A method and system for estimating leaf area index based on dynamic feature weighting

Abstract

This invention discloses a method and system for estimating leaf area index (LAI) based on dynamic feature weighting. The method includes fusing spectral, texture, depth, and phenological features of images to construct a multidimensional, multi-temporal remote sensing feature dataset, and dividing it into training and testing sets. Features from the training set and measured LAI values ​​are input into an extreme random tree model for training, resulting in an initial estimation model. The images in the training set are then nonlinearly weighted to obtain reconstructed images. These reconstructed images and K-fold cross-validation are used to train the initial estimation model, yielding a new LAI estimation model. Finally, the test set is input into the new LAI estimation model to estimate the LAI. This invention establishes a dynamic feature weighting mechanism based on SHAP values. By using a weight matrix to differentially scale each feature, the contribution of important features to the model's estimation results is strengthened, effectively suppressing the interference of noisy features.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing technology, and in particular relates to a method and system for estimating leaf area index based on dynamic feature weighting. Background Technology

[0002] Leaf area index (LAI) is a dimensionless parameter, representing the ratio of the total projected area of ​​plant leaves per unit area of ​​land to the total land area. As a core indicator characterizing vegetation canopy structure, it is of great significance. For vegetation, LAI directly reflects the effective area of ​​light energy intercepted by the canopy's leaves, thus determining the efficiency of photosynthetic product accumulation. Furthermore, dynamic changes in LAI can provide important information for diagnosing field nutrient status, pest and disease stress, and other issues. Remote sensing technology, with its significant advantages of wide coverage, temporal dynamics, and multi-dimensional correlation, has demonstrated great potential in monitoring vegetation leaf area index.

[0003] Current methods for leaf area index inversion mainly include the following four approaches: model statistical method, vegetation radiation transfer model inversion method, data assimilation method, and methods based on machine learning and deep learning. These methods each have their own characteristics: While the statistical model method is simple to operate, its application effect heavily relies on empirical models built from ground-measured data, and it requires parameter adjustments for different vegetation types, resulting in poor universality; the vegetation radiation transfer model is based on physical mechanisms and is not limited by changes in vegetation type and environmental background, exhibiting good versatility; however, its calculation process is often complex, and it is difficult to fully utilize the rich feature information provided by multi-source remote sensing data; data assimilation technology can significantly improve inversion accuracy by fusing remote sensing observations and ground-measured data, but this method has strict requirements for data quality and fails to fully explore the deep features in multi-temporal and multi-angle remote sensing data; machine learning and deep learning methods can automatically mine nonlinear relationships in data, showing advantages in inversion efficiency and accuracy, but existing methods generally suffer from insufficient feature engineering and inadequate high-dimensional feature extraction, making it difficult to effectively utilize the multi-dimensional feature information of remote sensing data, such as spectral, texture, temporal, and spatial context, and the models lack interpretability, making it difficult to quantify the contribution of each feature to the inversion results. Therefore, there is an urgent need to develop a new leaf area index estimation method that can fully extract deep features from multi-source remote sensing data, integrate feature weight analysis, and has strong interpretability, so as to improve the accuracy and applicability of leaf area index inversion. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing a leaf area index estimation method based on dynamic feature weighting, comprising the following steps:

[0005] Step 1: Use a drone equipped with a multispectral sensor to acquire multi-temporal multispectral images of plants throughout their entire growth period and preprocess them; simultaneously measure the leaf area index and observe phenology to obtain measured leaf area index values ​​and phenological characteristic data.

[0006] Step 2: Based on the preprocessed multi-temporal multispectral images, obtain the spectral features, texture features and depth features of the images, fuse these three features with phenological feature data to construct a multi-dimensional multi-temporal remote sensing feature dataset, and divide it into training set and test set;

[0007] Step 3: Input the features and measured leaf area index values ​​from the training set into the extreme random tree model for training. Use the trained model as the initial estimation model, use SHAP analysis to quantify the importance of various features, and calculate the SHAP mean and SHAP standard deviation of all features.

[0008] Step 4: Nonlinearly weight the images in the training set to obtain reconstructed images. Use the reconstructed images and K-fold cross-validation to train the initial estimation model and obtain a new leaf area index estimation model.

[0009] Step 4.1: Divide the images in the training set into K disjoint image subsets with similar number of images;

[0010] Step 4.2: Select K-1 image subsets to participate in subsequent calculations;

[0011] Step 4.3: Perform non-linear weighting based on the SHAP values, SHAP mean, and SHAP standard deviation of all features to obtain dynamic feature weights;

[0012] Step 4.4: Reconstruct the images in the K-1 image subsets selected in Step 4.2 according to the dynamic feature weights to obtain the reconstructed images;

[0013] Step 4.5: Train the initial estimation model using the reconstructed images. When the set number of training iterations is reached, the model will automatically stop, and the current training round will end.

[0014] Step 4.6: Repeat steps 4.2-4.5 until K rounds of training are completed, using a different K-1 subset of images for each round of training to obtain a new leaf area index estimation model.

[0015] Step 5: Input the images from the test set into the new leaf area index estimation model to estimate the leaf area index, and evaluate the accuracy of the new leaf area index estimation model by combining the measured leaf area index values.

[0016] Furthermore, the multispectral images acquired in step 1 include five bands: blue, red, green, red-edge, and near-infrared. Preprocessing includes geometric calibration and radiometric correction. The measured leaf area index is obtained by the experimenter using a portable leaf area meter placed above and below the canopy of the target plant area under uniform diffuse light conditions. The phenological characteristic data refers to the number of days after sowing corresponding to each key growth stage of the target plant area as observed and recorded by the experimenter on-site.

[0017] Furthermore, the spectral characteristics of the image in step 2 are obtained through band calculation, including the green chlorophyll index and the optimized soil-regulated vegetation index, and the specific calculation method is as follows:

[0018] (1)

[0019] (2)

[0020] In the formula, The green chlorophyll index is represented by OSAVI, which is the soil-optimized vegetation index. The canopy spectral reflectance is in the near-infrared band. The canopy spectral reflectance is in the green band. The red band represents the spectral reflectance of the canopy.

[0021] The texture features of the image include six statistical measures for each band: second angular moment, contrast, correlation, dissimilarity, energy, and uniformity. The formulas for calculating each statistical measure are as follows:

[0022] (3)

[0023] (4)

[0024] (5)

[0025] (6)

[0026] (7)

[0027] (8)

[0028] In the formula, ASM The second moment of the angle, Contrast For contrast, Correlation For correlation, Dissimilarity For opposites, Energy For energy, Homogeneity For uniformity, The gray-level co-occurrence matrix, i Number the rows of images pixels. jNumber the image pixel columns. N This represents the number of gray levels. and These are the row and column mean values, respectively. and These are the standard deviations of the rows and columns, respectively.

[0029] Six characteristic statistics of five bands (blue, red, green, red edge, and near-infrared) were compared. The texture characteristic statistics of the two bands with the lowest angular second moment, energy, dissimilarity, uniformity, and contrast, and the highest correlation, were selected for subsequent calculations.

[0030] The depth features of the images were extracted using a deep convolutional neural network (DCNN). The DCNN used a VGG-16 model pre-trained on ImageNet, with its fully connected layers removed and the input image size adjusted to fit the image patch size. A global average pooling layer was added after the bottleneck layer of the model to reduce the feature dimension. After the global average pooling layer, three fully connected layers were connected. Each fully connected layer used the sigmoid activation function, and dropout regularization was applied after the first two fully connected layers.

[0031] The spectral features, texture features, and depth features of all images are fused with the phenological features obtained in step 1. Then, the Z-score normalization method is used to normalize all features, constructing a multi-dimensional, multi-temporal remote sensing image feature dataset. The normalization calculation method is as follows:

[0032] (9)

[0033] In the formula, For the first m The first image after standardization l Such characteristics For the first m The first image l Such characteristics For all images l The mean of the features, For all images l The standard deviation of a characteristic.

[0034] The standardized feature dataset is divided into training and test sets.

[0035] Furthermore, in step 3, when training the extreme random tree model, the measured leaf area index (SHAP) value is used as the Y-value for regression training, and the multidimensional, multi-temporal remote sensing image features in the training set are used as the X-value for regression training. Training ends when the set number of training iterations is reached, and the model after training is used as the initial estimation model. The SHAP value is calculated as follows:

[0036] (10)

[0037] In the formula, For the first l The SHAP value of this feature For the set of all features, For set F Size, For feature set F The middle does not include the first l A subset of the features, For subset Size, For using only subsets S The expected output when predicting based on the features in the dataset. For using subsets S Add features l Expected output when making predictions express factorial, express factorial.

[0038] The mean and standard deviation of the Shap for all features are calculated using the Shap value of each feature, as shown in the following formula:

[0039] (11)

[0040] (12)

[0041] In the formula, The number of types of features, The SHAP mean of all features. SHAP standard deviation for all features.

[0042] Furthermore, the calculation method for the dynamic feature weights in step 4.3 is as follows:

[0043] (13)

[0044] In the formula, For the first q Image No. l Dynamic feature weights for each feature; This is an adjustment factor, with a default value of 0.5; For the first q Image No. l The SHAP value of this feature The SHAP mean of all features; SHAP standard deviation for all features; The function is designed to ensure a smooth weight transition, and its calculation formula is as follows:

[0045] (14)

[0046] In the formula, and These are the hyperbolic sine function and the hyperbolic cosine function, respectively.

[0047] Furthermore, the specific calculation method for reconstructing the image using dynamic feature weights in step 4.4 is as follows:

[0048] (15)

[0049] In the formula, For the reconstructed first q Image; For the K-1 image subsets selected in step 4.2, the first... q Image; For the first q The dynamic feature weight matrix of the image; This is the Hadamard product operator; For the K-1 image subsets selected in step 4.2, the first... q The first image l Such characteristics For the first q Image number l Dynamic feature weights for various features.

[0050] The present invention also provides a leaf area index estimation system based on dynamic feature weighting, for implementing the leaf area index estimation method based on dynamic feature weighting as described above.

[0051] Furthermore, it includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a leaf area index estimation method based on dynamic feature weighting as described above.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] 1) This invention proposes a multi-dimensional feature fusion framework of "spectral-texture-depth-phenology", breaking through the limitations of traditional single spectral features; 2) This invention establishes a dynamic feature weighting mechanism based on SHAP values, and performs differentiated scaling on each feature through a weight matrix, strengthening the contribution of important features to the model estimation results and effectively suppressing the interference of noise features; 3) This invention realizes the automated process of "feature importance → weight allocation → model optimization" by coupling SHAP analysis with the machine learning modeling process in a closed loop, providing a new technical path for agricultural remote sensing parameter inversion. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the leaf area index estimation method based on dynamic feature weighting according to an embodiment of the present invention.

[0056] Figure 2 The following are comparative charts of six texture feature statistics for five bands (blue, green, red, red edge, and near-infrared) in an embodiment of the present invention. Among them, (a) is a comparative chart of the second moment feature statistics of the five bands, (b) is a comparative chart of the contrast feature statistics of the five bands, (c) is a comparative chart of the correlation feature statistics of the five bands, (d) is a comparative chart of the dissimilarity feature statistics of the five bands, (e) is a comparative chart of the energy feature statistics of the five bands, and (f) is a comparative chart of the uniformity feature statistics of the five bands.

[0057] Figure 3 This represents the average dynamic feature weights of each feature after a ten-fold crossover in this embodiment of the invention.

[0058] Figures 4(a) and 4(b) are feature importance ranking diagrams based on SHAP analysis in the embodiments of the present invention. Figure 4(a) is the feature importance ranking diagram of the initial estimation model, and Figure 4(b) is the feature importance ranking diagram of the input features after dynamic feature weighting.

[0059] Figure 5 The following are regression scatter plots comparing the predicted and measured values ​​of leaf area index (LAI) obtained using different combinations of input features: (a) is the regression scatter plot of the predicted and measured values ​​of LAI obtained using spectral features; (b) is the regression scatter plot of the predicted and measured values ​​of LAI obtained using spectral features + texture features; (c) is the regression scatter plot of the predicted and measured values ​​of LAI obtained using spectral features + texture features + depth features; and (d) is the regression scatter plot of the predicted and measured values ​​of LAI obtained using spectral features + texture features + depth features + phenological features.

[0060] Figure 6 A graph showing the leaf area index residuals estimated using the method proposed in this invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be further described below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0062] Example 1

[0063] like Figure 1 As shown, this embodiment of the invention provides a leaf area index estimation method based on dynamic feature weighting, which includes the following steps:

[0064] Step 1: Use a drone equipped with a multispectral sensor to acquire multi-temporal multispectral images of plants throughout their entire growth period and preprocess them; simultaneously carry out ground leaf area index measurement and phenological observation to obtain measured leaf area index values ​​and phenological characteristic data.

[0065] The acquired multispectral images included five bands: blue, red, green, red-edge, and near-infrared. Preprocessing of the multispectral images included geometric calibration and radiometric correction, with linear calibration employed for radiometric calibration. Under uniform diffuse light conditions, researchers used a portable leaf area meter placed above and below the canopy of the target plant area to measure the leaf area index. Researchers also conducted visual observations in the field and recorded the number of days after sowing corresponding to each key growth stage of the target plant area as phenological characteristics.

[0066] Step 2: Based on the multi-temporal multispectral images preprocessed in Step 1, obtain the spectral features, texture features, and depth features of the images, fuse them with the phenological feature data obtained in Step 1, construct a multi-dimensional multi-temporal remote sensing feature dataset, and divide it into training set and test set.

[0067] The spectral characteristics of the image were obtained through band calculations, including the Green Chlorophyll Index (CI). green The Optimized Soil-Adjusted Vegetation Index (OSAVI) and its specific calculation methods are as follows:

[0068] (1)

[0069] (2)

[0070] In the formula, The green chlorophyll index is represented by OSAVI, which is the soil-optimized vegetation index. The canopy spectral reflectance is in the near-infrared band. The canopy spectral reflectance is in the green band. The red band represents the spectral reflectance of the canopy.

[0071] Based on the multi-temporal multispectral images preprocessed in step 1, image texture features are calculated, including six statistical features for each band: Angular Second Moment (ASM), Contrast, Correlation, Dissimilarity, Energy, and Homogeneity.

[0072] The Angular Second Moment (ASM) reflects the uniformity of image texture; the larger the value, the more uniform the texture. The calculation formula is as follows:

[0073] (3)

[0074] In the formula, This is a gray-level co-occurrence matrix; i Number the rows of pixels in the image; j Number the image pixel columns; N N represents the number of gray levels, which is the number of possible gray values ​​in the image. It is related to the image bit depth. Generally, when the image bit depth is 8, N is 256.

[0075] Contrast reflects the sharpness of image texture and the depth of texture grooves. The higher the value, the sharper the texture. The calculation formula is as follows:

[0076] (4)

[0077] In the formula, This is a gray-level co-occurrence matrix; i Number the rows of pixels in the image; j Number the image pixel columns; N N represents the number of gray levels, which is the number of possible gray values ​​in the image. It is related to the image bit depth. Generally, when the image bit depth is 8, N is 256.

[0078] Correlation measures the similarity of image textures along the row or column direction. The higher the value, the stronger the correlation. The calculation formula is as follows:

[0079] (5)

[0080] In the formula, This is a gray-level co-occurrence matrix; i Number the rows of pixels in the image; j Number the image pixel columns;N N represents the number of gray levels, which is the number of possible gray values ​​in the image. It is related to the image bit depth. Generally, when the image bit depth is 8, N is 256. and These are the row and column mean values, respectively. and These are the standard deviations of the rows and columns, respectively.

[0081] Dissimilarity measures the degree of dissimilarity between image textures; the higher the value, the greater the texture difference. The calculation formula is as follows:

[0082] (6)

[0083] In the formula, This is a gray-level co-occurrence matrix; i Number the rows of pixels in the image; j Number the image pixel columns; N N represents the number of gray levels, which is the number of possible gray values ​​in the image. It is related to the image bit depth. Generally, when the image bit depth is 8, N is 256.

[0084] Energy, like the second moment of the angle, reflects the uniformity of image texture. The calculation formula is as follows:

[0085] (7)

[0086] In the formula, This is a gray-level co-occurrence matrix; i Number the rows of pixels in the image; j Number the image pixel columns; N N represents the number of gray levels, which is the number of possible gray values ​​in the image. It is related to the image bit depth. Generally, when the image bit depth is 8, N is 256.

[0087] Homogeneity reflects the uniformity of image texture; the higher the value, the more uniform the texture. The calculation formula is as follows:

[0088] (8)

[0089] In the formula, This is a gray-level co-occurrence matrix; i Number the rows of pixels in the image; j Number the image pixel columns; N N represents the number of gray levels, which is the number of possible gray values ​​in the image. It is related to the image bit depth. Generally, when the image bit depth is 8, N is 256.

[0090] Comparison of texture feature statistics across different bands revealed that the red-edge and near-infrared bands exhibited the best performance, with the lowest angular second moment, energy, dissimilarity, uniformity, and contrast, and the highest correlation. Figure 2 This embodiment preferably uses images in the red-edge band and near-infrared band.

[0091] Based on the multi-temporal, multispectral images preprocessed in step 1, depth features are extracted using a deep convolutional neural network. This embodiment uses a VGG-16 model pre-trained on ImageNet, removing its fully connected layers and adjusting the input image size. To adapt to the image patch size, a global average pooling layer is added after the bottleneck layer of the model to reduce the feature dimensionality. Following the global average pooling layer are three fully connected layers, each using a sigmoid activation function to fuse features and perform regression prediction. To avoid overfitting, dropout regularization is applied after the first two fully connected layers with a dropout rate of 20%. The number of neurons in these three fully connected layers are 512, 512, and 1, respectively.

[0092] Spectral features of all images The spectral features OSAVI, six texture feature statistics in the red-edge band, six texture feature statistics in the near-infrared band, and depth features are fused with the phenological features observed in the field in step 1 to construct a multidimensional, multi-temporal remote sensing feature dataset. To eliminate the dimensional differences between features and accelerate the convergence of the Extra_Trees model, the "Z-score normalization" method is used to normalize all features. The specific calculation method is as follows:

[0093] (9)

[0094] In the formula, For the first m The first image after standardization l Such characteristics For the first m The first image l Such characteristics For all images l The mean of the features, For all images l The standard deviation of the feature, in this embodiment , .

[0095] The standardized feature dataset is divided into training and test sets in a 7:3 ratio.

[0096] Step 3: Input the features obtained in the training set in Step 2 and the measured leaf area index obtained in Step 1 into the Extra-Randomized Trees (Extra_Trees) model for training. Use the trained model as the initial estimation model, and use SHapley Additive Explanations (SHAP) analysis to quantify the importance of various features, and calculate the SHAP mean and SHAP standard deviation of all features.

[0097] When training the Extra_Trees model, the measured leaf area index (LAI) values ​​were used as the Y-values ​​for regression training, and the multidimensional, multi-temporal remote sensing image features from the training set were used as the X-values. The model automatically stopped training after 100 decision trees (Boosting rounds), and the trained model was used as the initial estimation model. SHAP analysis was used to quantify the importance of various initial features, and the SHAP mean and standard deviation of all features were calculated. The SHAP values ​​were calculated as follows:

[0098] (10)

[0099] In the formula, For the first l The SHAP value of this feature For the set of all features, For set F A set; For set F Size; For feature set F The middle does not include the first l A subset of the characteristics; For subset Size; For using only subsets S The expected output when predicting features in the dataset; For using subsets S Add features l The expected output when making a prediction; express factorial; express factorial.

[0100] The mean and standard deviation of the Shap for all features are calculated using the Shap value of each feature, as shown in the following formula:

[0101] (11)

[0102] (12)

[0103] In the formula, The number of feature types is 16 in this embodiment; The SHAP mean of all features. SHAP standard deviation for all features.

[0104] Step 4: Nonlinearly weight the images in the training set divided in Step 2 to obtain reconstructed images. Use the reconstructed images and K-fold cross-validation to train the initial estimation model and obtain a new leaf area index estimation model, New_Extra_Trees.

[0105] Step 4.1: Divide the images in the training set divided in Step 2 into K disjoint image subsets with similar number of images.

[0106] Step 4.2: Select K-1 image subsets to participate in subsequent calculations.

[0107] Step 4.3: Perform non-linear weighting based on the SHAP value, SHAP mean, and SHAP standard deviation of all features to obtain dynamic feature weights.

[0108] The dynamic feature weights are calculated as follows:

[0109] (13)

[0110] In the formula, For the first q Image No. l Dynamic feature weights for each feature; This is an adjustment factor, with a default value of 0.5; For the first q Image No. l The SHAP value of this feature The SHAP mean of all features; SHAP standard deviation for all features; The function is designed to ensure a smooth weight transition, and its calculation formula is as follows:

[0111] (14)

[0112] In the formula, and These are the hyperbolic sine function and the hyperbolic cosine function, respectively.

[0113] Step 4.4: Reconstruct the images in the K-1 image subsets selected in Step 4.2 according to the dynamic feature weights to obtain the reconstructed images.

[0114] The specific calculation method for reconstructing images using dynamic feature weights is as follows:

[0115] (15)

[0116] In the formula, For the reconstructed first q Image; For the K-1 image subsets selected in step 4.2, the first... q Image; For the first q The dynamic feature weight matrix of the image; This is the Hadamard product operator; For the K-1 image subsets selected in step 4.2, the first... q The first image l Such characteristics For the first q Image No. l Dynamic feature weights for various features.

[0117] Step 4.5: Train the initial estimation model using the reconstructed images. The model will automatically stop after training 100 decision trees (Boosting rounds), and the training round will end.

[0118] Step 4.6: Repeat steps 4.2-4.5 until K rounds of training are completed. Each time, use a different K-1 subset of images to participate in the training to obtain a new leaf area index estimation model New_Extra_Trees.

[0119] In this embodiment, K is 10, representing the number of images used in each training round. for ,but , The mean dynamic feature weights of each feature after the ten-fold crossover obtained in this embodiment are shown in the figure. Figure 3 The importance of each feature in the initial estimation model was quantified using SHAP analysis, and the comparison of the importance of each input feature after dynamic feature weighting is shown in Figures 4(a) and 4(b). From Figures 4(a) and 4(b), it can be found that phenological features, spectral vegetation index, and depth features contribute far more to the results than image texture.

[0120] Step 5: Input the test set from Step 2 into the new leaf area index estimation model New_Extra_Trees obtained in Step 4 to estimate the leaf area index, and evaluate the accuracy of the new leaf area index estimation model New_Extra_Trees by combining it with the measured leaf area index from Step 1.

[0121] Figure 5To compare the predicted leaf area index (LAI) results with the measured values ​​using spectral features, spectral features + texture features, spectral features + texture features + depth features, and spectral features + texture features + depth features + phenological features as inputs to New_Extra_Trees, a scatter plot was generated. The R² values ​​obtained from these four feature inputs are... 2 The values ​​were 0.51, 0.67, 0.75, and 0.83, respectively, and the MAE values ​​were 1.72, 1.13, 0.86, and 0.56, respectively. Figure 5 As can be seen, integrating multi-dimensional features of "spectrum-texture-depth-phenology" and adopting a dynamic feature weighting mechanism significantly improves the estimation accuracy of leaf area index. Figure 6 This is a map showing the leaf area index residuals estimated using the New_Extra_Trees model with spectral features, texture features, depth features, and phenological features in an embodiment of the present invention. It can be seen that the residual points are uniformly distributed along the straight line. The results show no obvious outliers, and the residual fluctuations do not change as the predicted values ​​increase. There is no heteroscedasticity, which indicates that the New_Extra_Trees model has consistent reliability in predicting all values. The residuals and predicted values ​​have no linear or nonlinear correlation, and the New_Extra_Trees model does not exhibit overfitting.

[0122] Example 2

[0123] Based on the same inventive concept, the present invention also provides a leaf area index estimation system based on dynamic feature weighting, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the leaf area index estimation method based on dynamic feature weighting as described above.

[0124] 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.

[0125] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for estimating leaf area index based on dynamic feature weighting, characterized in that, Includes the following steps: Step 1: Use a drone equipped with a multispectral sensor to acquire multi-temporal multispectral images of plants throughout their entire growth period, and then preprocess them. Simultaneously conduct ground leaf area index measurements and phenological observations to obtain measured leaf area index values ​​and phenological characteristic data; Step 2: Based on the preprocessed multi-temporal multispectral images, obtain the spectral features, texture features and depth features of the images, fuse these three features with phenological feature data to construct a multi-dimensional multi-temporal remote sensing feature dataset, and divide it into training set and test set; Step 3: Input the features and measured leaf area index values ​​from the training set into the extreme random tree model for training. Use the trained model as the initial estimation model, use SHAP analysis to quantify the importance of various features, and calculate the SHAP mean and SHAP standard deviation of all features. Step 4: Nonlinearly weight the images in the training set to obtain reconstructed images. Use the reconstructed images and K-fold cross-validation to train the initial estimation model and obtain a new leaf area index estimation model. Step 4.1: Divide the images in the training set into K disjoint image subsets with similar number of images; Step 4.2: Select K-1 image subsets to participate in subsequent calculations; Step 4.3: Perform non-linear weighting based on the SHAP values, SHAP mean, and SHAP standard deviation of all features to obtain dynamic feature weights; The dynamic feature weights are calculated as follows: （13） In the formula, For the first q Image No. l Dynamic feature weights for each feature; This is an adjustment factor, with a default value of 0.5; For the first q Image number l The SHAP value of this feature The SHAP mean of all features; SHAP standard deviation for all features; The function is designed to ensure a smooth weight transition, and its calculation formula is as follows: （14） In the formula, and These are the hyperbolic sine function and the hyperbolic cosine function, respectively; Step 4.4: Reconstruct the images in the K-1 image subsets selected in Step 4.2 according to the dynamic feature weights to obtain the reconstructed images; The specific calculation method for reconstructing images using dynamic feature weights is as follows: （15） In the formula, For the reconstructed first q Image; For the K-1 image subsets selected in step 4.2, the first... q Image; For the first q The dynamic feature weight matrix of the image; This is the Hadamard product operator; For the K-1 image subsets selected in step 4.2, the first... q The first image l Such characteristics For the first q Image No. l Dynamic feature weights for each feature; Step 4.5: Train the initial estimation model using the reconstructed images. When the set number of training iterations is reached, the model will automatically stop, and the current training round will end. Step 4.6: Repeat steps 4.2-4.5 until K rounds of training are completed, using a different K-1 subset of images for each round of training to obtain a new leaf area index estimation model. Step 5: Input the images from the test set into the new leaf area index estimation model to estimate the leaf area index, and evaluate the accuracy of the new leaf area index estimation model by combining the measured leaf area index values.

2. The leaf area index estimation method based on dynamic feature weighting as described in claim 1, characterized in that: The multispectral images acquired in step 1 include five bands: blue, red, green, red-edge, and near-infrared. Preprocessing includes geometric calibration and radiometric correction. The measured leaf area index was obtained by the experimenters using a portable leaf area meter placed above and below the canopy of the target plant area under uniform diffuse light conditions. Phenological characteristic data refers to the number of days after sowing corresponding to each key growth stage of the target plant area as observed and recorded by the experimenters on-site.

3. The leaf area index estimation method based on dynamic feature weighting as described in claim 1, characterized in that: The spectral characteristics of the image in step 2 are obtained through band calculation, including the green chlorophyll index and the optimized soil-regulated vegetation index. The specific calculation method is as follows: （1） （2） In the formula, The green chlorophyll index is represented by OSAVI, which is the soil-optimized vegetation index. The canopy spectral reflectance is in the near-infrared band. The canopy spectral reflectance is in the green band. The red band represents the spectral reflectance of the canopy.

4. The leaf area index estimation method based on dynamic feature weighting as described in claim 1, characterized in that: The texture features of the image in step 2 include six statistical measures for each band: angular second moment, contrast, correlation, dissimilarity, energy, and uniformity. The formulas for calculating each statistical measure are as follows: （3） （4） （5） （6） （7） （8） In the formula, ASM The second moment of the angle, Contrast For contrast, Correlation For correlation, Dissimilarity For opposites, Energy For energy, Homogeneity For uniformity, The gray-level co-occurrence matrix, i For row numbering, j For column numbering, N This represents the number of gray levels. and These are the row and column mean values, respectively. and These are the standard deviations of the rows and columns, respectively. Six characteristic statistics of five bands (blue, red, green, red edge, and near-infrared) were compared. The texture characteristic statistics of the two bands with the lowest angular second moment, energy, dissimilarity, uniformity, and contrast, and the highest correlation, were selected for subsequent calculations.

5. The leaf area index estimation method based on dynamic feature weighting as described in claim 1, characterized in that: In step 2, the depth features of the image are extracted using a deep convolutional neural network. The deep convolutional neural network uses the VGG-16 model pre-trained on ImageNet, with its fully connected layers removed and the input image size adjusted to fit the size of the image patch. A global average pooling layer is added after the bottleneck layer of the model to reduce the feature dimension. After the global average pooling layer, three fully connected layers are connected. Each fully connected layer uses the Sigmoid activation function, and a dropout regularization method is applied after the first two fully connected layers.

6. The leaf area index estimation method based on dynamic feature weighting as described in claim 1, characterized in that: In step 2, the spectral features, texture features, and depth features of all images are fused with the phenological features obtained in step 1, and the "Z-score normalization" method is used to normalize all features to construct a multidimensional, multi-temporal remote sensing image feature dataset. The normalization calculation method is as follows: （9） In the formula, For the first m The first image after standardization l Such characteristics For the first m The first image l Such characteristics For all images l The mean of the features, For all images l The standard deviation of the characteristic; The standardized feature dataset is divided into training and test sets.

7. The leaf area index estimation method based on dynamic feature weighting as described in claim 1, characterized in that: In step 3, when training the extreme random tree model, the measured leaf area index (LAI) value is used as the Y-value for regression training, and the multidimensional, multi-temporal remote sensing image features in the training set are used as the X-value for regression training. Training ends when the set number of training iterations is reached, and the model after training is used as the initial estimation model. The SHAP value is calculated as follows: （10） In the formula, For the first l The SHAP value of this feature For the set of all features, For set F Size, For feature set F The middle does not include the first l A subset of the features, For subset Size, For using only subsets S The expected output when predicting based on the features in the dataset. For using subsets S Add features l Expected output when making predictions express factorial, express factorial; The mean and standard deviation of the Shap for all features are calculated using the Shap value of each feature, as shown in the following formula: （11） （12） In the formula, The number of types of features, The SHAP mean of all features. SHAP standard deviation for all features.

8. A leaf area index estimation system based on dynamic feature weighting, characterized in that, It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the leaf area index estimation method based on dynamic feature weighting as described in any one of claims 1-7.

Images