Winter wheat LAI and SPAD estimation method based on lightweight semi-supervised model

Abstract

The application discloses a winter wheat LAI and SPAD estimation method based on a lightweight semi-supervised model, relates to the technical field of agricultural remote sensing monitoring, and comprises the following steps: acquiring multispectral image data of a key growth period of winter wheat and performing pretreatment to generate a standardized multi-channel vegetation index image; a lightweight semi-supervised model MCVI-SANet is constructed, and a VICReg-driven semi-supervised training strategy is adopted to perform self-supervised training on the MCVI-SANet; and the multi-channel vegetation index image is input into the trained MCVI-SANet to output quantitative estimation results of winter wheat LAI and SPAD. Through the combination of a saturation perception mechanism and semi-supervised learning, estimation deviation caused by dense canopy layer vegetation index saturation and data noise is effectively eliminated, the estimation precision and generalization ability in a complex agricultural scene are improved, and meanwhile, the lightweight deployment characteristics of the model are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural remote sensing monitoring, in particular to a winter wheat LAI and SPAD estimation method based on a lightweight semi-supervised model. BACKGROUND

[0002] The leaf area index LAI and the chlorophyll relative content SPAD of winter wheat are core physiological parameters reflecting crop growth, photosynthesis efficiency and nutritional status. Precise estimation of the two parameters is of great significance for guiding water and fertilizer management, predicting yield and evaluating disaster impact. Traditional estimation methods rely on manual field measurement, such as using a canopy analyzer to measure LAI and a chlorophyll meter to measure SPAD. Although these methods have high accuracy, they are time-consuming, labor-intensive, destructive and have limited coverage, which makes it difficult to meet the needs of large-scale field dynamic monitoring.

[0003] In recent years, unmanned aerial vehicle multi-spectral remote sensing technology has become a mainstream technology for estimating field parameters due to its flexibility, efficiency and low cost. Researchers have constructed estimation models by extracting vegetation indices (such as NDVI and EVI) from multi-spectral images and combining machine learning models (such as PLSR and RFR), achieving non-contact measurement of LAI and SPAD. With the development of deep learning, CNN-based models (such as ResNet and EfficientNet) have further improved estimation accuracy by automatically extracting deep features.

[0004] However, existing technologies still face many challenges in practical applications. First, the saturation problem of vegetation indices, which makes it difficult to distinguish crop growth differences and leads to increased estimation bias. Second, the scarcity of labeled samples, which limits the performance of supervised learning models due to the difficulty of obtaining sufficient labeled samples through field measurement. Third, the limitation of model deployment, which makes it difficult to adapt to real-time processing requirements of resource-constrained platforms such as unmanned aerial vehicles due to the large number of parameters and high computational complexity of mainstream deep learning models. Fourth, domain difference interference, which leads to insufficient model generalization ability due to the differences in image feature distribution under different growth stages and lighting conditions. Existing semi-supervised learning methods have few applications in the field of agricultural remote sensing, and most of them have not been optimized for the core problem of vegetation index saturation. Direct transfer may lead to insufficient feature learning, making it difficult to balance estimation accuracy and model lightweight requirements. SUMMARY

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides a winter wheat LAI and SPAD estimation method based on a lightweight semi-supervised model, comprising:

[0007] Obtain multispectral image data of winter wheat at key growth stages and preprocess to generate standardized multi-channel vegetation index images;

[0008] Construct a lightweight semi-supervised model MCVI-SANet and use a VICReg-driven semi-supervised training strategy to perform self-supervised training on the MCVI-SANet;

[0009] Input the multi-channel vegetation index images into the trained MCVI-SANet to output quantitative estimation results of winter wheat LAI and SPAD.

[0010] In a possible implementation, the obtaining of multispectral image data of winter wheat at key growth stages and the preprocessing to generate standardized multi-channel vegetation index images include:

[0011] Use a UAV to collect UAV multispectral images of winter wheat covering the jointing stage, booting stage, heading stage, flowering stage, and grain filling stage;

[0012] Preprocess the UAV multispectral images, including radiation correction, geometric correction, and orthophoto mosaic, to generate five-band orthophotos;

[0013] Based on the sampling plot boundary cropping, obtain standardized image blocks, calculate various vegetation indices, construct multi-channel vegetation index images, and simultaneously segment to generate unannotated image blocks for self-supervised pre-training;

[0014] Use a vegetation height-guided stratified sampling strategy to divide the dataset, and divide the training set, validation set, and test set according to the preset proportion through feature normalization, K-Means clustering, and majority voting mechanism.

[0015] In a possible implementation, the use of a vegetation height-guided stratified sampling strategy to divide the dataset, and the division of the training set, validation set, and test set according to the preset proportion through feature normalization, K-Means clustering, and majority voting mechanism include:

[0016] Let the original dataset be D contains |D| samples, each sample is composed of a feature matrix of three growth parameters LAI, SPAD, and vegetation height VH, first normalize the features:

[0017]

[0018] wherein: and are the mean and standard deviation of each feature, respectively;

[0019] K-Means clustering is used to capture the distribution structure of samples in the growth parameter space. The number of cluster repetitions N is set to reduce the sensitivity to random initialization and generate a label matrix. ;

[0020] Each sample is determined by majority vote. j Final clustering labels :

[0021]

[0022] in: For indicator functions, Let be the label of the j-th sample in the i-th cluster;

[0023] In each final cluster Within the training set, stratified sampling is performed according to the preset ratio of the training set, validation set, and test set, and the number of samples in each subset is calculated to ensure that the samples are evenly distributed across the reproductive stages.

[0024] In one possible implementation, constructing the lightweight semi-supervised model MCVI-SANet and performing self-supervised training on MCVI-SANet using a VICReg-driven semi-supervised training strategy includes:

[0025] The MCVI-SANet is constructed as follows: a vegetation index saturation sensing module VI-SABlock, a lightweight backbone network, an extender, and a regressor. The VI-SABlock is used to alleviate the vegetation index saturation problem, and the lightweight backbone network is constructed based on inverted residual blocks (IRB).

[0026] A semi-supervised training strategy driven by VICReg is adopted, and the encoder and extender are first pre-trained in a self-supervised manner using unlabeled images.

[0027] Then, fine-tune the regressor on the labeled dataset and learn a general feature representation by using a triple regularization loss constraint model.

[0028] In one possible implementation, the VI-SABlock, as the first layer of the network, sequentially includes: a batch normalization (BN) layer, a feature descriptor extraction module, a feature recalibration excitation (FRE) network, and a deep spatial attention module (DSAM).

[0029] The BN layer normalizes the multi-channel vegetation index input, stabilizes the heterogeneous feature distribution, and adapts to the differences in the numerical range of different vegetation indices.

[0030] The feature descriptor extraction module combines adaptive average pooling (AAP) and standard deviation (STD) to capture multi-exponential statistical information:

[0031]

[0032]

[0033] wherein: is the mean, denotes an adaptive average pooling operation, is the input processed by the BN layer, is the pixel value of the (i, j) position of the c-th channel image, i,j is the standard deviation, is the standard deviation, H and W is the height and width of the feature map, is the numerical stability index, STD retains the standard deviation information in the saturated state;

[0034] The feature recalibration excitation (FRE) network learns the nonlinear dependence between statistical information and adaptively reweights the VI channels through 1×1 convolution and Mish activation function:

[0035]

[0036]

[0037]

[0038] wherein: is the spliced statistical feature tensor, denotes a splicing operation, is the channel attention weight tensor, denotes a feature recalibration excitation network, and denotes an activation function, and the Mish activation function avoids truncating negative values, is the feature map after attention weighting, denotes element-wise multiplication, and is the 1×1 convolution layer weight, and r is the compression ratio;

[0039] The DSAM learns fine-grained spatial weights through a 3×3 depth convolution, and realizes bidirectional modulation through a tanh activation function:

[0040]

[0041]

[0042]

[0043] wherein: is the spatial attention weight, denotes a deep convolution operation, is a corresponding deep convolution matrix, is a feature map after a deep convolution operation, is a feature map after feature expansion, denotes a feature expansion layer, is the number of channels after feature expansion.

[0044] In a possible implementation, the lightweight backbone network is stacked by a plurality of inverted residual blocks (IRBs), all of which are provided with a preset stride to maintain residual connection and reserve fine-grained canopy structure features; the backbone network inputs a first size feature map, and outputs a second size high-dimensional feature map after processing by the plurality of IRBs.

[0045] In a possible implementation, the semi-supervised training strategy driven by VICReg first performs self-supervised pre-training on the encoder and the extender by using unlabeled images, including:

[0046] In the self-supervised pre-training stage, two enhanced views are generated by randomly horizontally or vertically flipping the unlabeled images Figure X and X ;

[0047] The enhanced views are mapped into high-dimensional embeddings Z and Z' by the encoder and the extender, and a triple regularization loss is used to constrain training:

[0048]

[0049]

[0050]

[0051]

[0052] wherein: is a consistency loss function, is a variance loss function, is a covariance loss function, , , are weight coefficients corresponding to the three loss functions, is a view obtained after data enhancement of the i-th batch, denotes the square of the L2 norm, B is the batch size, and D is the embedding dimension, is a constant target value of the standard deviation, is the batch variance of the d-th dimension embedding, and Cov(Z) is the covariance matrix of the embedding;

[0053] Pre-training parameter settings, including: batch size, iteration rounds, initial learning rate, projection head and output dimension;

[0054] Remove the projection head, freeze the encoder weights, and fine-tune the regressor on the labeled dataset for fine-tuning training.

[0055] In one possible implementation, the regressor contains two fully connected layers, the first layer maps the high-dimensional feature map output by the backbone network to a low-dimensional latent space, and the second layer outputs the LAI or SPAD estimate;

[0056] During the estimation process, the multi-channel vegetation index image is enhanced in feature by VI-SABlock, input into the lightweight backbone network to extract spatial-spectral features, and then processed by adaptive average pooling and flattening, and input into the regressor to obtain the final estimation result.

[0057] In one possible implementation, the performance evaluation of the MCVI-SANet uses the following indicators:

[0058]

[0059]

[0060] Wherein: R2 is the coefficient of determination, RMSE is the root mean square error, n is the total number of samples, Yi is the measured value of the i-th sample, Yi is the corresponding predicted value, is the mean of all measured values.

[0061] In a second aspect, the embodiments of the present application provide a winter wheat LAI and SPAD estimation system based on a lightweight semi-supervised model, comprising:

[0062] A data acquisition and processing module is configured to acquire multispectral image data of winter wheat at key growth stages and pre-process the data to generate standardized multi-channel vegetation index images;

[0063] An estimation model construction module is configured to construct a lightweight semi-supervised model MCVI-SANet and use a VICReg-driven semi-supervised training strategy to perform self-supervised training on the MCVI-SANet;

[0064] An estimation result output module is configured to input the multi-channel vegetation index image into the trained MCVI-SANet to output quantitative estimation results of winter wheat LAI and SPAD.

[0065] In this embodiment, multi-band images of winter wheat during key growth stages are acquired using UAV multispectral remote sensing technology. A stratified sampling strategy is employed to divide the dataset using vegetation height information, resulting in multi-channel vegetation index images representative of the growth stages. These multi-channel vegetation index images are then input into a trained MCVI-SANet model. Based on the regression results output by the model, quantitative estimations of winter wheat LAI and SPAD are performed. By combining a saturation sensing mechanism with semi-supervised learning, estimation biases caused by dense canopy vegetation index saturation and data noise are effectively eliminated. This approach improves estimation accuracy and generalization ability in complex agricultural scenarios while maintaining the model's lightweight deployment characteristics. Attached Figure Description

[0066] Figure 1 A flowchart illustrating a method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model, provided for embodiments of this application;

[0067] Figure 2 This is a schematic diagram of a UAV multispectral data acquisition scenario provided in an embodiment of this application;

[0068] Figure 3 This is a schematic diagram of an RGB orthophoto image obtained by stitching together, provided in an embodiment of this application.

[0069] Figure 4 An architecture diagram of MCVI-SANet provided for embodiments of this application;

[0070] Figure 5 This is a schematic diagram of the structure of VI-SABlock provided in an embodiment of this application;

[0071] Figure 6 This is a schematic diagram of the structure of the FRE network provided in an embodiment of this application;

[0072] Figure 7 A schematic diagram of the structure of DSAM provided in the embodiments of this application;

[0073] Figure 8 This is a schematic diagram of the lightweight backbone network provided in the embodiments of this application;

[0074] Figure 9 A comparison chart of LAI predicted values ​​and actual values ​​of SVR based on vegetation index and texture features on the test set, provided for embodiments of this application;

[0075] Figure 10 A comparison chart of ResNet18's LAI predictions and actual values ​​on the test set, provided for embodiments of this application;

[0076] Figure 11A comparison chart of LAI prediction values ​​and actual values ​​of MCVI-SANet on the test set provided for embodiments of this application;

[0077] Figure 12 A comparison chart of SPAD predicted values ​​and actual values ​​of PLSR based on vegetation index on the test set, provided for embodiments of this application;

[0078] Figure 13 A comparison chart of SPAD predictions and actual values ​​of ResNet50 on the test set provided for embodiments of this application;

[0079] Figure 14 A comparison chart of SPAD predictions and actual values ​​of MCVI-SANet on the test set provided for embodiments of this application;

[0080] Figure 15 A schematic diagram of a winter wheat LAI and SPAD estimation system based on a lightweight semi-supervised model provided in this application embodiment. Detailed Implementation

[0081] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0082] See Figure 1 The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model provided in this embodiment includes:

[0083] S101: Acquire multispectral image data of winter wheat during key growth periods and preprocess it to generate standardized multichannel vegetation index images.

[0084] In this embodiment, the key growth periods of winter wheat include: jointing stage, booting stage, heading stage, flowering stage, and grain filling stage. The collected multispectral image data includes leaf area index (LAI), relative chlorophyll content (SPAD), and vegetation height (VH).

[0085] To acquire multispectral image data, see [link / reference]. Figure 2 A typical winter wheat experimental field was selected and divided into 330 3m×3m sampling plots, covering five key growth stages. A drone equipped with a multispectral camera was used to collect images in five bands (red, green, blue, red-edge, and near-infrared) at a height of 15m under clear, windless conditions. The LAI, SPAD, and VH of each plot were simultaneously measured, and the average values ​​from multiple measurements were taken.

[0086] Image acquisition was conducted using a DJI M300 RTK drone equipped with a RedEdge-P multispectral camera. Flight parameters were set as follows: flight altitude 15m, forward overlap 80%, lateral overlap 80%, and ground control points (GCPs) were used to ensure geometric accuracy. Data acquisition took place from 10:00 to 14:00 under clear, low-cloud, and windless conditions. The camera simultaneously acquired five spectral bands: red (R), green (G), blue (B), red-edge (RE), and near-infrared (NIR). Before each flight, a ground-based radiation calibration board was photographed for radiation correction, and a downlink optical sensor simultaneously recorded solar irradiance to correct for light variations. Ground-based measurements were performed using standardized multi-point measurements: for LAI (Light Intensity Index), an AccuPAR LP-80 canopy analyzer was used to measure at 5 points in each plot, and the arithmetic mean was taken; for SPAD (Light Intensity Index), a SPAD-502 chlorophyll meter was used to measure at 4 representative plants in each plot, and the average was taken; for VH (Light Intensity Index), a measuring tape was used to measure the height of tall, medium, and short representative plants in each plot, and the average was taken. Ground measurements were performed synchronously with the drone flight. Figure 3 This is an orthophoto from the acquired multispectral image data.

[0087] After the multispectral image data was acquired, orthophotos were cropped at 3m×3m sampling plot boundaries using ArcGIS software to obtain 192×192 pixel image blocks, generating 330 labeled samples (paired LAI, SPAD, VH values). The remaining field area was uniformly divided to generate 2700 unlabeled five-band image blocks of the same size for self-supervised pre-training. Based on the five-band images, 11 vegetation indices were calculated, including NDVI, CIgreen, DVI, EVI, OSAVI, RVI, GNDVI, SAVI, VARI, NDRE, and MCARI, constructing a multi-channel vegetation index image.

[0088] To ensure the representativeness and balanced distribution of the dataset, a stratified sampling strategy was adopted to divide the dataset: LAI, SPAD, and vegetation height features were normalized, and K-Means clustering (K=10) was used to capture the sample distribution. The sample cluster labels were determined by majority voting, and within each cluster, the training set, validation set, and test set were divided in a 9:1:1 ratio.

[0089] Specifically, in this embodiment, the original dataset is set as follows: D Include |D| Each sample consists of a feature matrix composed of three growth parameters: LAI, SPAD, and vegetation height (VH). First, the features are normalized:

[0090]

[0091] in: and These are the mean and standard deviation for each feature.

[0092] K-Means clustering was used ( K =10) Capture the distribution structure of samples in the growth parameter space. To reduce the sensitivity of random initialization, clustering is repeated N>5 times to generate a label matrix. .

[0093] Each sample is determined by majority vote. j Final clustering labels :

[0094]

[0095] in: For indicator functions, Let be the label of the j-th sample in the i-th cluster;

[0096] In each final cluster Within the training set, stratified sampling is performed according to the preset ratio of the training set, validation set, and test set, and the number of samples in each subset is calculated to ensure that the samples are evenly distributed across the reproductive stages.

[0097] S102, construct a lightweight semi-supervised model MCVI-SANet and perform self-supervised training on the MCVI-SANet using a VICReg-driven semi-supervised training strategy.

[0098] See Figure 4 The MCVI-SANet core constructed in this embodiment includes a vegetation index saturation sensing module VI-SABlock, a lightweight backbone network, an extender, and a regressor.

[0099] See Figure 5 The VI-SABlock, serving as the first layer of the network, sequentially includes a batch normalization (BN) layer, a feature descriptor extraction module, a feature recalibration activation (FRE) network, and a deep spatial attention module (DSAM). VI-SABlock integrates the mean and standard deviation features of the vegetation index to alleviate the saturation problem; the backbone network is constructed based on inverted residual blocks to control the model's lightweight nature.

[0100] The BN layer normalizes the multi-channel vegetation index input, stabilizes the heterogeneous feature distribution, and adapts to the differences in the numerical range of different vegetation indices.

[0101] The feature descriptor extraction module combines adaptive average pooling (AAP) and standard deviation (STD) to capture multi-exponential statistical information:

[0102]

[0103]

[0104] in: The mean, This indicates an adaptive average pooling operation. The input is processed by the BN layer. For the c-th channel image ( i,j The pixel value at position ) Standard deviation H and W For the feature map height and width, STD is a numerical stability index that preserves the standard deviation information under saturation conditions.

[0105] See Figure 6 The Feature Recalibration Activated FRE network learns the nonlinear dependencies between statistical information and adaptively reweights the VI channels through 1×1 convolutions and the Mish activation function.

[0106]

[0107]

[0108]

[0109] in: The concatenated statistical feature tensor This indicates a splicing operation. For channel attention weight tensors, This indicates the feature recalibration excitation network. and This represents the activation function. The Mish activation function avoids truncating negative values. This is the feature map after attention weighting. This indicates element-wise multiplication. and is the weight of a 1×1 convolutional layer, and r is the compression ratio.

[0110] See Figure 7 The DSAM uses 3×3 deep convolution to learn fine-grained spatial weights and achieves bidirectional modulation through the tanh activation function.

[0111]

[0112]

[0113]

[0114] in: Spatial attention weights, This represents a depthwise convolution operation. For the corresponding depth convolution matrix, The feature map has undergone depthwise convolution. The feature map is obtained after feature expansion. Indicates the feature expansion layer. The number of channels after feature expansion.

[0115] See Figure 8 The lightweight backbone network is composed of multiple inverted residual blocks (IRBs) stacked together. All IRBs are set with a preset step size to maintain residual connections and preserve fine-grained canopy structure features. The backbone network takes a first-size feature map as input, and after processing by multiple IRBs, outputs a high-dimensional feature map of a second size.

[0116] In this embodiment, the backbone network consists of 12 stacked inverted residual blocks (IRBs). All IRBs are set to a stride of 1 to maintain residual connectivity and preserve fine-grained canopy structure features. The IRB parameters are configured as follows: the first 4 IRBs have an expansion factor t=1 and an output channel number c=64; the last 8 IRBs have an expansion factor t=1 and an output channel number c=96. The backbone network input is a 32×96×96 feature map. After processing by the 12 IRBs, it outputs a 96×96 high-dimensional feature map. The total number of parameters is controlled within 0.10M, meeting the requirements for lightweight deployment.

[0117] The regressor consists of two fully connected layers. The first layer maps the 96-dimensional features output by the backbone network to a 32-dimensional latent space, and the second layer outputs the LAI or SPAD estimate. During the estimation process, the multi-channel vegetation index image is enhanced with VI-SABlock features, then input into a lightweight backbone network to extract spatial spectral features. These features are then processed by adaptive average pooling and flattening before being input into the regressor to obtain the final estimation result.

[0118] The model outputs LAI estimation accuracy of R²=0.8123 and RMSE=0.4796, SPAD estimation accuracy of R²=0.6846 and RMSE=2.4222, model parameters of 0.10M, model size of 0.46MB, single-sample inference time of 17.05ms, and can be deployed on resource-constrained platforms such as drones.

[0119] After MCVI-SANet is constructed, semi-supervised model training is performed. Specifically, this embodiment adopts a VICReg-driven semi-supervised training strategy. First, the encoder and expander are pre-trained in self-supervised mode using unlabeled images. Then, the regressor is fine-tuned on labeled datasets, and the model learns general feature representations by constraining the model through triple regularization loss.

[0120] In this embodiment, the VICReg-driven semi-supervised training strategy includes:

[0121] Self-supervised pre-training phase: Data augmentation is performed on unlabeled images by random horizontal / vertical flipping and 90° rotation, generating two augmented views. X and X' ;

[0122] The augmented view is mapped to a high-dimensional embedding using an encoder and extender. Z and Z' Training is performed using triple regularization loss constraints:

[0123]

[0124]

[0125]

[0126]

[0127] in: For consistency loss function, Let Variance loss function be used. Let covariance loss function be used. , , These are the weight coefficients corresponding to the three loss functions. This is the viewpoint obtained after data augmentation for the i-th batch. Let B represent the square of the L2 norm, B be the batch size, and D be the embedding dimension. The constant target value is the standard deviation. Let Z be the batch variance of the d-th dimension embedding, and Cov(Z) be the covariance matrix of the embedding.

[0128] Pre-training parameter settings: batch size 100, 500 iterations, initial learning rate 1×10⁻⁶. -3 The projection head consists of two fully connected layers with BN and ReLU, with an output dimension of 256.

[0129] Fine-tuning phase: Remove the projector head, freeze the encoder weights, fine-tune the regressor on the labeled dataset, and estimate the initial learning rate as 5×10 using LAI. -4 SPAD estimates an initial learning rate of 5 × 10⁻⁶. -5 The batch size is 32, the iteration is 200 rounds, and an early stopping strategy is used to prevent overfitting.

[0130] S103, input the multi-channel vegetation index image into the trained MCVI-SANet to output the quantitative estimation results of winter wheat LAI and SPAD.

[0131] The preprocessed multi-channel images are input into the trained model, which outputs LAI and SPAD estimates. The model performance is evaluated using the coefficient of determination (R²) and root mean square error (RMSE), and the estimation accuracy is verified.

[0132] The performance evaluation metrics are expressed as follows:

[0133]

[0134]

[0135] in: As the coefficient of determination, Let n be the root mean square error, and n be the total number of samples. Let be the measured value of the i-th sample. For the corresponding predicted value, This is the mean of all measured values.

[0136] To verify the estimation method proposed in the above embodiments, this embodiment further compares four traditional machine learning methods and seven benchmark deep learning methods. Traditional machine learning methods include Support Vector Regression (SVR), Partial Least Squares Regression (PLSR), Random Forest Regression (RFR), and Extreme Gradient Boosting Machine (XGBoost), which combine vegetation indices (VIs) or vegetation indices with texture features (VIs+TFs). Benchmark deep learning methods include EfficientNet B0, EfficientNet V2-S, MobileNet V2, MobileNet V3, ResNet18, ResNet50, and ShuffleNet V2. Based on 10 repeated experiments, the best-performing model among the traditional machine learning and deep learning methods is selected. Then, the optimal result of the optimal model in the repeated experiments is selected, and its predicted-to-true value comparison on the LAI and SPAD test sets is compared with the experimental results of MCVI-SANet in this embodiment.

[0137] The traditional best machine learning method for LAI estimation is SVR, which combines vegetation index and texture features, while the best benchmark deep learning model is ResNet18. Figure 9 This is a comparison chart of the predicted and actual LAI values ​​of SVR based on vegetation index and texture features on the test set. Figure 10 This is a comparison chart of ResNet18's predicted LAI values ​​and actual values ​​on the test set. Figure 11 This is a comparison chart of the predicted and actual LAI values ​​of MCVI-SANet on the test set.

[0138] The traditional best machine learning method for SPAD estimation is PLSR, which uses only vegetation indices, while the best benchmark deep learning model is ResNet50. Figure 12 This is a comparison chart of the predicted and actual values ​​of SPAD based on the PLSR (Plant Growth Rate Scale) on the test set. Figure 13 This is a comparison chart of the LAI (Label Articulation Point) values ​​of ResNet50's SPAD predictions versus the actual values ​​on the test set. Figure 14 A comparison chart of SPAD predictions and actual values ​​for MCVI-SANet on the test set.

[0139] from Figures 9-14 As can be seen from this, for LAI prediction, MCVI-SANet's R... 2 The R value is 0.8619, and the RMSE is 0.4118. Compared to SVR and ResNet18, R... 2 It has the largest R value and the smallest RMSE in LAI. For SPAD predictions, MCVI-SANet's R value is... 2 The R value is 0.7756, and the RMSE is 2.0442. Compared to PLSR and ResNet50, R... 2 It has the largest R value and the smallest RMSE. Therefore, the MCVI-SANet in this embodiment has the highest R value in LAI and SPAD prediction. 2 With RMSE, it far surpasses other methods and can better predict LAI and SPAD of winter wheat.

[0140] Corresponding to the method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model provided in the embodiments of this application, this application also provides an embodiment of a system for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model.

[0141] See Figure 15 The estimation of LAI and SPAD for winter wheat based on a lightweight semi-supervised model in this embodiment includes:

[0142] The data acquisition and processing module 201 is used to acquire multispectral image data of winter wheat during key growth periods and preprocess it to generate standardized multichannel vegetation index images.

[0143] The estimation model building module 202 is used to build a lightweight semi-supervised model MCVI-SANet and perform self-supervised training on the MCVI-SANet using a VICReg-driven semi-supervised training strategy.

[0144] The estimation result output module 203 is used to input the multi-channel vegetation index image into the trained MCVI-SANet to output the quantitative estimation results of winter wheat LAI and SPAD.

[0145] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0146] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model, characterized in that, include: Acquire multispectral image data of winter wheat during key growth periods and preprocess it to generate standardized multichannel vegetation index images. A lightweight semi-supervised model, MCVI-SANet, is constructed and trained using a VICReg-driven semi-supervised training strategy. The process includes: constructing MCVI-SANet, which comprises a vegetation index saturation sensing module (VI-SABlock), a lightweight backbone network, an expander, and a regressor. The VI-SABlock, as the first layer of the network, sequentially includes a batch normalization (BN) layer, a feature descriptor extraction module, a feature recalibration excitation (FRE) network, and a deep spatial attention module (DSAM). The VI-SABlock is used to alleviate the vegetation index saturation problem. The lightweight backbone network is constructed based on inverted residual blocks (IRBs) and trained using a VICReg-driven semi-supervised strategy. First, the encoder and expander are pre-trained using unlabeled images. Then, the regressor is fine-tuned on a labeled dataset, and a general feature representation is learned by constraining the model through a triple regularization loss. The multi-channel vegetation index images are input into the trained MCVI-SANet, which outputs quantitative estimation results of winter wheat LAI and SPAD.

2. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 1, characterized in that, The process of acquiring multispectral image data of key growth stages of winter wheat and preprocessing it to generate standardized multichannel vegetation index images includes: UAVs were used to collect multispectral images of winter wheat covering the jointing, booting, heading, flowering, and grain-filling stages. The UAV multispectral images are preprocessed, including radiometric correction, geometric correction and orthophoto stitching, to generate a five-band orthophoto image; Standardized image blocks are obtained by cropping the boundaries of sampled cells. Multiple vegetation indices are calculated and multi-channel vegetation index images are constructed. At the same time, unlabeled image blocks are segmented and generated for self-supervised pre-training. The dataset is divided into training, validation and test sets according to a preset ratio by using a vegetation height-guided stratified sampling strategy. This is achieved through feature normalization, K-Means clustering and majority voting.

3. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 2, characterized in that, The dataset is partitioned using a vegetation height-guided stratified sampling strategy. Through feature normalization, K-Means clustering, and majority voting, the training, validation, and test sets are divided according to a preset ratio, including: Let the original dataset be... D Include |D| Each sample consists of a feature matrix composed of three growth parameters: LAI, SPAD, and vegetation height (VH). First, the features are normalized: in: and These are the mean and standard deviation for each feature; K-Means clustering with K=10 clusters is used to capture the distribution structure of samples in the growth parameter space. The number of cluster repetitions N is set to reduce the sensitivity to random initialization, and a label matrix is ​​generated. ; Each sample is determined by majority vote. j Final clustering labels : in: For indicator functions, Let be the label of the j-th sample in the i-th cluster; In each final cluster Within the training set, stratified sampling is performed according to the preset ratio of the training set, validation set, and test set, and the number of samples in each subset is calculated to ensure that the samples are evenly distributed across the reproductive stages.

4. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 1, characterized in that, The BN layer takes a multi-channel vegetation index input image sequence. X Normalization yields It has a stable and heterogeneous distribution of vegetation characteristics, and can adapt to the differences in the numerical range of different vegetation indices. The feature descriptor extraction module combines adaptive average pooling (AAP) and standard deviation (STD) to capture multi-exponential statistical information: in: The mean, This indicates an adaptive average pooling operation. The input is processed by the BN layer. For the c-th channel image ( i,j The pixel value at position ) Standard deviation, H and W For the feature map height and width, As a numerical stability index, STD retains the standard deviation information under saturation conditions; The Feature Recalibration Activated FRE network learns the nonlinear dependencies between statistical information and adaptively reweights the VI channels through 1×1 convolutions and the Mish activation function. in: The concatenated statistical feature tensor This indicates a splicing operation. For the channel attention weight tensor, This indicates the feature recalibration excitation network. and This represents the activation function. The Mish activation function avoids truncating negative values. This is the feature map after attention weighting. This indicates element-wise multiplication. and represents the weights of a 1×1 convolutional layer, and r is the compression ratio; The DSAM employs 3×3 deep convolution to learn fine-grained spatial weights and uses the tanh activation function to achieve bidirectional modulation. in: Spatial attention weights, This represents a depthwise convolution operation. For the corresponding depth convolution matrix, The feature map has undergone depthwise convolution. The feature map is obtained after feature expansion. Indicates the feature expansion layer. The number of channels after feature expansion.

5. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 4, characterized in that, The lightweight backbone network is composed of multiple inverted residual blocks (IRBs) stacked together. All IRBs are set with a preset step size to maintain residual connections and preserve fine-grained canopy structure features. The lightweight backbone network takes a first-size feature map as input, processes it through multiple IRBs, and outputs a high-dimensional feature map of a second size.

6. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 4 or 5, characterized in that, The semi-supervised training strategy driven by VICReg first uses unlabeled images to perform self-supervised pre-training on the encoder and extender, including: During the self-supervised pre-training phase, the unlabeled image is randomly flipped horizontally or vertically to generate two enhanced views, X and X'. The augmented view is mapped to high-dimensional embeddings Z and Z' through an encoder and expander, and trained with triple regularization loss constraints: in: For consistency loss function, Let Variance loss function be used. Let covariance loss function be used. , , These are the weight coefficients corresponding to the three loss functions. This is the viewpoint obtained after data augmentation for the i-th batch. Let B represent the square of the L2 norm, B be the batch size, and D be the embedding dimension. The constant target value is the standard deviation. Let Z be the batch variance of the d-th dimension embedding, and Cov(Z) be the covariance matrix of the embedding. Pre-training parameter settings include: batch size, number of iteration rounds, initial learning rate, projection head, and output dimension; Remove the projector head, freeze the encoder weights, and fine-tune the regressor on the labeled dataset for fine-tuning training.

7. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 6, characterized in that, The regressor contains two fully connected layers. The first layer maps the high-dimensional feature map output by the backbone network to a low-dimensional latent space, and the second layer outputs LAI or SPAD estimates. During the estimation process, the multi-channel vegetation index images are enhanced by VI-SABlock, then input into a lightweight backbone network to extract spatial spectral features. After adaptive average pooling and flattening, the images are input into a regressor to obtain the final estimation results.

8. The method for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model according to claim 1, characterized in that, The metrics used for evaluating the performance of MCVI-SANet are as follows: in: As the coefficient of determination, Let n be the root mean square error, and n be the total number of samples. Let be the measured value of the i-th sample. For the corresponding predicted value, This is the mean of all measured values.

9. A system for estimating LAI and SPAD of winter wheat based on a lightweight semi-supervised model, characterized in that, include: The data acquisition and processing module is used to acquire multispectral image data of winter wheat during key growth periods and preprocess it to generate standardized multichannel vegetation index images. The estimation model building module is used to construct a lightweight semi-supervised model MCVI-SANet and perform self-supervised training on MCVI-SANet using a VICReg-driven semi-supervised training strategy, including: The MCVI-SANet is constructed as follows: a vegetation index saturation sensing module VI-SABlock, a lightweight backbone network, an extender, and a regressor; The VI-SABlock, as the first layer of the network, sequentially includes: a batch normalization (BN) layer, a feature descriptor extraction module, a feature recalibration excitation (FRE) network, and a deep spatial attention module (DSAM). The VI-SABlock is used to alleviate the vegetation index saturation problem, and the lightweight backbone network is constructed based on inverted residual blocks (IRB). A semi-supervised training strategy driven by VICReg is adopted, and the encoder and extender are first pre-trained in a self-supervised manner using unlabeled images. Then, fine-tune the regressor on the labeled dataset and learn a general feature representation by using a triple regularization loss constraint model; The estimation result output module is used to input the multi-channel vegetation index image into the trained MCVI-SANet to output the quantitative estimation results of winter wheat LAI and SPAD.

Images