A method for estimating crop leaf area index by coupling physical model and deep learning

Abstract

The present invention discloses a crop leaf area index estimation method by coupling physical models and deep learning, comprising the following steps: step 1: data acquisition and preprocessing; step 2: simulating the reflectivity of the Sentinel‑2 channel; step 3: optimizing the spectral feature combination sensitive to the leaf area index; step 4: constructing a PROSAIL‑Transformer coupling model to estimate the crop leaf area index. The present invention makes full use of the advantages of the canopy radiation transfer model PROSAIL and the deep learning model Transformer, and selects the sample data closest to the Sentinel‑2 image by adopting a sample matching method, aiming to more accurately capture the changes in the real-world environment. The feature selection method of maximum information coefficient (MIC) and recursive feature elimination (RFE) is combined to identify the most sensitive spectral feature combination in the estimation of leaf area index of different crops. The coupling model is trained using the most representative and accurate data to improve the efficiency and accuracy of model training.

Description

Technical Field

[0001] The invention discloses a crop leaf area index estimation method by coupling a physical model and deep learning, and belongs to the fields of quantitative remote sensing technology and big data technology. Background Art

[0002] Crop growth monitoring is an important part of modern agricultural management and is of great significance for increasing crop yields, optimizing resource utilization and implementing precision agriculture. Leaf Area Index (LAI) is an important parameter for crop growth monitoring, which is used to evaluate crop growth status, manage pests and diseases, and predict crop yields. At present, the methods for obtaining leaf area index mainly include ground in-situ measurement and satellite remote sensing inversion. The ground in-situ measurement method has high accuracy, but it is time-consuming and labor-intensive, and the sample data provided is scattered, which is difficult to meet the needs of regional research. With the development of satellite remote sensing technology, quantitative inversion of leaf area index using satellite remote sensing has become the main method for estimating leaf area index.

[0003] Scholars at home and abroad have developed a variety of crop leaf area index estimation methods, mainly including empirical statistical methods, methods based on physical models, and methods based on machine learning. The empirical statistical method inverts the crop leaf area index by establishing an empirical relationship between the crop leaf area index and the band reflectance or spectral index. This method is simple and easy to implement, but it is limited by ground measurement data and has poor stability. The method based on the physical model uses the understanding of physical processes such as reflection, scattering and absorption of sunlight on the vegetation surface to establish a mathematical model that links remote sensing data with crop structure and optical properties to invert the crop leaf area index. This type of method is based on an in-depth understanding of the physical process between vegetation and sunlight, provides a reliable physical basis, has strong interpretability, but has high model complexity and expensive computational cost. The advantage of the method based on machine learning is that it has a low requirement for prior knowledge of the interaction process between vegetation spectral reflectance and solar radiation, and can better adapt to various data characteristics, but has limited ability to capture complex nonlinear relationships in real environments. Therefore, the existing technology has limitations and needs to be improved. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a crop leaf area index estimation method that couples physical models and deep learning in response to the deficiencies in the prior art.

[0005] The technical solution of the present invention is as follows:

[0006] A crop leaf area index estimation method based on coupling physical model and deep learning includes the following steps:

[0007] Step 1: Data acquisition and preprocessing

[0008] Sentinel-2 Level-2A (L2A) satellite remote sensing data were obtained from the European Space Agency Copernicus Open Access Center (https: / / scihub.copernicus.eu / ); the multispectral imager MSI on the Sentinel-2 satellite includes 13 spectral bands with spatial resolutions of 10m, 20m and 60m respectively; the band data with spatial resolutions of 20m and 60m were resampled to 10m to ensure the consistency of spatial resolution of all band data; bands B2, B3, B4, B5, B6, B7, B8, B8A, B11 and B12 were selected for spectral analysis; the quality_scene_classification quality flag was used to mask clouds and shadows; the Sentinel-2 remote sensing images were mosaicked and cropped to obtain clear and high-quality images;

[0009] Sentinel-2 remote sensing images and the random forest (RF) algorithm are used to generate crop type products with a spatial resolution of 10m. By establishing a crop sample database and constructing a continuous Sentinel-2 time series, spectral and phenological characteristics are extracted from it. A training set is constructed in combination with the crop sample database, and the RF algorithm is used for crop classification.

[0010] Step 2: Simulate Sentinel-2 channel reflectivity

[0011] The canopy radiation transfer model PROSAIL is used to simulate the channel reflectance of Sentinel-2. The model integrates the leaf model PROSPECT and the canopy structure model SAIL to effectively capture the physical relationship between the spectral reflectance of the vegetation canopy and the leaf area index (LAI). Based on the experimentally measured crop biochemical parameters and the parameter settings in the existing literature, the value ranges of the key biochemical parameters of the PROSAIL model are determined, including chlorophyll (Cab), carotenoids (Car), anthocyanins (Cant), brown pigments (Cbp), leaf equivalent water content (Cw), leaf dry matter content (Cm), leaf structure (N), average leaf inclination angle (ALA), leaf area index (LAI), hot spot parameters (Hot) and soil parameters (Psoil). The required geometric parameter information is extracted from the Sentinel-2 image.

[0012] The Monte Carlo method was used to generate 500,000 simulation samples, and the parameter values ​​of each sample were randomly generated within a preset range. The PROSAIL model was used to simulate the spectral reflectance of each sample from 400nm to 2500nm. According to the spectral response function of Sentinel-2, the simulated spectral reflectance was converted into the corresponding channel reflectance. A database containing 500,000 Sentinel-2 channel reflectances was constructed.

[0013] In the sample matching and reflectance screening stage, Sentinel-2 images were sampled for three crops: corn, soybean, and rice, and 18,700 representative sample points were obtained for each crop. For each image sample, the reflectance difference between the Sentinel-2 image sample and the PROSAIL simulation sample was quantified by calculating the sum of squared differences (S):

[0014]

[0015] Where N represents the total number of original spectral bands of Sentinel-2 images. and Represent the image reflectance and simulated reflectance of the i-th band respectively; screen the simulated sample database and identify the first 50 groups of samples with the minimum sum of squared differences; for these 50 groups of samples, extract their LAI values ​​and calculate the average value, which is used as the estimated LAI value of the corresponding Sentinel-2 image sample; this process is iterated for all sample points to ensure that each sample point can obtain an accurate LAI value;

[0016] Step 3: Optimize the spectral feature combination sensitive to leaf area index

[0017] 25 sensitive vegetation indices reflecting the physiological condition, coverage and growth status of vegetation were selected, including: Normalized Difference Vegetation Index NDVI, Ratio Vegetation Index RVI, Difference Vegetation Index DVI, Optimized Soil Adjusted Vegetation Index OSAVI, Modified Soil Adjusted Vegetation Index MSAVI, Renormalized Difference Vegetation Index RDVI, Improved Nonlinear Index MNLI, Normalized Difference Infrared Index NDII, Normalized Multiband Drought Index NMDI, Normalized Difference Phenology Index NDPI, Red Edge Normalized Vegetation Index (REDEI) Indices NDVIre1, NDVIre2 and NDVIre3, red edge position REP, normalized difference red edge index NDRE1, new inverted red edge chlorophyll index IRECI, enhanced vegetation index EVI, green normalized vegetation index GNDVI, blue normalized vegetation index BNDVI, modified simple ratio MSR, triangular vegetation index TVI, greenness simple ratio GSR, modified green-red vegetation index MGRVI, visible atmospheric resistance index VARI and atmospheric impedance vegetation index ARVI.

[0018] Since different indices have different spectral responses, the combination of multiple indices can fully capture crop information and enhance the accuracy and reliability of leaf area index inversion; the original spectral bands can provide more comprehensive information and help to accurately capture the subtle changes of crops; the spectral feature set is constructed by combining the original image spectral bands and vegetation indices;

[0019] The maximum information coefficient (MIC) method was used to measure the correlation between various spectral features and leaf area index, which was used to evaluate the impact of different spectral features on leaf area index estimation; each feature and leaf area index were ranked and transformed, and the Spearman rank correlation coefficient of the transformed data was calculated; for each feature, all possible permutations were generated; for each permutation, the Spearman rank correlation coefficient was recalculated; MIC was the maximum value of the coefficient among all permutations; MIC values ​​ranged from 0 (no correlation) to 1 (complete correlation); the above steps were repeated for all features to evaluate the correlation between each feature and leaf area index;

[0020] The recursive feature elimination (RFE) algorithm was used for further spectral feature selection. The spectral feature combinations sensitive to leaf area index were identified by iteratively training the model and gradually eliminating the features with the least impact on model performance. The RFE algorithm was used to determine the spectral feature combinations that were sensitive to leaf area index of corn, soybean and rice, including the ratio vegetation index RVI, the improved nonlinear index MNLI, the normalized difference infrared index NDII, the normalized multi-band drought index NMDI, the normalized difference phenological index NDPI, the red-edge normalized vegetation index NDVIre3, the new inverted red-edge chlorophyll index IRECI, the enhanced vegetation index EVI, the modified green-red vegetation index MGRVI, the visible atmospheric resistance index VARI and the atmospheric impedance vegetation index ARVI.

[0021] Step 4: Construct PROSAIL-Transformer coupling model to estimate crop leaf area index

[0022] The spectral feature combination sensitive to the leaf area index was selected as the coupling point, and the PROSAIL radiation transfer mechanism model and the Transformer deep learning model were coupled to construct the PROSAIL-Transformer coupling model; the model structure includes five key layers: data layer, input embedding layer, position encoding layer, multiple Transformer encoding layers and output layer;

[0023] (1) Data Layer

[0024] The data layer consists of a combination of spectral features sensitive to leaf area index obtained from PROSAIL simulations;

[0025] (2) Input Embedding Layer

[0026] The input samples are converted into sequence data, and each sample is represented as a feature vector through linear projection, which contains various parameters and output information from PROSAIL simulation; these samples are regarded as input sequences;

[0027] (3) Positional encoding

[0028] Add a position code to each sample in the input sequence to indicate its position in the sequence; the introduction of position code helps to distinguish the order of different samples;

[0029] (4) Multiple Transformer encoding layers

[0030] The input sequence and position encoding are input into the Transformer encoder; the Transformer encoder consists of four Transformer encoder layers; each Transformer encoder layer includes two sublayers: a multi-head self-attention layer and a multi-layer perceptron layer; the multi-head self-attention layer captures the global dependencies between different samples in the sequence and computes multiple attention heads in parallel, each head focusing on different aspects of the features; the multi-layer perceptron layer performs nonlinear transformation and combination of features at each position, thereby enhancing the model's fitting ability;

[0031] (5) Output layer

[0032] The output of the last Transformer encoder is passed to the output layer for final prediction; in the output layer, the output from the Transformer encoder is further processed using a fully connected layer and an activation function to obtain the leaf area index estimate;

[0033] The leaf area index inversion was performed using the trained PROSAIL-Transformer coupling model using the spectral feature combination of Sentinel-2 images as input. The Sentinel-2 image features at each time point were input into the coupling model to obtain the corresponding leaf area index estimation value, forming the time series Sentinel-2 leaf area index inversion result.

[0034] The beneficial effects of the present invention are: making full use of the advantages of the canopy radiation transfer model PROSAIL and the deep learning model Transformer, and selecting the sample data closest to the Sentinel-2 image by adopting the sample matching method, aiming to more accurately capture the changes in the real-world environment. Combining the feature selection method of maximum information coefficient (MIC) and recursive feature elimination (RFE), the most sensitive spectral feature combination in the estimation of leaf area index of different crops is identified. The coupled model is trained using the most representative and accurate data to improve the efficiency and accuracy of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Flowchart of crop leaf area index estimation method based on Sentinel-2 imagery and PROSAIL-Transformer coupling model;

[0036] Figure 2Spatial distribution map of leaf area index in time series based on Sentinel-2 image inversion;

[0037] Figure 3 Evaluation of the accuracy of Sentinel-2 leaf area index estimation using ground-truth leaf area index. (a) corn, (b) soybean, (c) rice, (d) all crops;

[0038] Figure 4 Comparison of leaf area index estimation accuracy using different methods DETAILED DESCRIPTION

[0039] The present invention is described in detail below in conjunction with specific embodiments.

[0040] Step 1: Data acquisition and preprocessing

[0041] Sentinel-2 Level-2A (L2A) satellite remote sensing data were obtained from the European Space Agency Copernicus Open Access Center (https: / / scihub.copernicus.eu / ). The multispectral imager MSI on the Sentinel-2 satellite includes 13 spectral bands with spatial resolutions of 10m, 20m and 60m respectively. The band data with spatial resolutions of 20m and 60m were resampled to 10m to ensure the consistency of spatial resolution of all band data. Bands B2, B3, B4, B5, B6, B7, B8, B8A, B11 and B12 were selected for spectral analysis. The quality_scene_classification quality flag was used to mask clouds and shadows. The Sentinel-2 remote sensing images were mosaicked and cropped to obtain clear and high-quality images.

[0042] Sentinel-2 remote sensing images and random forest (RF) algorithm are used to generate crop type products with a spatial resolution of 10m. By establishing a crop sample database and constructing a continuous Sentinel-2 time series, spectral and phenological characteristics are extracted from it. A training set is constructed based on the crop sample database, and crop classification is performed using the RF algorithm.

[0043] Step 2: Simulate Sentinel-2 channel reflectivity

[0044] The canopy radiation transfer model PROSAIL is used to simulate the channel reflectance of Sentinel-2. The model integrates the leaf model PROSPECT and the canopy structure model SAIL, effectively capturing the physical relationship between the spectral reflectance of the vegetation canopy and the leaf area index (LAI). Based on the experimentally measured crop biochemical parameters and the parameter settings in the existing literature, the value ranges of the key biochemical parameters of the PROSAIL model are determined (see Table 1), including chlorophyll (Cab), carotenoids (Car), anthocyanins (Cant), brown pigments (Cbp), leaf equivalent water content (Cw), leaf dry matter content (Cm), leaf structure (N), average leaf inclination angle (ALA), leaf area index (LAI), hot spot parameters (Hot) and soil parameters (Psoil). The required geometric parameter information is extracted from the Sentinel-2 image.

[0045] Table 1 List of PROSAIL model parameter settings

[0046]

[0047]

[0048] The Monte Carlo method was used to generate 500,000 simulation samples, and the parameter values ​​of each sample were randomly generated within a preset range. The PROSAIL model was used to simulate the spectral reflectance of each sample from 400nm to 2500nm. According to the spectral response function of Sentinel-2, the simulated spectral reflectance was converted into the corresponding channel reflectance. A database containing 500,000 Sentinel-2 channel reflectances was constructed.

[0049] In the sample matching and reflectance screening stage, Sentinel-2 images were sampled for three crops: corn, soybean, and rice, and 18,700 representative sample points were obtained for each crop. For each image sample, the reflectance difference between the Sentinel-2 image sample and the PROSAIL simulated sample was quantified by calculating the sum of squared differences (S):

[0050]

[0051] Where N represents the total number of original spectral bands of Sentinel-2 images. and Represent the image reflectance and simulated reflectance of the i-th band respectively. The simulated sample database is screened to identify the top 50 groups of samples with the smallest sum of squared differences. For these 50 groups of samples, their LAI values ​​are extracted and the average value is calculated, which is used as the estimated LAI value of the corresponding Sentinel-2 image sample. This process is iterated for all sample points to ensure that each sample point can obtain an accurate LAI value.

[0052] Step 3: Optimize the spectral feature combination sensitive to leaf area index

[0053] Based on the previous research results and the correlation between leaf area index and spectral vegetation index, 25 sensitive vegetation indices reflecting the physiological condition, coverage and growth condition of vegetation were screened, including: normalized difference vegetation index NDVI, ratio vegetation index RVI, difference vegetation index DVI, optimized soil adjusted vegetation index OSAVI, modified soil adjusted vegetation index MSAVI, renormalized difference vegetation index RDVI, improved nonlinear index MNLI, normalized difference infrared index NDII, normalized multi-band drought index NMDI, normalized difference physical ... Climate index NDPI, red-edge normalized vegetation index NDVIre1, NDVIre2 and NDVIre3, red-edge position REP, normalized difference red-edge index NDRE1, new inverted red-edge chlorophyll index IRECI, enhanced vegetation index EVI, green light normalized vegetation index GNDVI, blue light normalized vegetation index BNDVI, modified simple ratio MSR, triangular vegetation index TVI, greenness simple ratio GSR, modified green-red vegetation index MGRVI, visible light atmospheric resistance index VARI and atmospheric impedance vegetation index ARVI. Among these indices, the normalized difference vegetation index NDVI and enhanced vegetation index EVI mainly reveal the coverage and growth status of vegetation, and are positively correlated with the leaf area index. Indices such as the normalized difference red-edge index NDRE1 and the red-edge normalized vegetation index NDVIre1 reflect the physiological health of vegetation, and their fluctuations may affect the estimation of the leaf area index. Optimizing the soil-adjusted vegetation index OSAVI and the improved nonlinear index MNLI can help mitigate soil effects and enhance vegetation signals. Since different indices have different spectral responses, combining multiple indices can fully capture crop information and enhance the accuracy and reliability of leaf area index inversion. The original spectral bands can provide more comprehensive information and help accurately capture small changes in crops. The spectral feature set is constructed by combining the original image spectral bands and vegetation indices.

[0054] The maximum information coefficient (MIC) method was used to measure the correlation between various spectral features and leaf area index to evaluate the impact of different spectral features on leaf area index estimation. A ranking transformation was performed on each feature and leaf area index, and the Spearman rank correlation coefficient of the transformed data was calculated. For each feature, all possible permutations were generated. For each permutation, the Spearman rank correlation coefficient was recalculated. The MIC is the maximum value of the coefficient among all permutations. The MIC value ranges from 0 (no correlation) to 1 (perfect correlation). Repeat the above steps for all features to evaluate the correlation between each feature and the leaf area index.

[0055] The recursive feature elimination (RFE) algorithm was used for further spectral feature selection. By iteratively training the model and gradually eliminating the features with the least impact on the model performance, the spectral feature combinations sensitive to the leaf area index were identified. Using the RFE algorithm, the spectral feature combinations that were sensitive to the leaf area index of maize, soybean and rice were identified, including the ratio vegetation index RVI, the improved nonlinear index MNLI, the normalized difference infrared index NDII, the normalized multiband drought index NMDI, the normalized difference phenological index NDPI, the red-edge normalized vegetation index NDVIre3, the new inverted red-edge chlorophyll index IRECI, the enhanced vegetation index EVI, the modified green-red vegetation index MGRVI, the visible atmospheric resistance index VARI and the atmospheric impedance vegetation index ARVI.

[0056] Step 4: Construct PROSAIL-Transformer coupling model to estimate crop leaf area index

[0057] The spectral feature combination sensitive to the leaf area index was selected as the coupling point, and the PROSAIL radiation transfer mechanism model and the Transformer deep learning model were coupled to construct the PROSAIL-Transformer coupling model. The model structure includes five key layers: data layer, input embedding layer, position encoding layer, multiple Transformer encoding layers and output layer.

[0058] (1) Data Layer

[0059] The data layer consists of a combination of spectral features sensitive to leaf area index obtained from PROSAIL simulations.

[0060] (2) Input Embedding Layer

[0061] The input samples are converted into sequence data, and each sample is represented as a feature vector through linear projection, which contains various parameters and output information from PROSAIL simulation. These samples are regarded as input sequences.

[0062] (3) Positional encoding

[0063] Add a position code to each sample in the input sequence to indicate its position in the sequence. The introduction of position code helps to distinguish the order of different samples.

[0064] (4) Multiple Transformer encoding layers

[0065] The input sequence and position encoding are fed into the Transformer encoder. The Transformer encoder consists of four Transformer encoder layers. Each Transformer encoder layer consists of two sublayers: a multi-head self-attention layer and a multi-layer perceptron layer. The multi-head self-attention layer captures the global dependencies between different samples in the sequence and computes multiple attention heads in parallel, each focusing on different aspects of the features. The multi-layer perceptron layer performs nonlinear transformation and combination of features at each position, thereby enhancing the model's fitting ability.

[0066] (5) Output layer

[0067] The output of the last Transformer encoder is passed to the output layer for final prediction. In the output layer, the output from the Transformer encoder is further processed using a fully connected layer and activation function to obtain the leaf area index estimate.

[0068] Using the combination of Sentinel-2 image spectral features as input, the leaf area index is inverted through the trained PROSAIL-Transformer coupling model. The Sentinel-2 image features at each time point are input into the coupling model to obtain the corresponding leaf area index estimation value, forming the Sentinel-2 leaf area index inversion result of the time series.

[0069] Figure 2 This is the spatial distribution diagram of the leaf area index of the Sentinel-2 time series estimated by the method of the present invention. Figure 2 It can be seen that the estimated Sentinel-2 time series leaf area index effectively reflects the seasonal fluctuations and growth dynamics of crop leaf area index. Figure 3 To evaluate the accuracy of Sentinel-2 leaf area index estimation using ground-truth leaf area index. Figure 3 It can be seen that the R between the ground-measured leaf area index and the satellite-derived leaf area index is 2 The estimated value is 0.87, and the RMSE is 0.48, indicating a high degree of consistency between the estimated and measured values ​​and the reliability of the method. Figure 4 The method of the present invention is compared with three other methods: empirical model method (linear regression), physical model method (PROSAIL model), machine learning method (support vector machine SVM) in R 2, RMSE and MAE. The method of the present invention outperforms the other three methods in all evaluation indicators. The results show that it has certain advantages in model performance and can provide more accurate and reliable estimation results.

[0070] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A crop leaf area index estimation method based on coupling physical model and deep learning, characterized in that: The following steps are involved: Step 1: Data acquisition and preprocessing Acquire Sentinel-2 satellite remote sensing image data Level-2A from the European Space Agency's Copernicus Open Access Center; the multispectral imager MSI on the Sentinel-2 satellite includes 13 spectral bands with spatial resolutions of 10m, 20m and 60m respectively; resample the band data with spatial resolutions of 20m and 60m to 10m to ensure the consistency of spatial resolution of all band data; Select bands B2, B3, B4, B5, B6, B7, B8, B8A, B11 and B12 for spectral analysis; use the quality_scene_classification quality flag to mask clouds and shadows; mosaic and crop the Sentinel-2 satellite remote sensing images to obtain clear, high-quality images; The crop type products with 10m spatial resolution are generated using Sentinel-2 satellite remote sensing images and random forest algorithm. By establishing a crop sample database and constructing a continuous Sentinel-2 time series, spectral and phenological characteristics are extracted from it. The training set is constructed in combination with the crop sample database, and the RF algorithm is used for crop classification. Step 2: Simulate Sentinel-2 channel reflectivity The canopy radiation transfer model PROSAIL is used to simulate the channel reflectance of Sentinel-2. This canopy radiation transfer model integrates the leaf model PROSPECT and the canopy structure model SAIL, effectively capturing the physical relationship between the spectral reflectance of the vegetation canopy and the leaf area index. Extract the required geometric parameter information from Sentinel-2 satellite remote sensing images; The Monte Carlo method was used to generate 500,000 simulation samples, and the parameter values ​​of each sample were randomly generated within a preset range. The PROSAIL model was used to simulate the spectral reflectance of each sample from 400nm to 2500nm. According to the spectral response function of Sentinel-2, the simulated spectral reflectance was converted into the corresponding channel reflectance. A database containing 500,000 Sentinel-2 channel reflectances was constructed. In the sample matching and reflectance screening stage, Sentinel-2 satellite remote sensing images were sampled for three crops: corn, soybean, and rice, and 18,700 representative sample points for each crop were obtained. For each image sample, the reflectance difference between the Sentinel-2 satellite remote sensing image sample and the PROSAIL simulation sample was quantified by calculating the sum of squared differences S: Where N represents the total number of original spectral bands of Sentinel-2 satellite remote sensing images. and Represent the image reflectance and simulated reflectance of the i-th band respectively; screen the simulated sample database and identify the first 50 groups of samples with the minimum sum of squared differences; for these 50 groups of samples, extract their LAI values ​​and calculate the average value, which is used as the estimated LAI value of the corresponding Sentinel-2 satellite remote sensing image sample; this process is iterated for all sample points to ensure that each sample point can obtain an accurate LAI value; Step 3: Selecting a combination of spectral features that are sensitive to leaf area index Screen 25 sensitive vegetation indices that reflect the physiological status, coverage and growth status of vegetation; use the maximum information coefficient method to measure the correlation between various spectral features and leaf area index to evaluate the impact of different spectral features on leaf area index estimation; perform ranking transformation on each spectral feature and leaf area index, and calculate the Spearman rank correlation coefficient of the transformed data; for each spectral feature, generate all possible permutations; for each permutation, recalculate the Spearman rank correlation coefficient; MIC is the maximum value of the coefficient in all permutations; the range of MIC values ​​is from 0 to 1; repeat the above steps for all spectral features to evaluate the correlation between each spectral feature and leaf area index; A recursive feature elimination algorithm was used for further spectral feature selection. The spectral feature combinations sensitive to leaf area index were identified by iteratively training the model and gradually eliminating the spectral features with the least impact on model performance. The RFE algorithm was used to determine the spectral feature combinations that were sensitive to leaf area index of corn, soybean, and rice. Step 4: Construct PROSAIL-Transformer coupling model to estimate crop leaf area index The spectral feature combination sensitive to the leaf area index is selected as the coupling point, and the PROSAIL radiation transfer mechanism model and the Transformer deep learning model are coupled to construct the PROSAIL-Transformer coupling model; the PROSAIL-Transformer coupling model structure includes five key layers: data layer, input embedding layer, position encoding layer, multiple Transformer encoding layers and output layer; The spectral feature combination of Sentinel-2 satellite remote sensing images is used as input, and the leaf area index is inverted through the trained PROSAIL-Transformer coupling model; the spectral features of the Sentinel-2 satellite remote sensing images at each time point are input into the coupling model to obtain the corresponding leaf area index estimation value, forming the time series Sentinel-2 leaf area index inversion result.

2. The method for estimating crop leaf area index by coupling physical model and deep learning according to claim 1, characterized in that: In step 4: (1) Data Layer The data layer consists of a combination of spectral features sensitive to leaf area index obtained from PROSAIL simulations; (2) Input Embedding Layer The input samples are converted into sequence data, and each input sample is represented as a feature vector through linear projection, which contains various parameters and output information from PROSAIL simulation; these input samples are regarded as input sequences; (3) Position encoding layer Add a position code to each input sample in the input sequence to indicate its position in the sequence; the introduction of position code helps to distinguish the order of different samples; (4) Multiple Transformer encoding layers The input sequence and position encoding are input into the Transformer encoder; the Transformer encoder consists of four Transformer encoding layers; each Transformer encoding layer includes two sublayers: a multi-head self-attention layer and a multi-layer perceptron layer; the multi-head self-attention layer captures the global dependencies between different input samples in the sequence and calculates multiple attention heads in parallel, each of which focuses on different aspects of the features; the multi-layer perceptron layer performs nonlinear transformation and combination of features at each position, thereby enhancing the model's fitting ability; (5) Output layer The output of the last Transformer encoder is passed to the output layer for final prediction; at the output layer, the output from the Transformer encoder is further processed using a fully connected layer and activation function to obtain the leaf area index estimate.

3. The crop leaf area index estimation method based on coupled physical model and deep learning according to claim 1, characterized in that: In step 2, based on the experimentally measured crop biochemical parameters and the parameter settings in the existing literature, the value ranges of the key biochemical parameters of the PROSAIL model are determined, including chlorophyll (Cab), carotenoids (Car), anthocyanins (Cant), brown pigments (Cbp), leaf equivalent water content (Cw), leaf dry matter content (Cm), leaf structure (N), average leaf inclination angle (ALA), leaf area index (LAI), hot spot parameters (Hot) and soil parameters (Psoil).

4. The method for estimating crop leaf area index by coupling physical model and deep learning according to claim 1, characterized in that: In the step 3, the 25 sensitive vegetation indices reflecting the physiological condition, coverage and growth condition of vegetation include: normalized difference vegetation index NDVI, ratio vegetation index RVI, difference vegetation index DVI, optimized soil adjusted vegetation index OSAVI, modified soil adjusted vegetation index MSAVI, renormalized difference vegetation index RDVI, improved nonlinear index MNLI, normalized difference infrared index NDII, normalized multi-band drought index NMDI, normalized difference phenological index NDPI, red edge normalization Vegetation indices NDVIre1, NDVIre2 and NDVIre3, red edge position REP, normalized difference red edge index NDRE1, new inverted red edge chlorophyll index IRECI, enhanced vegetation index EVI, green normalized vegetation index GNDVI, blue normalized vegetation index BNDVI, modified simple ratio MSR, triangular vegetation index TVI, greenness simple ratio GSR, improved green-red vegetation index MGRVI, visible atmospheric resistance index VARI and atmospheric impedance vegetation index ARVI.

5. The method for estimating crop leaf area index by coupling physical model and deep learning according to claim 1, characterized in that: In the step 3, the spectral feature combination that is sensitive to the leaf area index of corn, soybean and rice includes the ratio vegetation index RVI, the improved nonlinear index MNLI, the normalized difference infrared index NDII, the normalized multi-band drought index NMDI, the normalized difference phenological index NDPI, the red-edge normalized vegetation index NDVIre3, the new inverted red-edge chlorophyll index IRECI, the enhanced vegetation index EVI, the modified green-red vegetation index MGRVI, the visible light atmospheric resistance index VARI and the atmospheric impedance vegetation index ARVI.

Images