or vegetation indices. However, during dense canopy or panicle development stages, vegetation indices tend to saturate, becoming insensitive to changes in lower and middle leaf layers and shading by non-leaf organs, leading to increased bias in the estimation of the true LAI.

[0003] Indirect optical methods, based on canopy gap ratio, typically yield quantities closer to the effective plant area index (PAI) than the true leaf area index (LAI). The gap ratio, as used in this invention, refers to the proportion of the target canopy not obscured by leaves, stems, spikes, or other canopy elements along the corresponding observation direction, and also includes continuous proxy quantities constructed based on spectral characteristics. Leaf aggregation, differences in plant type, stem occlusion, spike occlusion, background variations, and differences in phenological stages can all cause dynamic deviations between the PAI and the LAI.

[0004] Existing multi-angle remote sensing methods can increase the geometric information of observations, but often require customized inclined flight paths, circular flight paths, or additional observation equipment; existing three-dimensional radiative transfer models have strong physical significance, but in real-world scenarios, they suffer from large parameter spaces, ill-conditioned inversion, and difficulties in transferring simulations to real-world measurements; existing deep learning methods have nonlinear fitting capabilities, but are often used as black-box regressors, making it difficult to prove that their output conforms to the physical relationship between leaf area index and gap ratio.

[0005] Therefore, there is an urgent need for a technical solution that can make full use of multi-angle observations under conventional UAV flight paths and connect the effective plant area index with the true leaf area index through clear physical constraints. Summary of the Invention

[0006] This invention aims to solve the following technical problems: optical saturation of dense canopy leads to spectral features being insensitive to the true leaf area index; there is a dynamic deviation between the effective plant area index and the true leaf area index caused by leaf aggregation and non-leaf organ occlusion; customized multi-angle image acquisition increases operational costs; the true leaf area index label relies on destructive sampling, resulting in a limited number of samples; and pure data-driven models have insufficient generalization and lack physical interpretability across crop, year, and planting density scenarios.

[0007] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: A crop leaf area index inversion method based on multi-angle gap ratio and physical constraints is proposed. This method acquires target area images, camera intrinsic and extrinsic parameters, plot boundaries, and canopy height collected by a UAV via a standard vertical flight path. The plot boundaries are back-projected onto multiple overlapping original images to extract multi-angle regions of interest (ROI) images of the same plot and determine the observation direction. For each observation direction, the segmentation gap ratio is obtained based on the ratio of gap pixels to the total number of effective pixels, and the spectral gap ratio proxy is obtained based on the complement of the normalized vegetation index. The gap ratios from multiple observation directions are aggregated to obtain the effective plant area index of the segmented path and the effective plant area index of the spectral proxy path. The effective plant area index of the segmented path, the effective plant area index of the spectral proxy path, the canopy height, and the vegetation index are input into a shared feature encoder to obtain a shared latent vector; the shared latent vector is input into a first prediction head and a second prediction head respectively to output a predicted true leaf area index and a predicted structure conversion coefficient; through a physical reconstruction layer without training parameters, the predicted true leaf area index is divided by the predicted structure conversion coefficient to obtain the reconstructed effective plant area index; During training, the supervised anchor loss is calculated for samples with true leaf area index labels, and the physical reconstruction loss is calculated for samples with effective plant area index of spectral proxy paths, so that unlabeled samples can participate in training through the physical reconstruction loss; the true leaf area index of the target region is output using the trained model.

[0008] Furthermore, the back projection includes projecting the three-dimensional coordinates of the boundary points of the land parcel area onto the distortion-free original image according to the pinhole camera model, calculating the observation zenith angle and observation azimuth angle based on the spatial vector from the camera center to the geometric center of the land parcel, and retaining only the observation directions where the projected polygon is located within the effective range of the image and the area is greater than a preset threshold.

[0009] Furthermore, the segmentation gap ratio is obtained by determining the gap pixels for multispectral region of interest images through an adaptive threshold; the spectral gap ratio proxy is a subtraction of the normalized vegetation index.

[0010] Moreover, the shared feature encoder is a feature labeler Transformer, which linearly maps each input feature to an embedding label and adds a classification label. After processing by a multi-head self-attention module, it outputs a shared latent vector.

[0011] Moreover, the shared feature encoder is a multilayer perceptron encoder, which outputs a shared latent vector through multiple linear layers and activation functions.

[0012] Moreover, both the first and second prediction heads are multilayer perceptrons. The output layer of the first prediction head uses the Softplus function to ensure that the predicted true leaf area index is positive, while the output layer of the second prediction head uses the mapping Sigmoid function to limit the prediction structure conversion coefficient within a preset biologically reasonable range. The prediction structure conversion coefficient characterizes the structural deviation between the effective plant area index and the true leaf area index caused by at least one of leaf aggregation, stem shading, ear shading, background changes, and phenological changes.

[0013] Furthermore, a three-dimensional canopy scene is constructed and solar geometry and atmospheric parameters are set to generate simulation samples for pre-training; then, the pre-trained model is fine-tuned using measured samples, wherein measured samples without real leaf area index labels participate in training through the physical reconstruction loss.

[0014] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints as described above.

[0015] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints as described above.

[0016] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints as described above.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects.

[0018] First, this invention utilizes the different observation directions naturally formed by the camera's field of view in a standard vertical flight path to obtain multi-angle observations of the same plot of land without the need for additional route circling, thus improving data acquisition efficiency.

[0019] Second, the present invention simultaneously constructs segmented gap ratio and spectral gap ratio proxy quantities, and aggregates them to obtain two types of effective plant area indices, so that geometric gap information and continuous spectral proxy information complement each other, thereby alleviating the instability of single gap ratio estimation under shade, mixed pixels, saturated canopy or spike shading.

[0020] Third, this invention explicitly outputs the true leaf area index and structural conversion coefficient through a dual-head prediction structure, and reconstructs the effective plant area index through a physical reconstruction layer without training parameters, so that the neural network output is constrained by interpretable physical relationships.

[0021] Fourth, this invention introduces three-dimensional radiative transfer simulation pre-training and measured few-label fine-tuning, which improves the generalization ability across crops, years and different planting densities while reducing the need for destructive sampling.

[0022] Fifth, the experimental results show that the overall coefficient of determination of the effective plant area index estimated by multi-angle multispectral images is higher than that of RGB images; the model using the feature labeler Transformer, dual-head prediction and physical reconstruction layer achieves higher accuracy in the inversion of true leaf area index on mixed rice and sorghum data, and shows better robustness in small sample, cross-year extrapolation and feature ablation experiments.

[0023] The present invention is simple and convenient to implement, highly practical, and solves the problems of low practicality and inconvenience in actual application of related technologies. It can improve user experience and has significant market value. Attached Figure Description

[0024] Figure 1 This is a system structure block diagram according to an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of the inversion method based on multi-angle gap ratio and physical constraints according to an embodiment of the present invention.

[0026] Figure 3 This is a network structure diagram of the P2LAI-FTT model and the P2LAI-MLP model with physical consistency constraints according to an embodiment of the present invention.

[0027] Figure 4 This is a flowchart illustrating the simulation pre-training and experimental fine-tuning process in an embodiment of the present invention.

[0028] Figure 5 This is a comparison chart of the multi-angle effective plant area index estimation results of an embodiment of the present invention.

[0029] Figure 6 This is a few-sample performance trajectory diagram of an embodiment of the present invention.

[0030] Figure 7 This is a performance graph of cross-year extrapolation in an embodiment of the present invention.

[0031] Figure 8 This is a diagram showing the ablation results of a specific embodiment of the present invention.

[0032] Figure 9 The diagram illustrates the feature contributions of an embodiment of the present invention. Detailed Implementation

[0033] The following will further explain the concept, specific structure and technical effects of the present invention in conjunction with the accompanying drawings and embodiments, so as to fully understand the purpose, features and effects of the present invention.

[0034] Example 1: This invention discloses a crop leaf area index (LAI) inversion method based on multi-angle gap ratio and physical constraints. The method backprojects the plot boundary onto the overlapping original imagery of a standard UAV vertical flight path, extracting multi-angle regions of interest (ROIs) for the same plot. Two types of effective plant area indices are obtained by aggregating segmented gap ratio and spectral gap ratio surrogate values, respectively, and input together with canopy height and vegetation index into a shared feature encoder. The true LAI and structural transformation coefficients are output through two prediction heads, and the effective LAI is reconstructed through a physical reconstruction layer without training parameters. The model is trained using the true LAI to supervise anchor point loss and the physical reconstruction loss using the effective LAI as anchor points for spectral surrogate paths, allowing samples without true LAI labels to participate in parameter updates. This invention can mitigate inversion biases caused by canopy optical saturation, leaf aggregation, and panicle occlusion, and reduces the need for destructive sampling labels.

[0035] See Figure 2 This invention provides a method for inverting crop leaf area index based on multi-angle gap ratio and physical constraints. S1, acquires images of the target area, camera internal and external parameters, land area boundaries, and canopy height collected by the UAV along a standard vertical flight path; S2, based on the camera's internal and external parameters, back-project the boundary of the land area to multiple original images, extract multi-angle region of interest images of the same land area, and determine the corresponding observation direction, i.e., the observation zenith angle of each multi-angle region of interest image; S3, For the multi-angle region of interest image, calculate the segmentation gap ratio and spectral gap ratio proxy amount respectively; In one possible embodiment, for the first i For each region of interest image in each observation direction, the segmentation gap ratio is obtained based on the ratio of the number of gap pixels to the total number of effective pixels. P seg ( θ i ), and the spectral gap rate surrogate quantity P obtained from the supplement of the normalized vegetation index. VI ( θ i ); S4, based on the Beer-Lambert relation, aggregates the gap rate of multiple observation directions to obtain the effective plant area index of the segmented path and the effective plant area index of the spectral proxy path, respectively; In one possible implementation, aggregation is based on the Bill-Lambert relation.n The effective vegetation area index (PAIe) for each effective observation direction was obtained. seg And the effective plant area index PAIe of the spectral proxy path VI Among them, the use of the first i Blade projection coefficients for each observation direction G ( θ i ); S5, input the effective plant area index of the segmented path, the effective plant area index of the spectral proxy path, the canopy height and the vegetation index into the shared feature encoder, use the first prediction head to output the predicted true leaf area index, and use the second prediction head to output the predicted structure conversion coefficient. In one possible embodiment, the PAIe will be included seg The PAIe VI The input feature set, including the canopy height and at least one vegetation index, is input into a shared feature encoder to obtain a shared latent vector. This shared latent vector is then input into a first prediction head and a second prediction head, respectively. The first prediction head outputs a predicted true leaf area index greater than zero. The predicted structure conversion coefficients output by the second prediction head are limited to a preset biologically reasonable range. .

[0036] S6. Based on the predicted true leaf area index and the predicted structure conversion coefficient, the reconstructed effective plant area index is obtained through a physical reconstruction layer without training parameters. In one possible embodiment, a physical reconstruction layer without trainable parameters is used, according to Obtain the reconstructed effective plant area index .

[0037] S7. During the model training phase, the supervision anchor point loss is calculated based on the true leaf area index label, and the effective plant area index of the spectral proxy path is used as the observation anchor point. The physical reconstruction loss is calculated based on the deviation between the reconstructed effective plant area index and the effective plant area index of the spectral proxy path, so that the measured samples without true leaf area index labels can participate in the model parameter update through the physical reconstruction loss. In one possible embodiment, the supervisory anchor loss between the predicted true leaf area index and the true leaf area index label is calculated for samples with true leaf area index labels, and the physical reconstruction loss between the reconstructed effective plant area index and the spectral proxy path effective plant area index is calculated for samples with the spectral proxy path effective plant area index, thereby enabling samples without true leaf area index labels to participate in model parameter updates through the physical reconstruction loss.

[0038] S8 uses the trained model to output the true leaf area index of the target region.

[0039] The above steps form a mutually coupled technical closed loop: the effective plant area index of the segmented path and the effective plant area index of the spectral proxy path are used together as the structural input of the shared feature encoder, wherein the effective plant area index of the spectral proxy path is further used as the observation anchor point of the physical reconstruction loss; the outputs of the first prediction head and the second prediction head are recoupled through a deterministic physical reconstruction layer without training parameters, so that samples with true leaf area index labels are supervised and anchored, and samples without true leaf area index labels but with UAV observation features can participate in model training through physical reconstruction loss.

[0040] Preferably, in step S2, the corner points of the plot and the image pixel coordinates satisfy the pinhole camera projection relationship:

[0041] in, s As a scale factor, K For the camera intrinsic parameter matrix, R Let be a rotation matrix. t It is a translation vector. X , Y , Z Represents the coordinates of the boundary points of the land parcel in the three-dimensional world coordinate system: X This represents the horizontal coordinate of the point in the world coordinate system. Y This represents the vertical coordinate of the point in the world coordinate system. Z This represents the elevation or height coordinates of the point in the world coordinate system. u , v This represents the two-dimensional pixel coordinates of the 3D point projected onto the camera image plane: u The column coordinates in the image; v These are the row coordinates in the image.

[0042] Preferably, in step S3, the segmentation gap ratio is calculated based on the vegetation and gap pixel classification results; the spectral gap ratio surrogate is constructed based on the normalized vegetation index.

[0043] in, θ i For the i-th observation direction, VI norm ( θ i ) is the first i Normalized vegetation index under each observation direction.

[0044] Preferably, in step S4, the effective plant area index satisfies:

[0045] Where n is the number of effective observation directions, P ( θ i ) represents the clearance ratio in the i-th direction. G ( θ i ) represents the blade projection coefficient.

[0046] Preferably, in steps S5 and S6, the structural transformation coefficients and the physical reconstruction layer satisfy:

[0047] Among them, PAIe recon To reconstruct the effective plant area index, To predict the true leaf area index, To predict the structural transformation coefficients.

[0048] Preferably, the shared feature encoder is a feature labeler Transformer or a multilayer perceptron. The feature labeler Transformer maps each input feature to a continuous embedding label and learns the interaction relationship between the effective plant area index of the segmentation path, the effective plant area index of the spectral proxy path, the canopy height, and the vegetation index through a multi-head self-attention module.

[0049] In one possible embodiment, the shared feature encoder is a feature labeler Transformer; the feature labeler Transformer linearly maps each input feature to a 32-dimensional continuous embedding label, adds a learnable classification label before the feature label sequence, and processes the feature label sequence using a multi-head self-attention module including a feedforward network with four attention heads, a GELU activation function, and a dropout layer with a dropout rate of 0.1, and uses the output of the classification label as the shared latent vector.

[0050] In another possible embodiment, the shared feature encoder is a multilayer perceptron encoder, which includes two linear layers with output dimensions of 64 and 32 respectively, and after the first linear layer, layer normalization, SiLU activation function and dropout layer with dropout rate of 0.1 are set in sequence; the multilayer perceptron encoder, together with the first prediction head, the second prediction head and the physical reconstruction layer, forms the same dual-output physical reconstruction constraint as in claim 1.

[0051] Preferably, both the first prediction head and the second prediction head are two-layer multilayer perceptrons with a hidden layer dimension of 16 and employ the SiLU activation function; the output layer of the first prediction head employs the Softplus function; the output layer of the second prediction head employs the mapped Sigmoid function, limiting the prediction structure conversion coefficient to within [0.5, 1.5]; the prediction structure conversion coefficient characterizes the structural deviation between the effective plant area index and the true leaf area index caused by at least one of leaf aggregation, stem occlusion, ear occlusion, background changes, and phenological changes.

[0052] Preferably, before model training, a three-dimensional radiative transfer model is used to generate simulation samples containing crop leaves, stems, ears, background, planting structure, and phenological changes to pre-train the model; then, measured samples containing a small number of real leaf area index labels are used for fine-tuning; and unsupervised physical reconstruction loss is calculated for measured samples without real leaf area index labels but with UAV characteristics and observed effective plant area index.

[0053] Example 2: like Figure 1 As shown, this embodiment of the invention provides a crop leaf area index (LAI) inversion system, including an image and parameter acquisition module, a multi-angle region of interest (ROI) image extraction module, a gap ratio calculation module, an effective plant area index aggregation module, a shared feature encoding module, a dual-head prediction module, a physical reconstruction module, and a training or inference module. Each module works in sequence and collaboratively to output a true LAI prediction result based on multi-angle images of the target area, canopy height, and vegetation indices. Specifically: The image and parameter acquisition module is configured to acquire standard vertical flight path images of the UAV, camera intrinsic and extrinsic parameters, land area boundaries, and canopy height; The multi-angle region of interest backprojection module is configured to extract multi-angle region of interest images of the same plot from overlapping original images and determine the observation direction; The dual-path effective plant area index generation module is configured to generate the segmented path effective plant area index and the spectral proxy path effective plant area index based on the segmentation gap rate and the spectral gap rate proxy amount, respectively. The shared feature encoding and dual-head prediction module is configured to generate a shared latent vector based on two types of effective plant area indices, canopy height, and vegetation indices, and output the predicted true leaf area index and the predicted structure transformation coefficient. A physical reconstruction module that does not contain trainable parameters is configured to obtain a reconstructed effective plant area index based on the ratio of the predicted true leaf area index to the predicted structure transformation coefficient. The training and inference module is configured to calculate the supervision anchor loss based on the true leaf area index label and the physical reconstruction loss based on the reconstructed effective plant area index and the spectral proxy path effective plant area index.

[0054] Example 3: See Figure 2 This invention provides a method for inverting crop leaf area index based on multi-angle gap ratio and physical constraints. A preferred implementation includes the following specific steps: Step S1: Acquire images of the target area, camera intrinsic and extrinsic parameters, plot boundaries, and canopy height. Step S1 specifically includes the following sub-steps: S1.1, A drone equipped with an RGB camera and a multispectral camera performs standard vertical flight path data acquisition on the target crop plot. In this embodiment, it is preferably recommended that the flight altitude be set to 12m or 30m, with a lateral overlap greater than 70% and a longitudinal overlap greater than 80%, and all images be acquired from the nadir direction. The multispectral camera may include green, red, red-edge, and near-infrared bands, for example, with center wavelengths of 560nm, 650nm, 730nm, and 860nm, respectively.

[0055] S1.2, Photogrammetric processing is performed on the acquired images to obtain camera intrinsic parameters, camera extrinsic parameters, distortion-free raw images, digital surface models, orthophotos, and point clouds. Multispectral images undergo radiometric correction, converting digital quantization values ​​into surface reflectance.

[0056] S1.3 Obtain the boundaries of the plot area. The plot boundaries can be provided by field surveys, manual delineation from orthophotos, geographic information vector files, or experimental plot design documents, and should be unified to the coordinate system of the photogrammetric results.

[0057] S1.4, Calculate the canopy height. In this embodiment, the canopy height model CHM is obtained by subtracting the digital surface model DSM from the bare ground digital terrain model DTM: CHM=DSM-DTM The CHM pixel values ​​within each plot area are statistically analyzed as canopy height. H In this embodiment, the 99th percentile is preferably used to reduce the impact of point cloud noise, outliers, or local occlusion on height features. In practice, in scenarios with sparse point clouds, short canopies, or many outliers, the value can also be determined within the range of 0.95 to 0.995 based on point cloud density and outlier ratio.

[0058] Step S2: Extract multi-angle region of interest images of the same plot and determine the observation direction. Step S2 specifically includes the following sub-steps: S2.1, for any plot of land, represent the four corner points or multiple boundary points of the plot area boundary as three-dimensional world coordinates. Read the camera intrinsic parameter matrix corresponding to each original image from the photogrammetric results. K Rotation matrix R Translation vector t .

[0059] S2.2, Perform back projection based on the pinhole camera model to map the land parcel boundaries onto multiple distortion-free original images:

[0060] in, s As a scale factor, K For the camera intrinsic parameter matrix, R Let be a rotation matrix. t This is the translation vector. X , Y , Z () represents the coordinates of the boundary points of the land parcel in the three-dimensional world coordinate system: X This represents the horizontal coordinate of the point in the world coordinate system. Y This represents the vertical coordinate of the point in the world coordinate system. Z This represents the elevation or height coordinates of the point in the world coordinate system. u , v () represents the two-dimensional pixel coordinates of the 3D point projected onto the camera image plane: u The column coordinates in the image; v These are the row coordinates in the image.

[0061] When the projected polygon falls within the effective range of the image, the region of interest image enclosed by the projected polygon is cropped; when the projection exceeds the image boundary, causes severe distortion, or the plot is incomplete, the observation direction is discarded.

[0062] S2.3 Calculate the observation zenith angle and observation azimuth angle based on the spatial vector from the camera center to the geometric center of the plot, and save the image of the region of interest, plot number, observation zenith angle, observation azimuth angle, and acquisition time in association. Since the same plot is located at different positions in the camera's field of view in adjacent aerial images, the standard vertical flight path can naturally generate a multi-angle observation set.

[0063] S2.4 If the number of effective observation directions for a certain plot is insufficient, it shall be marked as a low-confidence sample, or more directions shall be obtained by increasing the overlap of flight paths, supplementing the collection, or expanding the effective image screening range.

[0064] Step S3: Calculate the surrogate values ​​for the segmentation gap ratio and the spectral gap ratio. Step S3 specifically includes the following sub-steps: S3.1, For RGB region of interest images, calculate the Color Vegetation Extraction Index (CIVE): CIVE=0.441×Red-0.811×Green+0.385×Blue+18.787 Here, Red, Green, and Blue represent the pixel values ​​of the red, green, and blue channels, respectively. The segmentation threshold is determined based on the acquisition date or lighting conditions, and non-vegetation pixels are used as gap pixels.

[0065] S3.2 For multispectral region of interest images, calculate the near-infrared vegetation reflectance related index NIRv.

[0066] In one embodiment, the following may be adopted:

[0067]

[0068] in, Near-infrared reflectance, Red light reflectance, This is the normalized differential vegetation index.

[0069] Set an adaptive threshold for each region of interest image. :

[0070] in, This represents the maximum NIRv value of the effective pixels within the current region of interest. This is the minimum NIRv value of the valid pixels within the current region of interest. When the pixel value NIRv is less than or equal to... T NIRv When a pixel is found to be in gaps, it is classified as a gap pixel. The gap segmentation rate is the ratio of the number of gap pixels to the total number of valid pixels.

[0071] S3.3, for regions of interest images with coarse spatial resolution, deep shadows, many mixed pixels, or obvious canopy optical saturation, construct a spectral gap rate proxy.

[0072] First, calculate the spatial average of the vegetation index within the region of interest; if the theoretical range of the vegetation index is not within [0,1], use maximum-minimum normalization: VI norm ( θ i )=[VI( θ i )-VI min ] / [VI max -VI min ] Among them VI ( θ i ) is the first i The spatial average of vegetation indices within the region of interest under each observation direction, VI min and VI max These are the minimum and maximum values ​​of the same vegetation index within the training sample or a preset physical effective range, respectively; when VI max =VI min At that time, VI norm ( θ i Set it to 0.

[0073] After truncating the normalized result to [0,1], the spectral gap ratio proxy is obtained according to the following formula. :

[0074] in, θ i For the first i One observation direction, VI norm ( θ i ) is the first i Normalized vegetation index under each observation direction.

[0075] S3.4, vegetation indices may include one or more of the following: Normalized Difference Vegetation Index (NDVI), Normalized Difference Red Edge Index (NDRE), Modified Soil Adjusted Vegetation Index (MSAVI), Nonlinear Vegetation Index (NLI), Near-Infrared Vegetation Reflectance Correlation Index (NIRv), Kernel Normalized Difference Vegetation Index (kNDVI), and Normalized Green-Red Difference Index (NGRDI). For ease of implementation, some preferred feature calculation methods are given below:

[0076]

[0077]

[0078]

[0079]

[0080] in, ρ G , ρ R , ρ RE and ρ NIR These represent green light, red light, red edge, and near-infrared reflectance, respectively.

[0081] Step S4: Aggregate the gap ratios from multiple observation directions to obtain two types of effective plant area indices. Step S4 specifically includes the following sub-steps: S4.1, perform boundary processing on the clearance ratio obtained in step S3 for each direction, limiting it to be no less than the parameter epsilon and no greater than 1, where epsilon is a preset decimal greater than zero, used to avoid abnormal logarithmic calculations.

[0082] S4.2 Calculate the effective plant area index for each direction based on the Beer-Lambert relationship (Beer-Lambert Law), and average the results over multiple observation directions to obtain the effective plant area index. :

[0083] in, n For the number of effective observation directions, P ( θ i ) represents the clearance ratio in the i-th direction. G ( θ i ) represents the blade projection coefficient. G ( θ i This characterizes the average projection ratio of a unit leaf area onto a plane perpendicular to the observation direction; when the leaf tilt angle follows a spherical distribution and measured leaf tilt angle data are lacking, the preferred value is... G ( θ i The parameter value is set to 0.5 to reduce additional parameter estimation and maintain consistency with commonly used Beer-Lambert inversion assumptions. For crops with known leaf tilt angle distribution, row orientation, or canopy geometry, the parameters can be calculated based on the leaf tilt angle distribution function and the observation direction. G ( θ i Alternatively, the corresponding value can be obtained through actual measurement and calibration. Users can choose either method based on their actual data conditions.

[0084] S4.3, when P ( θ i When taking the segmentation gap rate, the effective plant area index PAIe of the segmentation path is obtained. seg ;when P ( θ i When taking the spectral gap rate proxy value, the effective plant area index PAIe of the spectral proxy path is obtained. VI In the embodiments, the multispectral segmentation path performed well in the structurally complex sorghum canopy, while the NLI-based spectral surrogate path performed well in the narrow-leaved rice canopy. Therefore, subsequent models also used PAIe. segand PAIe VI And preferably PAIe VI As an observation anchor point for physical reconstruction loss.

[0085] Step S5: Input the effective plant area index, canopy height, and vegetation index into the shared feature encoder, and output the predicted true leaf area index and predicted structure transformation coefficient. like Figure 3 As shown, the inversion model with physical consistency constraints includes two optional implementation methods: the P2LAI-FTT model and the P2LAI-MLP model. Figure 3 Part (a) shows the PAIe-LAI physical consistency inversion model (P2LAI-FTT) structure, which uses a feature labeler Transformer as the shared feature encoder. Part (b) shows the PAIe-LAI physical consistency inversion model (P2LAI-MLP) structure, which uses a multilayer perceptron as the shared feature encoder. Both implementations have the same input features, dual-head prediction target, and subsequent physical reconstruction constraints, but the shared feature encoders differ: P2LAI-FTT uses a feature labeler Transformer, while P2LAI-MLP uses a multilayer perceptron. During any training or inference process, one implementation is selected for execution.

[0086] Step S5 specifically includes the following sub-steps.

[0087] S5.1, Construct the input feature set: F={PAIe VI ,PAIe seg , H ,VIs} Among them, PAIe VI PAIe is the effective plant area index for spectral proxy paths. seg The effective plant area index for path segmentation. H VIs represents the canopy height, and VIs represents one or more vegetation indices.

[0088] S5.2, perform shared feature encoding on the input feature set to obtain a shared latent feature vector.

[0089] See Figure 3 In part (a), the working process of the P2LAI-FTT model is as follows: The input feature module 100 receives the input feature set.

[0090] Feature labeler 110 performs a linear mapping on each scalar feature in the input feature set, thereby mapping the first scalar feature to the second scalar feature. i Input features x i Transform into corresponding feature labelsT i : T i = x i W i + b i in W i and b i These are the mapping weights and biases corresponding to the feature, respectively. This indicates element-wise multiplication.

[0091] The feature label sequence module 120 places the learnable classification label [CLS] at the beginning of the sequence to aggregate global correlation information between input features during Transformer feature interaction. The feature labels are arranged in a predetermined and fixed order to form the feature label sequence:

[0092] in, T 0 represents a learnable classification label [CLS]. T 1. Effective Plant Area Index (PAIe) derived from spectral proxy pathways VI generate, T 2. Effective plant area index (PAIe) based on the segmented path seg generate, T 3. Based on canopy height H generate, T 4 to T k These are generated from various vegetation indices. In the example, k =25, meaning there are 25 feature labels besides the classification label, and the input sequence length of the Transformer is 26. The embedding dimension of each feature label is 32.

[0093] The Transformer feature interaction module 130 includes two sequentially connected Transformer layers. Each Transformer layer sequentially includes a multi-head self-attention sub-layer, a first residual connection, a first normalization layer, a feedforward network sub-layer, a second residual connection, and a second normalization layer. Preferably, the multi-head self-attention sub-layer includes four attention heads, the feedforward network uses the GELU activation function, and the dropout rate of both the attention sub-layer and the feedforward network is 0.1. The feature label sequence is input into the Transformer feature interaction module 130, and after multi-head self-attention calculation, the classification label [CLS] output by the second Transformer layer is taken as the 32-dimensional shared latent feature vector.

[0094] See Figure 3 In part (b), in the P2LAI-MLP model, the process is as follows: the input feature module 200 concatenates the input feature set into an input vector according to the same feature order as the P2LAI-FTT model, and then inputs it into the MLP encoding module 210. The MLP encoding module 210 includes linear layers with hidden dimensions of 64 and 32 in sequence; after the first linear layer, layer normalization, SiLU activation function, and dropout layer with a dropout rate of 0.1 are set in sequence; after the second linear layer, layer normalization and SiLU activation function are set to output a 32-dimensional shared latent feature vector.

[0095] S5.3 performs bi-head prediction on the shared latent feature vector to obtain the predicted true leaf area index and the predicted structure transformation coefficient, respectively.

[0096] For the P2LAI-FTT model, the output of the Transformer feature interaction module 130 is fed into the LAI prediction branch 141 and the correction coefficient branch 142, respectively.

[0097] For the P2LAI-MLP model, the shared latent feature vector output by the MLP encoding module 210 is input into the LAI prediction branch 221 and the correction coefficient branch 222, respectively.

[0098] Preferably, in both implementation methods described above, the LAI prediction branch and the correction coefficient prediction branch each sequentially include a first linear layer, a SiLU activation function, and a second linear layer. The first linear layer maps the 32-dimensional shared latent feature vector to 16-dimensional hidden features, and the second linear layer maps the 16-dimensional hidden features to a 1-dimensional linear output.

[0099] For the LAI prediction branch, a Softplus function is applied to its one-dimensional linear output to obtain the predicted true leaf area index. And ensure that it is positive. For the correction coefficient prediction branch, apply the mapping Sigmoid function to its 1D linear output to obtain the predicted structure transformation coefficients. And limit it to a predetermined biologically reasonable range. This embodiment uses the predicted structure transition coefficient... The value of is limited to the range of [0.5, 1.5]. This range is used to ensure that the model can learn differences in canopy structure while avoiding extreme values ​​of the structure transformation coefficient that do not conform to agronomic common sense.

[0100] Step S6: Obtain the reconstructed effective plant area index based on the predicted true leaf area index and the predicted structure conversion coefficient. Step S6 specifically includes the following sub-steps.

[0101] S6.1 introduces a structural transformation coefficient to characterize the dynamic proportional relationship between the true leaf area index and the effective plant area index:

[0102] Wherein, PAIe is the effective plant area index, and LAI is the true leaf area index. C conv This is the structural conversion coefficient. This coefficient is used to characterize structural deviations caused by leaf aggregation, stem shading, ear shading, and phenological changes.

[0103] S6.2 will predict the true leaf area index. and predicted structural transformation coefficients Input a physical reconstruction layer without training parameters and calculate the effective plant area index for reconstruction:

[0104] Among them, PAIe recon To reconstruct the effective plant area index, To predict the true leaf area index, This is for predicting structural transformation coefficients. The physical reconstruction layer does not contain trainable parameters; its purpose is to couple the two prediction heads with a deterministic physical relationship, avoiding mathematically fit but agronomically unreasonable combinations of network outputs.

[0105] See Figure 3 In the P2LAI-FTT implementation path, the physical reconstruction and loss calculation correspond to the physical consistency loss module 150; in the P2LAI-MLP implementation path, the physical reconstruction and loss calculation correspond to the physical consistency loss module 230. Both implementation paths use the same physical relationship, the only difference being that the shared feature encoder uses either the Transformer feature interaction module or the MLP encoding module.

[0106] Step S7: Calculate the supervised anchor point loss and the unsupervised physics reconstruction loss, and complete the model training. like Figure 4 As shown, the model training adopts a combination of simulation pre-training and experimental fine-tuning: First, based on the three-dimensional radiative transfer model, simulation samples containing different crop structures, phenological stages, planting densities, light conditions and observation conditions are constructed to pre-train the inversion model; then, experimental samples are used for fine-tuning, in which samples with true leaf area index labels participate in the calculation of supervised anchor point loss, and samples without true leaf area index labels participate in training through physical reconstruction loss.

[0107] Specifically, model training may include: constructing a three-dimensional canopy scene containing crop leaves, stems, ears, and soil or water background; randomly sampling the optical properties of leaves, stems, ears, and background; and setting solar geometry, atmospheric parameters, a virtual drone sensor, and a virtual upward fisheye sensor; generating multispectral reflectance features from the virtual drone sensor, generating an effective plant area index from the virtual upward fisheye sensor, and obtaining a true leaf area index label based on the geometric surface area of ​​the leaves in the three-dimensional canopy scene to form a three-dimensional radiative transfer simulation sample; pre-training the model using the three-dimensional radiative transfer simulation sample, and then fine-tuning the pre-trained model using measured samples containing a small number of true leaf area index labels and measured samples without true leaf area index labels but possessing drone observation features and the effective plant area index of the spectral proxy path.

[0108] Step S7 specifically includes the following processing: S7.1 calculates the supervision anchor loss for samples with destructive sampling of true leaf area index labels. The supervision anchor loss can be the mean square error between the predicted true leaf area index and the measured true leaf area index: L LAI =MSE( (LAI) S7.2 calculates the unsupervised physical reconstruction loss for samples with UAV observation characteristics and observed effective plant area index (PAIe). The physical reconstruction loss can be calculated as the reconstructed effective plant area index minus the observed PAIe. VI Mean square error between: L recon =MSE(PAIe recon ,PAIe VI ) Where MSE() represents the mean squared error.

[0109] Even if a sample does not have a true leaf area index label, as long as it has drone imagery characteristics and an effective plant area index, it can still be obtained through... L recon Participate in training.

[0110] See Figure 3 The monitoring anchor point loss is determined by the predicted true leaf area index. The gap ratio reconstruction loss was calculated from the true leaf area index label and the reconstructed effective plant area index PAIe. reconThe loss is calculated from the observed effective plant area index. For the P2LAI-FTT model, the loss is summarized by module 150; for the P2LAI-MLP model, the loss is summarized by module 230. Therefore, both labeled and unlabeled samples can participate in model training through the same physical consistency constraint.

[0111] S7.3, the total loss of the model is:

[0112] When it is necessary to adjust the contributions of the two losses, non-negative weighting coefficients can also be set, but in this example, they are directly added together.

[0113] In another possible implementation, the total model loss satisfies , in, L total B represents the total loss of the model. L B is the set of samples with true leaf area index labels in the training batch; B is the set of samples in the training batch that are either with or without true leaf area index labels; j is the sample index. Let LAI be the predicted true leaf area index for the j-th sample; j PAIe is the measured true leaf area index of the j-th labeled sample. recon,j PAIe is the reconstructed effective plant area index obtained by the j-th sample through a physical reconstruction layer without training parameters. VI,j The effective plant area index of the spectral proxy path is obtained for the j-th sample based on the spectral gap ratio proxy amount. λ For example, set the positive weighting coefficients for the physical reconstruction loss. λ =1, balancing the supervisory anchor loss and the physical reconstruction loss in the loss function, ensuring that the prediction of LAI does not deviate from physical laws. The first term is the supervisory anchor loss, applicable only to set B. L The first term is the labeled sample loss; the second term is the physical reconstruction loss, calculated for samples in set B. When B... L When the set is empty, the first term is not calculated, only the second term is calculated.

[0114] S7.4, Constructing 3D radiative transfer simulation samples. In one embodiment, a LESS 3D radiative transfer model is used to construct rice and sorghum scenarios, including leaves, stems, panicles, and a soil or water background. The rice simulation sample covers 11 growth stages, and the sorghum simulation sample covers 8 growth stages; the row orientation includes 0 degrees, 45 degrees, and 90 degrees; the true leaf area index of rice covers 0.06 to 13, the canopy height covers 0.24 m to 1.8 m, and the planting density covers 16.6 plants / m². 2 Up to 432.3 plants / m 2The true leaf area index of sorghum ranges from 0.09 to 4.8, the canopy height ranges from 0.4m to 3.8m, and the planting density ranges from 5.6 plants / m². 2 Up to 16.7 plants / m 2 .

[0115] S7.5 sets the optical and environmental parameters for the simulation scene. The optical properties of leaves, stems, ears, and background are randomly sampled from a measured spectral library; the solar zenith angle ranges from 10 to 48 degrees, and the solar azimuth angle ranges from 85 to 173 degrees; atmospheric parameters are coupled through a 6S model, using the MidlatitudeSummer atmospheric profile, Continental aerosol type, and a 550 nm aerosol optical thickness of 0.05 to 0.25.

[0116] S7.6, set up virtual sensors for the simulation scene. The virtual UAV sensor observes the canopy from the nadir direction, generating a bidirectional reflectance factor of 400nm to 1000nm with a 2nm interval, and convolves it with the spectral response function of the target multispectral camera. The virtual upward fisheye sensor is set to a 165-degree field of view and isostatic projection to calculate the effective plant area index, and 10% Gaussian noise can be added to simulate the uncertainty of actual measurements. The true leaf area index is accurately calculated from the geometric surface area of ​​the leaves in the 3D scene, and the canopy height is obtained from the 99th percentile height of the canopy height model of the simulated point cloud. A total of 13,000 simulation data points are generated in this example.

[0117] S7.7, perform simulation pre-training. The model is trained for 200 epochs on three-dimensional radiative transfer simulation samples with a batch size of 256, using Adam as the optimizer, an initial learning rate of 0.001, and cosine annealing as the learning rate scheduler. The validation loss early stopping patience value is set to 15 epochs.

[0118] S7.8, fine-tuning based on real-world testing. The real-world samples are stratified by crop type, into a 70% training set and a 30% test set. Fine-tuning is performed for a maximum of 200 epochs, using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and an early-stop patience value of 15 epochs. The true leaf area index label can be used only for a portion of the samples; the remaining unlabeled samples still participate in training through physical reconstruction loss.

[0119] Step S8: Use the trained model to output the true leaf area index of the target region. Step S8 specifically includes the following sub-steps.

[0120] S8.1, Perform steps S1 to S4 on the plot to be tested to obtain PAIe. seg 、PAIe VI Canopy height H And vegetation index VIs.

[0121] S8.2 Input the above features into the trained model, perform shared feature encoding, dual-head prediction, and physical reconstruction, and output the predicted true leaf area index. Simultaneously output the predicted structural transformation coefficients. This is used to explain the sources of deviation between the effective plant area index and the true leaf area index.

[0122] S8.3 can output the plot-level true leaf area index as a spatial distribution map, plot statistics table, or crop growth diagnosis results. For scenarios with multi-temporal images, the above steps can be repeated sequentially to obtain the dynamic changes of the true leaf area index and structure conversion coefficient.

[0123] To verify the technical effectiveness of the present invention, comparative experiments were conducted on measured datasets of rice and sorghum. Evaluation indicators included the coefficient of determination R0. 2 Root mean square error (RMSE), relative root mean square error (RMSE), and consistency correlation coefficient, etc. Figure 5 As shown, in terms of effective plant area index estimation, the overall R of multi-angle multispectral imagery... 2 The value is 0.728, higher than 0.500 for RGB images. Multispectral segmentation path PAIe seg_MS Achieving R on the structurally complex sorghum canopy 2 =0.811, NLI-based spectral proxy path PAIe VI_NLI R was achieved in the canopy of fine-leaved rice. 2=0.755. This result shows that the multi-angle gap ratio can provide more stable 3D structural constraints than the single spectral index. In this embodiment, the comparison models include ElasticNet, Random Forest (RF), Support Vector Regression (SVR), Feature Label Transformer (FTT), and Multilayer Perceptron (MLP). P2LAI-FTT refers to the PAIe to LAI physical consistency inversion model with the Feature Label Transformer as the shared feature encoder; P2LAI-MLP refers to the PAIe to LAI physical consistency inversion model with the Multilayer Perceptron as the shared feature encoder. Standard FTT and P2LAI-FTT use the same input feature set, feature label order, number of feature labels, embedding dimension, number of Transformer layers, number of attention heads, activation function, and dropout rate. The difference lies in that the standard FTT only retains the shared feature encoding module and the true leaf area index (TLAI) prediction module, meaning the classification label output is directly input into a single TLAI prediction branch. The standard FTT does not include a structure transformation coefficient prediction module, does not perform a physical reconstruction step based on the predicted TLAI and structure transformation coefficients, and does not calculate physical reconstruction loss; its training loss consists solely of the supervised loss from samples with TLAI labels. The standard MLP and P2LAI-MLP use the same input feature set, feature concatenation order, number of MLP encoding layers, hidden layer dimension, layer normalization, activation function, and dropout rate. The difference is that the standard MLP only retains the MLP encoding module and the TLAI prediction module, omitting the structure transformation coefficient prediction module, the physical reconstruction module without training parameters, and the physical reconstruction loss calculation module. Its training loss also consists solely of the supervised loss from samples with TLAI labels.

[0124] As shown in Table 1, in terms of true leaf area index inversion, ElasticNet, RF, SVR, and the standard FTT and MLP without physical consistency constraints are all used as comparison models. On the overall dataset, RF achieves the highest R-value. 2 The R value is 0.785, and the RMSE is 0.592; the standard FTT's R... 2 The R value is 0.798 and the RMSE is 0.574; the preferred R value of the P2LAI-FTT in this invention is... 2 With an R² value of 0.820 and an RMSE of 0.542, P2LAI-MLP demonstrates higher overall inversion accuracy. Regarding the ablation of physical reconstruction mechanisms, P2LAI-MLP shows a higher overall R² value compared to standard MLP. 2 The overall R-value improved from 0.790 to 0.802, while the RMSE decreased from 0.585 to 0.569; P2LAI-FTT further improved the overall R-value compared to standard FTT. 2=0.820. This result shows that dual-head prediction and a physical reconstruction layer without training parameters can improve the inversion performance under different encoders.

[0125] Table 1. Benchmark Comparison of True Leaf Area Index Inversion Models

[0126] like Figure 6 As shown, under limited sample conditions, when the proportion of actual leaf area index labels is only 1%, the R-values ​​of ElasticNet, SVR, and RF are significantly lower. 2 The R values ​​decreased to 0.074, 0.120, and 0.258 respectively, while P2LAI-FTT and P2LAI-MLP both maintained an R value of approximately 0.558. 2 When the tag ratio is 5%, the R of P2LAI-FTT 2 The result reaches 0.655, which is close to the performance of RF at a 20% tag ratio. This result demonstrates that the present invention can reduce the need for destructive sampling tags.

[0127] like Figure 7 As shown, in the inter-year extrapolation experiment, the traditional RF model reduced the CCC to 0.400 and increased the RMSE to 1.285 under different planting density scenarios in 2025; the P2LAI-FTT maintained a lower RMSE of 0.850 and a higher CCC of 0.510 under the same scenario. For sorghum test data in 2025, the correlation coefficient RMSE of P2LAI-FTT was 0.820, which was better than the RMSE of RF (1.272).

[0128] like Figure 8 As shown, in the characteristic ablation experiment, when only the vegetation index was used, the overall R of P2LAI-FTT was... 2 The R value of P2LAI-FTT decreased from 0.809 to 0.725 after removing PAIe in the structurally complex sorghum canopy. 2 The value decreased from 0.729 to 0.627. This result indicates that three-dimensional structural features such as PAIe and canopy height are important sources of the technical effectiveness of this invention.

[0129] like Figure 9 As shown, in interpretive analysis, traditional ElasticNet contributes up to 94.4% of spectral features, while RF and MLP contribute over 60%. However, P2LAI-FTT, driven by unsupervised physical reconstruction loss, reduces the contribution of spectral features to 49.2%, while also reducing PAIe... VI and PAIe seg The total contribution increased to 44.7%. This result shows that the present invention is not a simple black-box fitting, but rather shifts the decision focus to three-dimensional structural features that are more in line with the gap rate theory and agronomic logic.

[0130] In mature sorghum scenarios, changes in ear color, ear enlargement, stem bending, and leaf senescence can lead to PAIe based on spectral proxy. VI The value is lower than the true leaf area index. In the example, the structure conversion coefficient learned by the model can be increased to 1.43, which is used to compensate for the observation bias caused by shading of non-photosynthetic panicles and canopy aggregation, indicating that the structure conversion coefficient has clear agronomical interpretative significance.

[0131] The above embodiments are only used to illustrate the technical solution of the present invention. Those skilled in the art can adjust the specific parameters according to crop type, sensor configuration, plot size, number of simulation samples and computing resources. As long as the effective plant area index is still formed by multi-angle gap ratio and constrained by the physical reconstruction relationship between the predicted true leaf area index and the structural conversion coefficient, the technical objective of the present invention can be achieved.

[0132] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0133] The following describes the crop leaf area index inversion electronic device based on multi-angle gap ratio and physical constraints provided by the present invention. The crop leaf area index inversion electronic device based on multi-angle gap ratio and physical constraints described below can be referred to in correspondence with the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints described above.

[0134] The electronic device may include a processor, a communications interface, memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can call logical instructions from the memory to execute a crop leaf area index inversion method based on multi-angle gap ratio and physical constraints, mainly including the software processing part described above.

[0135] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the software processing part of the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints provided by the above methods.

[0137] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the software processing portion of the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints provided by the above methods.

[0138] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0139] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for inverting crop leaf area index based on multi-angle gap ratio and physical constraints, characterized in that: Acquire target area images, camera intrinsic and extrinsic parameters, plot boundaries, and canopy heights collected by a standard vertical flight path of an unmanned aerial vehicle (UAV); backproject the plot boundaries onto multiple overlapping original images, extract multi-angle region of interest images of the same plot, and determine the observation direction; for each observation direction, obtain the segmentation gap rate based on the ratio of gap pixels to the total number of effective pixels, and obtain the spectral gap rate proxy based on the complement of the normalized vegetation index; aggregate the gap rates of multiple observation directions to obtain the effective vegetation area index of the segmentation path and the effective vegetation area index of the spectral proxy path; The effective plant area index of the segmented path, the effective plant area index of the spectral proxy path, the canopy height, and the vegetation index are input into a shared feature encoder to obtain a shared latent vector; the shared latent vector is input into a first prediction head and a second prediction head respectively to output a predicted true leaf area index and a predicted structure conversion coefficient; through a physical reconstruction layer without training parameters, the predicted true leaf area index is divided by the predicted structure conversion coefficient to obtain the reconstructed effective plant area index; During training, the supervised anchor loss is calculated for samples with true leaf area index labels, and the physical reconstruction loss is calculated for samples with effective plant area index of spectral proxy paths, so that unlabeled samples can participate in training through the physical reconstruction loss; the true leaf area index of the target region is output using the trained model.

2. The crop leaf area index inversion method based on multi-angle gap ratio and physical constraints according to claim 1, characterized in that: The back projection includes projecting the three-dimensional coordinates of the boundary points of the land parcel area onto the distortion-free original image according to the pinhole camera model, calculating the observation zenith angle and observation azimuth angle based on the spatial vector from the camera center to the geometric center of the land parcel, and retaining only the observation directions where the projected polygon is located within the effective range of the image and the area is greater than a preset threshold.

3. The crop leaf area index inversion method based on multi-angle gap ratio and physical constraints according to claim 1, characterized in that: The segmentation gap ratio is obtained by determining the gap pixels for multispectral region of interest images through an adaptive threshold; the spectral gap ratio proxy is a subtraction of the normalized vegetation index.

4. The crop leaf area index inversion method based on multi-angle gap ratio and physical constraints according to claim 1, characterized in that: The shared feature encoder is a feature labeler Transformer. The feature labeler Transformer linearly maps each input feature to an embedding label and adds a classification label. After processing by a multi-head self-attention module, it outputs a shared latent vector.

5. The crop leaf area index inversion method based on multi-angle gap ratio and physical constraints according to claim 1, characterized in that: The shared feature encoder is a multilayer perceptron encoder, which outputs a shared latent vector through multiple linear layers and activation functions.

6. The crop leaf area index inversion method based on multi-angle gap ratio and physical constraints according to claim 1, characterized in that: Both the first prediction head and the second prediction head are multilayer perceptrons. The output layer of the first prediction head uses the Softplus function to ensure that the predicted true leaf area index is positive. The output layer of the second prediction head uses the mapping Sigmoid function to limit the predicted structure conversion coefficient within a preset biologically reasonable range. The predicted structure conversion coefficient represents the structural deviation between the effective plant area index and the true leaf area index caused by at least one of leaf aggregation, stem shading, ear shading, background changes, and phenological changes.

7. The crop leaf area index inversion method based on multi-angle gap ratio and physical constraints according to claim 1, characterized in that: A three-dimensional canopy scene is constructed and solar geometry and atmospheric parameters are set to generate simulation samples for pre-training. Then, the pre-trained model is fine-tuned using measured samples, where measured samples without real leaf area index labels participate in training through the physical reconstruction loss.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the crop leaf area index inversion method based on multi-angle gap ratio and physical constraints as described in any one of claims 1 to 7.

Images