Non-leaf green organ coupled crop photosynthetic area index radar inversion method and system

Abstract

A non-leaf green organ coupled crop photosynthetic area index radar inversion method comprises the steps that S1, a photosynthetic area index CAI is calculated, and the photosynthetic area index CAI is equal to the sum of a leaf area index LAI and a non-leaf green organ area index NFAI; s2, crop microwave scattering characteristics are analyzed, and polarization decomposition is carried out on radar remote sensing data; s3, screening out a polarization characteristic combination with the strongest correlation with a photosynthetic area index (CAI); and S4, on the basis of the selected feature combination, utilizing a machine learning algorithm to construct an inversion model of the photosynthetic area index CAI and the preferable polarization feature, and utilizing the inversion model to perform inversion on the photosynthetic area index of the regional crop to obtain a regional inversion result. The invention further provides a non-leaf green organ coupled crop photosynthetic area index radar inversion system. According to the method, the photosynthetic area index CAI is provided, the leaf area and the non-leaf green organ area are comprehensively considered, the mechanism process of crop growth change and yield formation is scientifically reflected, and the crop growth and yield can be accurately monitored.

Description

Technical Field

[0001] The present invention relates to the technical field of crop photosynthetic area index inversion, and more specifically, to a crop photosynthetic area index radar inversion method and system coupled with non-chlorophyll organs. Background Art

[0002] Remote sensing technology cannot directly obtain crop yield information. It is necessary to connect the relationship between remote sensing features and crop yield based on crop canopy parameters. At the same time, crop canopy parameters can also reflect the status and trend of crop growth and are one of the important indicators for monitoring crop growth. Leaf Area Index (LAI) is one of the most typical canopy parameters, which is defined as the sum of crop leaf areas per unit area. For gramineous crops, leaf area index is an important indicator reflecting the growth status of crop groups, and its size is directly related to the final yield. Direct measurement of leaf area index is time-consuming and labor-intensive, and it is impossible to conduct real-time and dynamic monitoring. Remote sensing technology can monitor leaf area index over a large area and quickly, and has become an important data support for large-scale and efficient monitoring of crop growth and yield. At present, optical remote sensing is one of the main means of leaf area index inversion. However, in southern my country, it is difficult to obtain stable and effective optical image data due to interference from clouds and rain. Radar remote sensing microwaves have the ability to penetrate clouds and fog, and can still effectively capture crop canopy information under cloudy and rainy weather conditions. In addition, radar microwaves are highly sensitive to crop structure and soil structure, especially to the three-dimensional structure of crop canopy components. Therefore, they can be used in large-scale, rapid monitoring of leaf area index applications.

[0003] For gramineous crops represented by wheat, rice, and corn, their canopies are mainly composed of leaves, and existing crop growth and yield remote sensing monitoring is mostly based on leaf area index. However, for some crops (such as rapeseed and soybean), the crop canopy components include not only leaves, but also non-leaf green organs such as stems, siliques, and pods. If crop growth and yield monitoring is based only on leaf area index, it will miss the consideration of non-leaf green organs and fail to accurately reflect the mechanism of crop growth changes and yield formation, which ultimately leads to deviations in the estimation of crop growth and yield changes. Taking rapeseed, a crop with the most active non-leaf green organs, as an example, with the growth and development of rapeseed, the rapeseed canopy gradually changes from leaves to siliques. Leaves are the main canopy component of rapeseed from seedling stage to flowering stage; after rapeseed blooms, siliques grow, and silique skins begin to photosynthesize. Leaves and silique skins are the canopy components of rapeseed together; after the silique stage, siliques grow and develop to a fixed shape, leaves decline, and the rapeseed canopy is mainly composed of siliques. Therefore, rapeseed growth monitoring and yield estimation based on remote sensing technology needs to comprehensively consider leaf area and non-leaf green organ area.

[0004] The inversion of crop canopy parameters based on radar remote sensing data is based on the radar backscattering characteristics and polarization decomposition characteristics of crops, and the inversion models are constructed to invert. The inversion models can be divided into three categories: empirical models, semi-mechanistic models and mechanistic models. Among them, the empirical model does not involve the formation mechanism of crop biomass and yield, but simply establishes a statistical model by combining radar polarization characteristic parameters with canopy parameters. This method is simple and highly operable, but the empirical relationship constructed by the model is difficult to achieve cross-time and cross-space applications, and the model has poor transplantability. The mechanistic model establishes a theoretical model of crop microwave scattering, combines the crop microwave scattering characteristics with mathematical and physical models, and reveals the scattering characteristics of each crop component and its influence mechanism. The mechanistic model is highly mechanistic but complex, and is rarely used in actual use. The semi-mechanistic model introduces the physical mechanism of crop scattering on the basis of the empirical model to increase the mechanistic nature of the empirical model. The semi-mechanistic model is gradually being used in the inversion of crop canopy parameters due to its relatively simple structure, easy implementation, and certain description of the formation mechanism of crop biomass. However, its accuracy needs to be further improved in regional crops with non-leaf green organs. Summary of the invention

[0005] In view of the problems in the background technology, the present invention proposes a radar inversion method for crop photosynthetic area index coupled with non-leaf green organs, comprising: S1, calculating the crop non-leaf green organ area index NFAI, combining with the ground measured leaf area index LAI, calculating the photosynthetic area index CAI, wherein the photosynthetic area index CAI is equal to the sum of the leaf area index LAI and the non-leaf green organ area index NFAI; S2, analyzing the crop microwave scattering characteristics, and performing polarization decomposition on the radar remote sensing data to obtain polarization characteristics with mechanical rationality; S3, analyzing the correlation between different polarization characteristics and the photosynthetic area index and optimizing the feature combination, and screening out the polarization feature combination with the strongest correlation with the photosynthetic area index CAI; S4, based on the selected feature combination, using a machine learning algorithm, constructing an inversion model of the photosynthetic area index CAI and the preferred polarization feature, using the inversion model, combining the radar polarization characteristics of different phases to invert the regional crop photosynthetic area index, and obtaining a regional inversion result.

[0006] On the other hand, the present invention also proposes a crop photosynthetic area index radar inversion system coupled with non-chlorophyll organs, comprising: a computer executable program, which implements the above-mentioned method when executed.

[0007] In order to accurately invert the canopy parameters of crops with active non-leaf green organs, the present invention proposes a crop canopy parameter photosynthetic area index (CAI), which takes into account not only the crop canopy leaves but also its non-leaf green organs; then, radar remote sensing polarization decomposition technology is used to extract mechanical polarization features; a machine learning algorithm is used in combination with ground sample data to invert the crop canopy parameter photosynthetic area index (CAI).

[0008] The beneficial effects of the present invention include:

[0009] 1) The crop canopy parameter leaf area index only considers the crop canopy leaves, and does not consider the crop canopy stems, siliques, pods and other non-leaf green organs, but non-leaf green organs are particularly important for crop growth and development and yield formation. Taking rapeseed as an example, the maximum area of ​​rapeseed silique peel is 1.54 times the maximum leaf area, and the net photosynthetic rate, transpiration rate and light radiation intensity of the rapeseed silique layer are higher than those of the leaf layer. The present invention proposes a canopy parameter photosynthetic area index for crops with active non-leaf green organs, which comprehensively considers the leaf area and the area of ​​non-leaf green organs, scientifically reflects the mechanism of crop growth changes and yield formation, and can accurately monitor crop growth and yield formation processes. In addition, for crops with inactive non-leaf green organs such as Gramineae, the photosynthetic area index is equal to the leaf area index, so the photosynthetic area index proposed in the present invention is still valid for Gramineae crops.

[0010] 2) The inversion method of the present invention selects an adaptive polarization decomposition method in combination with crop geometric characteristics to perform polarization decomposition, thereby obtaining polarization characteristics that are strongly correlated with crop canopy parameters, and using the polarization characteristics with mechanistic rationality to participate in empirical model modeling to obtain a semi-mechanistic model, thereby solving the problem of lack of mechanistic rationality in the empirical model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to make the present invention more easily understood, the present invention will be described in more detail by referring to the specific embodiments shown in the accompanying drawings. These drawings only depict typical embodiments of the present invention and should not be considered as limiting the scope of protection of the present invention.

[0012] Figure 1 is a flow chart of the method of the present invention.

[0013] Figure 2 The present invention provides a technical roadmap for the method of the present invention.

[0014] Figure 3 It is the decision tree structure of the method of the present invention.

[0015] Figure 4 The random forest algorithm flow chart of the method of the present invention.

[0016] Figure 5 This is an overview of the study area of ​​an application example of the method of the present invention.

[0017] Figure 6 The figure is a distribution diagram of field sampling points of an application example of the method of the present invention.

[0018] Figure 7 The feature importance ranking based on random forest is an application example of the method of the present invention.

[0019] Figure 8 The present invention provides a preferred characteristic test result of an application example of the method of the present invention.

[0020] Fig. 9 The present invention is a training set accuracy verification of an application example of the method of the present invention.

[0021] Fig.10 The CAI regional result of rapeseed during the critical growth period is an application example of the method of the present invention.

[0022] Fig.11 The figure is a test set accuracy verification result of an application example of the method of the present invention.

[0023] Fig.12 A radar chart of a CAI inversion test set which is an application example of the method of the present invention. DETAILED DESCRIPTION

[0024] The following describes the implementation modes of the present invention with reference to the accompanying drawings so that those skilled in the art can better understand the present invention and implement it. However, the enumerated embodiments are not intended to limit the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments may be combined with each other, wherein the same components are represented by the same figure marks.

[0025] like Figure 1 The method of the present invention shown comprises steps S1-S4.

[0026] S1, based on agronomic experience parameters, combined with the ground-measured leaf area index LAI, the ground-measured leaf length and width data, the ground-measured non-leaf organ length and diameter data, calculate the crop non-leaf organ area index NFAI and the photosynthetic area index CAI, where the photosynthetic area index CAI is equal to the sum of the leaf area index LAI and the non-leaf organ area index NFAI.

[0027] Existing remote sensing monitoring of crop growth and yield is mostly based on the leaf area index. For gramineous crops represented by wheat, rice, and corn, their canopies are mainly composed of leaves, and the accuracy of growth monitoring based on the leaf area index can be guaranteed. However, if it is directly extended to crops with active non-leaf green organs, it will not be able to accurately reflect the mechanism of crop growth changes and yield formation, which will ultimately lead to deviations in the estimation of crop growth and yield changes. The present invention studies the succession process of photosynthetic organs of crops with active non-leaf green organs, and the proposed photosynthetic area index CAI can fully express the mechanism of crop growth changes and yield formation, providing a new theoretical basis for growth monitoring of crops with active non-leaf green organs. .

[0028] S2, first analyze the crop microwave scattering characteristics, and perform polarization decomposition on the radar remote sensing data to obtain polarization characteristics with mechanical rationality, that is: obtain the corresponding polarization decomposition components through polarization decomposition.

[0029] Radar microwaves can pass through the crop canopy and scatter multiple times between crop flowers, crop siliques and crop leaves, containing information about each layer of the crop canopy. There are many studies on modeling using radar polarization decomposition characteristics and leaf area index, but there is a lack of research on modeling using radar polarization decomposition characteristics and photosynthetic area index CAI. Therefore, the crop microwave scattering characteristics are first analyzed, and the radar remote sensing data is polarized to obtain rational polarization characteristics, that is, the corresponding polarization decomposition components are obtained through polarization decomposition. Preferably, the radar remote sensing data is preprocessed before polarization decomposition: orbit correction, radiation calibration, band synthesis, terrain correction, etc.

[0030] S3, analyze the correlation between different polarization features and photosynthetic area index and optimize the feature combination, and screen out the polarization feature combination with the strongest correlation with photosynthetic area index CAI. The feature optimization method of the present invention innovatively uses the forward introduction method, that is, cyclically introduces the most important features, and performs comparative analysis to determine the optimal feature combination. The method is simple and easy to operate, which not only ensures that the preferred features are of high importance, but also obtains a flexible number of preferred features, solving the drawbacks of conventional fixed selection of the number of preferred features.

[0031] S4, based on the selected feature combination, uses machine learning algorithm to construct the inversion model of photosynthetic area index CAI and preferred polarization features. The inversion model is used to invert the regional crop photosynthetic area index in combination with the polarization characteristics of radars in different phases to obtain the regional inversion results.

[0032] like Figure 2 The technical roadmap of the present invention is shown. The data sources of the method of the present invention come from field collected data and radar remote sensing data. In step S1 as described above, the photosynthetic area index CAI is calculated for the data collected in the field. In step S2, the polarization decomposition method is used to obtain the polarization decomposition components of the radar remote sensing data, and the polarization decomposition methods adopted include: H / A / Alpha, Freeman-Durden, Compact-Pol, etc. Then, a model is constructed. Polarization features are optimized using polarization decomposition components and ground sample data. And machine learning inversion modeling is performed. Preferably, verification applications are performed based on the model: single point verification, regional application and verification. Finally, comparative analysis is performed: model comparison, polarization mode comparison, and polarization decomposition feature comparison.

[0033] In one embodiment, the step of calculating the photosynthetic area index CAI in step S1 comprises:

[0034] S11, based on the actual leaf length and width data measured on the ground, the crop leaf area S is calculated using the agronomic empirical formula (1) leaf ,

[0035] S leaf =2*a*L leaf *W leaf (1)

[0036] Among them, S leaf is the crop leaf area to be solved, L leaf is the blade length, W leaf is the leaf width, and a is the coefficient of the agronomic empirical formula for calculating the crop leaf area.

[0037] S12, using the leaf area index LAI data measured on the ground, a linear relationship model between leaf area and leaf area index LAI is established, as shown in formula (2).

[0038]

[0039] Among them, b0, b, … b n These are the coefficients of the linear relationship model between leaf area index LAI and leaf area.

[0040] S13, calculating the surface area of ​​the non-chlorophyll organs of the crop based on the measured data of the length and diameter of the non-chlorophyll organs and the agronomic empirical formula (3).

[0041] S nl =L nl *W nl *c+d (3)

[0042] Among them, S nl To obtain the surface area of ​​the non-chlorophyll organs of the crop, L nl is the length of the non-chlorophyll organ, W nl is the diameter / width / thickness of non-chlorophyll organs, c and d are the coefficients of the agronomic empirical formula for calculating the surface area of ​​non-chlorophyll organs.

[0043] S14, calculating the non-leaf green organ area index NFAI according to the relationship model between the leaf area and the leaf area index LAI established in S12.

[0044]

[0045] Among them, NFAI is the non-chlorophyll organ area index, b0, b,…b n These are the coefficients of the linear relationship model between leaf area index LAI and leaf area.

[0046] S15, combining the calculated non-leaf green organ area index NFAI and the measured leaf area index LAI, and calculating the crop photosynthetic area index CAI according to formula (5).

[0047] CAI=LAI+NFAI (5)

[0048] An implementation of the radar remote sensing data polarization decomposition method of step S2 is described below.

[0049] Polarization target decomposition can be divided into two categories: coherent decomposition and incoherent decomposition. Coherent decomposition is mainly based on the polarization scattering matrix, and representative methods include Krogager decomposition and Cameron decomposition. Incoherent decomposition methods are more suitable for the analysis of crop microwave scattering characteristics. Incoherent decomposition is mainly based on the polarization coherence matrix and the polarization covariance matrix, including polarization target decomposition based on eigenvalue decomposition and polarization target decomposition based on scattering model.

[0050] The following describes the implementation of the polarization target decomposition method based on eigenvalue decomposition. The technical idea is to perform eigenvalue decomposition on the polarization coherence matrix and use the intrinsic connection and physical meaning of each eigenvalue and eigenvector to explain the ground object scattering. Taking the Cloude-Pottier decomposition of full polarization data as an example, the core expression is:

[0051] T=U∑U -1 (6)

[0052] Where ∑ is a diagonal matrix, and the 3 eigenvalues ​​λ of the polarization coherence matrix T are i , i=1,2,3. U is a unitary matrix, which consists of the three eigenvectors u of the polarization coherence matrix T i , i=1,2,3. For the eigenvector u i By parameterization, we can get:

[0053]

[0054] In the formula, α, β, j, δ, γ are the characteristic vectors u i Parameter item.

[0055] Using the eigenvalue decomposition results, three commonly used polarization features (H / A / Alpha polarization decomposition) are obtained: polarization entropy H, polarization anti-entropy A, and angle parameter α. These three polarization features are described below.

[0056] 1) Polarization entropy H

[0057] In a scattering unit, there are usually multiple scatterers with different scattering types. Polarization entropy is used to describe the statistical randomness of different scatterers, which is defined as:

[0058]

[0059] Where λ is the eigenvalue of the polarization coherence matrix T.

[0060] The polarization entropy H ranges from [0,1]. If the polarization entropy H is low (H<0.3), the depolarization effect of the scattering unit is considered to be weak, so the main scattering mechanism can be extracted. If the polarization entropy is large, scatterers with different scattering types are mixed in one scattering unit. As the polarization entropy value increases further, the distinguishability of the scattering mechanism categories in the observed data decreases. When H=1, the observed data is completely random noise, and no useful polarization information can be extracted.

[0061] 2) Polarization anti-entropy A

[0062] Polarization anti-entropy A is used to supplement polarization entropy H, mainly to characterize the gap between two smaller eigenvalues. If the three eigenvalues ​​are sorted in descending order as λ1≥λ2≥λ3>0, polarization anti-entropy A is defined as:

[0063]

[0064] Where λ is the eigenvalue of the polarization coherence matrix T.

[0065] When the polarization entropy H>0.7, the polarization anti-entropy A can be used to further distinguish the scattering type of the scatterer. When the polarization entropy H is low, the polarization anti-entropy A is seriously interfered by noise.

[0066] 3) Angular parameter α

[0067] The angular parameter α is the core parameter for identifying the main scattering mechanism and is defined as:

[0068]

[0069] In the formula, p is as shown in formula (8), α is the characteristic vector u i Parameter item.

[0070] Specifically, when α→0, the corresponding scattering mechanism is mainly the odd-order scattering mechanism caused by the rough surface, also called the surface scattering mechanism; when α→π / 4, the corresponding scattering mechanism is the volume scattering mechanism; when α→π / 2, the corresponding scattering mechanism is the secondary scattering mechanism.

[0071] The following describes an implementation method of a polarization target decomposition method based on a scattering model. In the polarization target decomposition method based on a scattering model: 1) decompose the scattering mechanism of the scatterer into basic scattering components such as even scattering (dihedral scattering or secondary scattering), odd scattering (flat plate scattering or surface scattering), volume scattering and spiral scattering; 2) analyze the energy and other physical parameters of each basic scattering component to interpret the scattering mechanism of the scatterer. The advantage of this method is that it is directly based on the basic scattering model, and its decomposition results have clear physical meaning and are widely used. Taking the Freeman-Durden decomposition of fully polarized data as an example, the core expression is:

[0072] T=f vol T vol +f dbl T dbl +f odd T odd (11)

[0073] Where T is the polarization coherence matrix, T vol , T dbl and T odd They are volume scattering, secondary scattering and odd scattering models respectively. vol , f dbl and f odd They are the contribution coefficients of volume scattering, secondary scattering and odd scattering models respectively. vol , T dbl and T odd Specifically:

[0074]

[0075]

[0076]

[0077] Where α is the parameter of the quadratic scattering model and β is the parameter of the odd-order scattering model.

[0078] The following describes an implementation of the feature importance analysis and feature optimization method in step S3 of the method of the invention. The random forest algorithm is used in combination with ground sample data to analyze the correlation between different polarization features and photosynthetic area index and optimize feature combinations to screen out the polarization feature combination with the strongest correlation with photosynthetic area index CAI.

[0079] Random Forest (RF) can use a variety of rules to rank the importance of each feature, and use this as the basis for feature selection. Random Forest's methods for evaluating feature importance include: 1) Counting the number of times each feature in each decision tree of the random forest is used as a split feature. The more times it is used as a split feature, the more important the feature is; 2) When building a decision tree, calculate the impurity gain of the feature when it is used for splitting, such as using the Gini coefficient as an impurity metric. Because there are many decision trees in the random forest, the average of the impurity gain of the feature in different decision trees can be used as a measure of its importance; 3) Rearrange the feature in the out-of-bag (OOB) sample, and determine the importance of the feature by analyzing the results of the rearrangement.

[0080] Taking the Gini coefficient as an example, the standard for splitting child nodes during decision tree construction is the impurity gain value, which is the difference between the Gini coefficient of the node before the split and the Gini coefficient of the child node after the split. The larger the difference, the more obvious the effect of using the feature for segmentation. By counting the impurity gain values ​​corresponding to the feature in different decision trees and calculating their average or weighted average, the final importance score of the feature can be obtained.

[0081] Feature optimization is to select a true subset from a feature set, and only perform data mining or machine learning on this subset. Generally, the size of a feature subset is much smaller than the original feature set, and an excellent feature selection algorithm can filter out features that are more useful for model building from the original feature set and remove the noise features mixed in the original feature set, which not only reduces the time cost of model building, but also improves the accuracy of the model.

[0082] The characteristic optimization method of the present invention comprises:

[0083] 1) All features are analyzed and ranked by importance through random forest. The more important features are ranked higher, which is called feature sequence P.

[0084] 2) The current feature set Q is an empty set.

[0085] 3) Take the most important features from P and put them into Q, reduce the number of features in P by 1, and increase the number of features in Q by 1. Use random forest to model the Q features and analyze the results (modeling accuracy and speed).

[0086] 4) Repeat 3) until P is empty.

[0087] 5) Compare and analyze all the results in the process to obtain the optimal feature set.

[0088] The random forest used in step S3 is an integrated model, and the base model selected is the decision tree. The structure of the decision tree is as follows Figure 3 As shown in the figure, the decision tree is similar to the tree structure in the data structure. The decision tree model is divided into leaf nodes and non-leaf nodes according to the position of the nodes. The top node is called the root node. Each non-leaf node contains a judgment standard. The object to be evaluated in the data set first enters the root node. According to the judgment standard of the root node, one of the directed edges is selected to enter the next layer of nodes. Then, according to the judgment standard in each node, it goes down until it enters the leaf node to obtain the final prediction result.

[0089] The construction principle of the decision tree is as follows: first, all the features of the data set are screened, and according to the classification criteria formula, the feature with the best classification effect is selected as the root node, and the root node is used to classify the original data set, thereby splitting multiple child nodes, and then repeating the previous steps for each child node until the new node grown cannot be divided any further.

[0090] The split evaluation index of sub-nodes is very important in the process of decision tree construction. The most commonly used method is the impurity measurement method, including the Gini coefficient method, the gain ratio method, and the information gain method.

[0091] The formula for the Gini coefficient is:

[0092]

[0093] In the formula, N represents the total number of categories, p i It indicates the probability that the current sample data belongs to the i-th category. The Gini coefficient reflects the probability that a data is misclassified when it is randomly selected from a set of data. The smaller the Gini coefficient, the smaller the probability of misclassification, which means the higher the purity of the original data set. When the Gini coefficient is 0, it means that all data categories of the original data set are the same. The function expression of using the Gini coefficient to measure the impurity of a decision tree node is as follows:

[0094]

[0095] In the formula, N represents the total number of categories, and {P(i|a)} represents the conditional probability of being classified as the i-th category in node a. From this, we can get the impurity gain function:

[0096]

[0097] In the formula, N represents the total number of categories, P L With P RThey represent the probability of the parent node data being divided into the left node and the right node, that is, the proportion of sample data divided into the left node and the proportion of sample data divided into the right node. In order to achieve the optimal split of the decision tree, it is necessary to maximize the impurity gain, that is, to select the optimal split feature range to maximize ΔImpurity(a).

[0098] Refer to the following Figure 3-Figure 4 An embodiment of step S4 of the method of the present invention is described.

[0099] S4, inversion modeling through random forest algorithm. Model training is performed using the optimal polarization feature combination and photosynthetic area index CAI; when performing regional inversion, the optimal polarization feature combination is input into the trained model, and the model outputs the photosynthetic area index CAI.

[0100] Random forest is an integrated learning algorithm that improves the prediction accuracy of the overall model by building multiple decision tree models, which can effectively prevent overfitting problems. Random forest trains the base model by repeatedly randomly extracting some samples from the original data set: 1) The optimal feature combination obtained from S3 coupled with multi-phase data is used as the original set, and multiple samples are randomly extracted from the original set to form a temporary training set; 2) Multiple base models are trained using the temporary training sets extracted each time; 3) The voting results of multiple base models determine the final result.

[0101] Random forest algorithm flowchart see Figure 4 ,The algorithm process mainly includes the following steps.

[0102] 1) Randomly select samples with replacement from the total sample to generate multiple new training subsets. Due to sampling with replacement, each training subset may contain repeated samples.

[0103] 2) For each training subset, a decision tree is constructed. Since the directly generated decision tree is not pruned, consider using only some features to train the decision tree.

[0104] 3) Combine the prediction results of all decision trees and determine the final result by simple voting or taking the average.

[0105] The present invention also proposes a crop photosynthetic area index radar inversion system coupled with non-chlorophyll organs, comprising a computer executable program, and when the program is executed, the method described above is implemented.

[0106] The method of the present invention was verified. The rapeseed in Hengyang City, Hunan Province, the main rapeseed producing area in China, was used as the research object to verify the feasibility and applicability of the proposed radar inversion method for regional crop photosynthetic area index coupled with non-leaf green organs. The terrain of the study area is hilly and located in the subtropical monsoon climate zone. The soil type is mainly red soil. The main crop planting system is winter rapeseed-one-season rice double-cropping system. The general map of the study area is shown in Figure 5 .

[0107] Remote sensing data: Winter rapeseed in the study area is sown from September each year to harvested in mid-May of the following year: October to December is the seedling stage, and mid-January, late February, early April and early May of the following year are the bud stage, flowering stage, silique stage and maturity stage of winter rapeseed respectively.

[0108] The experimental remote sensing data uses Sentine-1SLC (Single Look Complex) data covering the study area. The data is in Interferometric Wide swath (IW) mode, and the polarization mode is VV+VH dual polarization. The present invention uses a total of 5 Sentine-1 time-series remote sensing images, and the imaging times are January 14, 2021, February 19, 2021, March 15, 2021, April 8, 2021, and May 2, 2021. The present invention uses SNAP software to perform orbit correction, radiation calibration, band synthesis (Deburst), terrain correction and other preprocessing on the downloaded Sentinel-1 data.

[0109] Ground sampling points: According to the distribution of rapeseed planting in the study area, this paper arranges 67 ground sampling points (see Figure 6 ), crop canopy parameters were collected during the satellite transit times corresponding to the five key growth periods of rapeseed. The data acquisition time periods for ground sample points are from January 12 to January 15, 2021, from February 21 to February 23, 2021, from March 14 to March 17, 2021, from April 3 to April 6, 2021, and from May 1 to May 3, 2021.

[0110] Radar data processing and polarimetric decomposition: The Sentnel-1 data in the study area from January to May 2021 were preprocessed, including orbit correction, radiation calibration, band synthesis, terrain correction and other preprocessing, and then polarimetric decomposition methods such as H / A / Alpha, Raney, and Compact-Pol were used. After preliminary screening, 11-dimensional crop-related polarimetric features were finally obtained, namely compact_l1, compact_ms, compact_mv, compact_p1, compact_Ps, compact_Pv, compact_tau, Raney_Dbl, Raney_m, Raney_Odd, and Raney_Rnd.

[0111] Random forest-based feature importance ranking and feature optimization: This paper combines the random forest method and the Gini coefficient to evaluate the importance of polarization features and analyze the importance of each polarization feature. The results of the random forest-based feature importance ranking are shown in Figure 7 On this basis, the feature optimization scheme proposed in the present invention was used to conduct experiments and obtain the feature optimization test results. Figure 8 The analysis results show that the polarization feature Raney_Rnd can most effectively capture the changes in the photosynthetic area index CAI, and the four effective features (Raney_Rnd, compact_l1, Raney_Dbl, Raney_Odd) can fully capture the changing trend of the photosynthetic area index CAI.

[0112] Machine learning inversion modeling and accuracy verification: For the data of the entire period, a total of 335 samples were divided into training set and test set with a ratio of 7:3, with a total of 234 samples in the training set and 101 samples in the test set. The training set was used for inversion modeling based on the random forest algorithm, and the test set was used to verify the accuracy of the model. The measured values ​​of the training set were compared and verified with the model prediction values, and the determination coefficient R 2 is 0.71, see Fig. 9 , indicating that the model accuracy is relatively reliable.

[0113] Using the trained model, the data of five key growth periods of rapeseed were predicted and the regional inversion results were obtained. Fig.10 It can be observed that in the early stage of rapeseed growth, the photosynthetic area index CAI value is low, and the color in the figure is brown; as the rapeseed grows and develops, the photosynthetic area index CAI value gradually increases, and the photosynthetic area index CAI value reaches its peak at maturity, which is consistent with the actual situation.

[0114] For the test set data, the model prediction value was obtained from the regional map, and the measured value of the test set was compared with the model prediction value. The determination coefficient R 2 is 0.43, reaching a significant pass. Fig.11 , which illustrates the reliability of the model accuracy.

[0115] Furthermore, a radar chart is generated for the measured values ​​of the test set and the model predicted values, see Fig.12 It can be observed that in the red box, the model can effectively capture the high and low values ​​of the photosynthetic area index CAI, which further illustrates the reliability of the model.

[0116] The embodiments described above are only preferred specific implementations of the present invention. This specification uses the phrases "in one embodiment", "in another embodiment", "in yet another embodiment" or "in other embodiments" to refer to one or more of the same or different embodiments according to the present disclosure. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for inverting crop photosynthetic area index radar coupled with non-chlorophyll organs, characterized in that: include: S1, calculate the crop non-leaf green organ area index NFAI, and calculate the photosynthetic area index CAI by combining the ground measured leaf area index LAI, where the photosynthetic area index CAI is equal to the sum of the leaf area index LAI and the non-leaf green organ area index NFAI; S2, analyze the microwave scattering characteristics of crops and perform polarization decomposition on radar remote sensing data to obtain polarization characteristics with mechanical rationality; S3, analyzing the correlation between different polarization characteristics and photosynthetic area index and optimizing the characteristic combination, and screening out the polarization characteristic combination with the strongest correlation with photosynthetic area index CAI; S4, based on the selected feature combination, uses machine learning algorithm to construct the inversion model of photosynthetic area index CAI and preferred polarization features. The inversion model is used to invert the regional crop photosynthetic area index in combination with the polarization characteristics of radars in different phases to obtain the regional inversion results.

2. The inversion method according to claim 1, characterized in that: Step S1 includes: S11, based on the actual leaf length and width data measured on the ground, the crop leaf area S is calculated by the following formula leaf , S leaf =2*a*L leaf *W leaf , Among them, S leaf is the crop leaf area to be solved, L leaf is the blade length, W leaf is the leaf width, a is the coefficient of the agronomic empirical formula for calculating the crop leaf area; S12, using the leaf area index LAI data measured on the ground, a linear relationship model between leaf area and leaf area index LAI is established according to the following formula: Where b0, b, … b n are the coefficients of the linear relationship model between leaf area index LAI and leaf area; S13, calculating the surface area of ​​the non-chlorophyll organs of the crop based on the measured data of the length and diameter of the non-chlorophyll organs and the following formula. S nl =L nl *W nl *c+d, Among them, S nl To obtain the surface area of ​​the non-chlorophyll organs of the crop, L nl is the length of the non-chlorophyll organ, W nl is the diameter / width / thickness of non-chlorophyll organs, c and d are the coefficients of the agronomic empirical formula for calculating the surface area of ​​non-chlorophyll organs; S14, according to the relationship model between leaf area and leaf area index LAI established in S12, the non-leaf green organ area index NFAI is calculated according to the following formula: Where NFAI is the non-chlorophyll organ area index, b0, b, …b n are the coefficients of the linear relationship model between leaf area index LAI and leaf area; S15, combining the calculated non-leaf green organ area index NFAI and the measured leaf area index LAI, and calculating the crop photosynthetic area index CAI according to the following formula: CAI = LAI + NFAI.

3. The inversion method according to claim 1, characterized in that: The polarization characteristics of step S2 include three characteristics: polarization entropy H, polarization anti-entropy A and angle parameter α: Polarization entropy H is used to describe the statistical randomness of different scatterers and is defined as: Where λ is the eigenvalue of the polarization coherence matrix T; Polarization anti-entropy A is used to supplement polarization entropy H to characterize the gap between two smaller eigenvalues. If the three eigenvalues ​​are sorted in descending order as λ1≥λ2≥λ3>0, polarization anti-entropy A is defined as: Where λ is the eigenvalue of the polarization coherence matrix T; The angular parameter α is the core parameter for identifying the main scattering mechanism and is defined as: In the formula, α is the characteristic vector u i Parameter item.

4. The inversion method according to claim 3, characterized in that: In step S2, the polarization target decomposition method based on the scattering model includes: 1) Decompose the scattering mechanism of the scatterer into basic scattering components of even scattering, odd scattering, volume scattering and spiral scattering; 2) Analyze the energy and other physical parameters of each basic scattering component and interpret the scattering mechanism of the scatterer.

5. The inversion method according to claim 1, characterized in that: Step S3 includes evaluating the feature importance, including: 1) Count the number of times each feature in each decision tree of the random forest is used as a segmentation feature. The more times it is used as a segmentation feature, the more important the feature is; 2) When building a decision tree, calculate the impurity gain when the feature is used for splitting; 3) Rearrange the features in the out-of-bag data samples and determine the importance of the feature by analyzing the results of the rearrangement.

6. The inversion method according to claim 5, characterized in that: In step S3, the decision tree construction method includes: screening all features of the data set, selecting the feature with the best classification effect as the root node according to the classification evaluation criterion formula, using the root node to classify the original data set, splitting multiple child nodes, and then repeating the previous steps for each child node until the new node grown cannot be divided any further.

7. The inversion method according to claim 6, characterized in that: In step S3, the feature optimization method includes: 1) Use random forest to analyze the importance of all features and sort them. The more important features are ranked first, which is called feature sequence P; 2) The current feature set Q is an empty set; 3) Take the most important features from P and put them into Q, reduce the number of features in P by 1, increase the number of features in Q by 1, use random forest to model the Q features and analyze the results; 4) Repeat step 3) until P is empty; 5) Compare and analyze all the results in the process to obtain the optimal feature set.

8. The inversion method according to claim 1, characterized in that: In step S4, the base model is trained by repeatedly randomly extracting some samples from the original data set using a random forest algorithm, including: 1) The optimal feature combination obtained from S3 coupled multi-temporal data is used as the original set, and multiple samples are randomly selected from the original set to form a temporary training set; 2) Use the temporary training set extracted each time to train multiple base models; 3) The final result is determined by the voting results of multiple base models.

9. The inversion method according to claim 1, characterized in that: In step S4, the base model is trained by repeatedly randomly extracting some samples from the original data set using a random forest algorithm, and the random forest algorithm includes: 1) Randomly extract samples from the total sample with replacement to generate multiple new training subsets; 2) For each training subset, build a decision tree and use some features to train the decision tree; 3) Combining the prediction results of all decision trees, the final result is determined by simple voting or taking the average.

10. A crop photosynthetic area index radar inversion system coupled with non-chlorophyll organs, characterized in that: include: A computer executable program, which, when executed, implements the method according to any one of claims 1 to 9.

Images