CN114755189B - Feature-optimized self-attention-based hyperspectral satellite LAI inversion method - Google Patents
Feature-optimized self-attention-based hyperspectral satellite LAI inversion method Download PDFInfo
- Publication number
- CN114755189B CN114755189B CN202210338869.1A CN202210338869A CN114755189B CN 114755189 B CN114755189 B CN 114755189B CN 202210338869 A CN202210338869 A CN 202210338869A CN 114755189 B CN114755189 B CN 114755189B
- Authority
- CN
- China
- Prior art keywords
- model
- leaf area
- area index
- spectrum
- inversion
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
- G01B11/00—Measuring arrangements characterised by the use of optical techniques
- G01B11/28—Measuring arrangements characterised by the use of optical techniques for measuring areas
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N2021/1793—Remote sensing
Abstract
The invention discloses a feature-optimized self-attention mechanism hyperspectral satellite LAI inversion method which is used for coupling a PROSAIL radiation transmission model with the feature-optimized self-attention mechanism and realizing the leaf area index inversion of a high-spectrum image on a regional scale. The method comprises the following steps: 1) Simulating a vegetation canopy spectrum using a PROSAIL radiation transmission model; 2) Resampling and Gaussian noise adding are carried out on the simulated spectrum; 3) Screening characteristic wave bands by using a random forest; 4) Constructing a leaf area index inversion model based on the self-attention mechanism model; 5) And respectively evaluating the precision of the leaf area index inversion model for the ground measured spectrum and the hyperspectral image. Compared with the traditional method, the method effectively improves the inversion speed, enhances the stability and generalization capability of the model, reduces the dependence on ground survey data, and is suitable for the leaf area index inversion of the regional scale hyperspectral image.
Description
Technical Field
The invention relates to vegetation radiation transmission model simulation and remote sensing data inversion processing, in particular to a self-attention-mechanism hyperspectral satellite Leaf Area Index (LAI) inversion method for coupling a PROSAIL radiation transmission model and performing characteristic optimization, which is a method for realizing the model construction based on simulation data to be applied to regional scale hyperspectral image fast Leaf Area Index (LAI) inversion and belongs to the field of satellite remote sensing technology and application.
Background
The Leaf Area Index (LAI) is an important composition parameter in models of carbon circulation, water circulation and the like of an ecological system, has important significance for analyzing photosynthesis and transpiration of vegetation, and the change of the Leaf Area Index (LAI) restricts regional microclimates. The accurate and effective estimation of the regional vegetation leaf area index plays an important indication role in monitoring the health degree of the ecological system.
The traditional method for measuring the leaf area index is generally implemented by using an instrument to manually measure the leaf area index in a vegetation coverage area, a large amount of manpower and financial resources are consumed, the obtained ground data are limited, the leaf area index of the regional scale is not enough to be rapidly obtained, and the real-time dynamic vegetation growth state monitoring can not be achieved. There are significant advantages to inverting LAI using remote sensing means or the like.
The current leaf area index inversion methods are mainly classified into three categories: empirical models, physical models, hybrid approaches of physical models and machine learning algorithms. The empirical model is mainly used for constructing a regression equation such as linear regression and partial least square regression by using the spectral reflectivity of the vegetation and the derived spectral index thereof to invert the leaf area index. The physical model is mainly based on a radiation transmission model and a geometric optical model, and realizes the simulation of parameters such as leaf area index and the like by simulating the transmission process of light on the basis of strict mathematics and physical knowledge. The physical model is more a PROSAIL model formed by combining PROSPECT and SAIL. The hybrid method of the physical model and the machine learning algorithm actively learns the complex features from the physical model by means of the high fitting of the machine learning to the nonlinear data, and achieves a good inversion effect.
Although the above method has a certain practical significance, it also has certain disadvantages. The regression equation constructed by using the empirical model is too simple, the suitability is good in a small area range, the universality of the model is poor when the vegetation type, the sensor and the like are changed, and the large-range popularization is difficult. The physical model needs to input a large number of parameters, the physical mechanism is complex, the time cost of the inversion process is large, and certain challenges exist in wide application. The hybrid method of the physical model and the machine learning algorithm is mainly limited by the performance and generalization capability of different machine learning algorithms, but the hybrid method can effectively solve data redundancy caused by a large amount of data, can deeply mine hidden semantic information by using a proper and advanced machine learning algorithm, and has important significance for quickly and accurately inverting the leaf area index of the hyperspectral image of the regional scale.
Disclosure of Invention
Aiming at the problems of low inversion speed, weak stability and poor generalization capability in the process of inverting the leaf area index by hyperspectral data, the invention provides a self-attention mechanism hyperspectral satellite leaf area index inversion method for coupling a PROSAIL radiation transmission model and performing characteristic optimization.
A self-attention mechanism hyperspectral satellite leaf area index inversion method for coupling a PROSAIL radiation transmission model and performing feature optimization is characterized by comprising the following steps of: the method comprises the following steps:
Step 2, resampling the spectrum of the simulated vegetation canopy, and convolving the simulated spectrum to the spectrum resolution ratio same as that of the hyperspectral satellite by combining the spectrum response function of the hyperspectral satellite;
step 3, adding Gaussian noise, randomly adding Gaussian noise to the vegetation canopy spectrum convolved in the step 2, and simulating the influence of the satellite platform, the attitude, the periodic offset of the sensor, the electromagnetic interference among load components and the like in the imaging process;
step 4, using a random forest to carry out characteristic waveband screening, constructing a sensitive waveband of a canopy spectrum under the asynchronous long-leaf area index of the decision tree evaluation simulation to the leaf area index, and screening out a waveband with the accumulated contribution rate of more than 80%;
step 5, performing inverse regression on the generated huge simulation data by using an attention machine model, further reducing band redundancy on the characteristic bands screened in the step 4 by using an attention machine model, capturing key information, and constructing a model suitable for leaf area index inversion;
step 6, performing the operations in the steps 2 and 3 on the field actual measurement canopy spectrum data, screening out the corresponding characteristic wave band in the step 4, predicting by using the self-attention mechanism model in the step 5, and performing model precision evaluation and inspection by combining the field actual measurement leaf area index;
and 7, preprocessing the hyperspectral image, optimizing model parameters aiming at the inversion result of the field actually-measured spectral data in the step 6, performing leaf area index inversion by using the model in the step 5 based on the optimized parameters, and evaluating the inversion result by combining ground survey data to realize the migration of the simulation data to the real image data.
In step 1, further comprising: setting observation zenith angle parameters in the range of 0-50 degrees by taking 10 as a step length according to parameters required to be combined by the PROSAIL radiation transmission model; setting an average blade inclination angle parameter within the range of 30-70 degrees by taking 10 as a step length; setting reference soil reflectivity parameters according to different soil types; setting structural parameters of the blades within the range of 1.215-3; at 1.326-98.8 mu g/cm 2 Setting chlorophyll content parameters within the range; 0.0003659-0.05249 g/cm 2 Setting an equivalent water thickness parameter within the range; 0.0084-0.0547 g/cm 2 Setting dry matter content parameters in the range; finally, setting leaf area index within the range of 0-7 by taking 0.1 as step length, and obtaining 310000 planted canopy spectral curves through parameter combination cumulative simulation, wherein the spectral range is 400-2500 nm. The range of values of the above parameters to be combined is determined from the LOPEX93 data set, and meets the basic vegetation growth characteristics.
In step 2, further comprising: converting the simulated vegetation canopy data into the reflectivity within the hyperspectral satellite wave band range, wherein the conversion expression is as follows:
wherein rho (lambda) is the wave band reflectivity of the hyperspectral satellite after spectrum convolution; lambda [ alpha ] max 、λ min Respectively the maximum and minimum wavelength ranges of different wave bands; ρ (λ [. Lamda. ]) i ) The high spectral reflectivity of the vegetation canopy simulated by the PROSAIL model is obtained;is the spectral response function of the corresponding hyperspectral satellite.
In step 3, further comprising: the simulated spectrum and the real image spectrum have certain difference under the influence of atmospheric radiation transmission, the characteristics of a sensor and the like, and Gaussian noise is randomly added to eliminate the influence, wherein the expression of the Gaussian noise is as follows:
G(x,y)=f(x,y)+n(x,y)
g (x, y) is a vegetation spectral curve added with random Gaussian noise; f (x, y) is the original vegetation spectral curve; n (x, y) is Gaussian noise that conforms to the normal probability distribution of F (x; mu, sigma), mu being the mean and sigma 2 Is the variance, x is the band range, and y is the spectral reflectance at the corresponding band range.
In step 4, further comprising: and a random forest is used for constructing a decision tree to screen leaf area index characteristic wave bands, so that wave bands with high colinearity are eliminated, and dimension disaster is reduced. Adopting an out-of-bag data (OOB) error rate to evaluate the characteristic Importance VI (Variable Importance) of each band for all the input bands, wherein the characteristic Importance calculation expression is as follows:
wherein errOOB1 is the initial out-of-bag data error rate of the random forest; errOOB2 is the error rate of the data outside the bag after random noise is added to all input sample characteristics; n is that the importance degree of the characteristic wave band is high if the error rate is changed greatly after random noise is added to the number of trees constructed in the random forest model.
In step 5, further comprising: the characteristic wave band is trained and learned by using a self-attention mechanism model, and the model focuses more on characteristic information with high correlation to the leaf area index. The expression of the self-attention mechanism model is as follows:
Q=X×W Q
K=X×W K
V=X×W V
wherein, self-Attention (Q, K, V) is the output result of the Self-Attention mechanism; q is a query vector; k is a key vector; v is a vector of values; q, K and V are multiplied by W by an embedded vector X through three matrixes without weights respectively Q 、W K 、W V Obtaining; k is T Transposing the key vector K; d k Is the dimension of the query vector Q; softmax is a normalized exponential function used to classify the results. And inquiring and matching corresponding keys, and finishing model learning training by indexing corresponding values through the keys.
In step 6, further comprising: and (3) carrying out precision evaluation on the leaf area inversion model on the ground actual measurement spectrum data by using the self-attention mechanism model, and verifying the suitability of the model on the field actual measurement spectrum.
In step 7, further comprising: and (4) performing radiometric calibration, atmospheric correction and geometric correction pretreatment on the hyperspectral image. And meanwhile, in step 7, performing leaf area index inversion on the pixel points on the hyperspectral image by using the proposed self-attention mechanism model, and verifying the feasibility of the model for inverting the leaf area index on the regional-scale hyperspectral image.
Inverting the measured spectrum and hyperspectral image of the leaf area index on the ground and adopting a decision coefficient R 2 And performing precision evaluation on the root mean square error RMSE, wherein the precision evaluation expression is as follows:
wherein m is the number of input samples; y is i The real pixel leaf area index value is obtained;obtaining a predicted pixel leaf area index value for model training inversion;and inputting the average value of the real pixel leaf area index of the sample.
The invention has the advantages and beneficial effects that: the invention designs a self-attention mechanism method for coupling a PROSAIL radiation transmission model and performing characteristic optimization based on a hybrid method of a physical model and a machine learning algorithm, realizes the inversion of the leaf area index of a hyperspectral image on a regional scale, and improves the overall generalization and universality capacity of the model. Based on the above detailed implementation steps of the present invention, those skilled in the art should be able to make any improvements and modifications (such as changing application scenarios) to the present invention without departing from the basic principles and steps of the present invention.
Drawings
Fig. 1 is a flow chart of a method for performing feature-optimized self-attention-based hyperspectral image LAI inversion coupled with a PROSAIL radiation transmission model.
FIG. 2 is a precision result diagram of leaf area index inversion of the field ground actual measurement vegetation spectrum data by using the constructed model.
FIG. 3 is a result diagram of leaf area index inversion of ZY1-02D hyperspectral images by using the constructed model.
Detailed Description
Aiming at the problems of low inversion speed, weak stability and poor generalization capability in the process of inverting the leaf area index by hyperspectral data, the invention provides a self-attention-mechanism hyperspectral image LAI inversion method (shown in figure 1) for coupling a PROSAIL radiation transmission model and performing feature optimization. In order to make the technical scheme and the technical problem solved by the invention clearer, the invention will be further described by taking a ZY1-02D hyperspectral satellite as an example.
Step 2, resampling the spectrum of the simulated vegetation canopy, and convolving the simulated spectrum to the interval and the wave band number of the same wave band of the ZY1-02D hyperspectral satellite by combining the spectral response function of the ZY1-02D hyperspectral satellite;
step 3, adding Gaussian noise, randomly adding the Gaussian noise to the vegetation canopy spectrum convolved in the step 2, and simulating the influence of a satellite platform, an attitude, periodic deviation of a sensor, electromagnetic interference among load components and the like in the imaging process;
step 4, using a random forest to carry out characteristic waveband screening, constructing a sensitive waveband of a canopy spectrum under the asynchronous long-leaf area index of the decision tree evaluation simulation to the leaf area index, and screening out a waveband with the accumulated contribution rate of more than 80%;
step 5, performing inverse regression on the generated huge simulation data by using an attention machine model, further reducing band redundancy on the characteristic bands screened in the step 4 by using an attention machine model, capturing key information, and constructing a model suitable for leaf area index inversion;
step 6, performing the operations in the steps 2 and 3 on the field actual measurement canopy spectrum data, screening out the corresponding characteristic wave band in the step 4, predicting by using the self-attention mechanism model in the step 5, and performing model precision evaluation and inspection by combining the field actual measurement leaf area index;
and 7, preprocessing the ZY1-02D hyperspectral image, optimizing model parameters according to an inversion result of the field measured spectral data in the step 6, performing leaf area index inversion by using the model in the step 5 based on the optimized parameters, and evaluating the inversion result by combining ground survey data to realize the migration of the simulation data to the real image data.
Parameters to be combined according to PROSAIL radiation delivery modelSetting observation zenith angle parameters within the range of 0-50 degrees by taking 10 as a step length; setting an average blade inclination angle parameter within the range of 30-70 degrees by taking 10 as a step length; setting reference soil reflectivity parameters according to different soil types; setting structural parameters of the blades within the range of 1.215-3; at 1.326-98.8/mu g/cm -2 Setting chlorophyll content parameters within the range; in the range of 0.0003659-0.05249 g/cm 2 Setting equivalent water thickness parameters within the range; 0.0084-0.0547 g/cm 2 Setting dry matter content parameters within the range; finally, setting leaf area index within the range of 0-7 by taking 0.1 as step length, and obtaining 310000 planted canopy spectral curves through parameter combination cumulative simulation, wherein the spectral range is 400-2500 nm. The range of parameter values to be combined is determined from the LOPEX dataset and corresponds to the basic vegetation growth characteristics.
Converting the simulated vegetation canopy data into the reflectivity in the ZY1-02D hyperspectral satellite wave band range, wherein the conversion expression is as follows:
wherein ρ zy-1D (lambda) is the wave band reflectivity of the ZY1-02D hyperspectral satellite after spectrum convolution; lambda [ alpha ] max 、λ min Respectively the maximum and minimum wavelength ranges of different wave bands; ρ (λ) i ) The high spectral reflectivity of the vegetation canopy simulated by the PROSAIL model is obtained;is the spectral response function of the corresponding ZY1-02D hyperspectral satellite.
The simulated spectrum and the real image spectrum have a certain difference under the influence of atmospheric radiation transmission, the self characteristics of a sensor and the like, and Gaussian noise needs to be randomly added to eliminate the influence, wherein the expression of the Gaussian noise is as follows:
G(x,y)=f(x,y)+n(x,y)
g (x, y) is a vegetation spectral curve added with random Gaussian noise; f (x, y) is the original vegetation spectral curve; n (x, y) is Gaussian noise which conforms to the normal probability distribution of F (x; mu, sigma), mu is the mean value, sigma 2 Is the variance, x is the band range, and y is the spectral reflectance at the corresponding band range.
And (3) a random forest is used for constructing a decision tree to screen leaf area index characteristic wave bands, so that wave bands with high collinearity are eliminated, and dimension disaster is reduced. Adopting an out-of-bag data (OOB) error rate to evaluate the characteristic Importance VI (Variable Importance) of each band for all the input bands, wherein the characteristic Importance calculation expression is as follows:
wherein errOOB1 is the initial out-of-bag data error rate of the random forest; errOOB2 is the error rate of the data outside the bag after random noise is added to all input sample characteristics; and N is the number of trees constructed in the random forest model, and if the error rate is changed greatly after random noise is added, the importance degree of the characteristic wave band is high.
The characteristic wave band is trained and learned by using a self-attention mechanism model, and the model focuses more on characteristic information with high correlation to the leaf area index. The expression of the self-attention mechanism model is as follows:
Q=X×W Q
K=X×W K
V=X×W V
wherein, self-Attention (Q, K, V) is the output result of the Self-Attention mechanism; q is a query vector; k is a key vector; v is a vector of values; q, K and V are multiplied by W by an embedded vector X through three matrixes without weights respectively Q 、W K 、W V Obtaining; k is transposition of the key vector K; d k Is the dimension of the query vector Q; softmax is a normalized exponential function used to classify the results. And querying and matching corresponding keys, and completing model learning training by indexing corresponding values through the keys.
And (3) carrying out precision evaluation on the leaf area inversion model on the ground actually-measured spectrum data by using the self-attention mechanism model, and verifying the suitability of the model on the field actually-measured spectrum.
And (4) carrying out radiometric calibration, atmospheric correction and geometric correction on the ZY1-02D hyperspectral image. Meanwhile, in step 7, the provided self-attention mechanism model is used for evaluating the leaf area index inversion accuracy of the image element points on the hyperspectral image, and the feasibility of the model for inverting the leaf area index on the regional scale hyperspectral image is verified.
Inverting the actual measurement spectrum of the leaf area index on the ground and the ZY1-02D hyperspectral image, and adopting a decision coefficient R 2 And performing precision evaluation on the root mean square error RMSE, wherein the precision evaluation expression is as follows:
wherein m is the number of input samples; y is i The real pixel leaf area index value is obtained;obtaining a predicted pixel leaf area index value for model training inversion;and inputting the average value of the real pixel leaf area index of the sample.
In order to better illustrate the technical scheme and the application of the method in an actual scene, a typical grassland area of an inner Mongolia temperate zone is taken as a research area, a plurality of vegetation canopy spectral data and actually measured leaf area indexes of the typical grassland are collected, and ZY1-02D hyperspectral satellite image data with the same time as the ground data collection time are obtained. By utilizing the technical scheme provided by the invention, the following results are obtained:
(1) Random forest characteristic wave band screening
A decision tree is constructed by random forests, importance evaluation is carried out on each input characteristic wave band through out-of-bag errors, and the result shows that: nearly thirty characteristic wave bands sensitive to leaf area indexes, such as 670nm, 421nm, 911nm, 919nm, 791nm, 842nm, 851nm, 765nm, 782nm and the like, are screened, and the cumulative contribution rate of the characteristic wave bands reaches 80%. Data redundancy brought by full wave bands is effectively reduced through characteristic wave band screening.
(2) Inversion evaluation of field actual measurement spectrum leaf area index
Based on the method provided by the invention, 192 canopy spectra of different vegetation types collected in a typical grassland area of inner Mongolia are input into a model for leaf area index inversion, and the result shows that the correlation coefficient R is 0.892, and the coefficient R is determined 2 At 0.797, the root mean square error RMSE was calculated to be 0.140 (fig. 2). The method provided by the invention can achieve the leaf area index inversion result with higher precision aiming at different vegetation types.
(3) Inversion of ZY1-02D hyperspectral image
Based on the method model provided by the invention, ZY1-02D hyperspectral images are used for leaf area index inversion, the universality of the model in the regional scale is evaluated, an inner Mongolia typical grassland area is selected as an evaluation area, and the leaf area index inversion of the hyperspectral remote sensing images of the inner Mongolia typical grassland area is obtained by coupling a PROSAIL radiation transmission model and a self-attention mechanism (figure 3).
Claims (1)
1. A feature-optimized self-attention mechanism hyperspectral satellite LAI inversion method is characterized by comprising the following steps: the method comprises the following steps:
step 1, using a PROSAIL radiation transmission model, and simulating to obtain a vegetation canopy spectrum under asynchronous longleaf area indexes by carrying out multiple combinations on input model parameters, wherein the parameters to be combined are as follows: observing zenith angleMean blade inclination angle (LIDF) a ,LIDF b ) Soil reflectance (r) soil ) Leaf structure parameter (N), chlorophyll content (C) ab ) Equivalent water thickness (C) w ) Dry matter content (C) m );
Step 2, resampling the spectrum of the simulated vegetation canopy, and convolving the simulated spectrum to the spectrum resolution which is the same as that of the hyperspectral satellite by combining the spectral response function of the hyperspectral satellite;
step 3, adding Gaussian noise, randomly adding Gaussian noise to the vegetation canopy spectrum convolved in the step 2, and simulating the influence of the satellite platform, the attitude, the periodic offset of the sensor and the electromagnetic interference among load components in the imaging process;
step 4, using a random forest to carry out characteristic waveband screening, constructing a sensitive waveband of a canopy spectrum under the asynchronous long-leaf area index of the decision tree evaluation simulation to the leaf area index, and screening out a waveband with the accumulated contribution rate of more than 80%;
step 5, performing inversion regression on the generated huge simulation data by using an attention mechanism model, further reducing band redundancy on the characteristic bands screened in the step 4 by using an attention mechanism, capturing key information, and constructing a model suitable for leaf area index inversion;
step 6, performing the operations in the steps 2 and 3 on the field actual measurement canopy spectrum data, screening out the corresponding characteristic wave band in the step 4, predicting by using the self-attention mechanism model in the step 5, and performing model precision evaluation and inspection by combining the field actual measurement leaf area index;
step 7, preprocessing the hyperspectral image, optimizing model parameters aiming at the inversion result of the field actually-measured spectral data in the step 6, performing leaf area index inversion by using the model in the step 5 based on the optimized parameters, and evaluating the inversion result by combining ground survey data to realize the migration of the simulation data to the real image data;
in step 1, further comprising: according to PParameters required to be combined by the ROSAIL radiation transmission model, and observation zenith angle parameters are set in a range of 0-50 degrees by taking 10 as a step length; setting an average blade inclination angle parameter within the range of 30-70 degrees by taking 10 as a step length; setting a reference soil reflectivity parameter according to different soil types; setting structural parameters of the blades within the range of 1.215-3; at 1.326-98.8 mu g/cm 2 Setting chlorophyll content parameters within the range; in the range of 0.0003659-0.05249 g/cm 2 Setting an equivalent water thickness parameter within the range; 0.0084-0.0547 g/cm 2 Setting dry matter content parameters within the range; finally setting leaf area index within the range of 0-7 by taking 0.1 as step length, and obtaining 310000 vegetation canopy spectral curves through parameter combination cumulative simulation, wherein the spectral range is 400-2500 nm;
in step 2, further comprising: converting the simulated vegetation canopy data into the reflectivity in the hyperspectral satellite wave band range, wherein the conversion expression is as follows:
wherein rho (lambda) is the wave band reflectivity of the hyperspectral satellite after the spectrum convolution; lambda max 、λ min Respectively the maximum and minimum wavelength ranges of different wave bands; ρ (λ [. Lamda. ]) i ) The high spectral reflectivity of the vegetation canopy simulated by the PROSAIL model is obtained;is the spectral response function of the corresponding hyperspectral satellite;
in step 3, further comprising: the simulated spectrum and the real image spectrum have certain difference under the influence of atmospheric radiation transmission and the characteristics of the sensor, and Gaussian noise is randomly added to eliminate the influence, wherein the expression of the Gaussian noise is as follows:
G(x,y)=f(x,y)+n(x,y)
g (x, y) is a vegetation spectral curve added with random Gaussian noise; f (x, y) is the original vegetation spectral curve; n (x, y) is Gaussian noise which conforms to the normal probability distribution of F (x; mu, sigma), mu is the mean value, sigma 2 Is the variance, x is the band range, y is the spectral reflectance under the corresponding band range;
in step 4, further comprising: a random forest is used for constructing a decision tree to screen leaf area index characteristic wave bands, the wave bands with high colinearity are eliminated, and dimension disaster is reduced; and evaluating the characteristic importance VI of each band by adopting the OOB error rate of the data outside the bag for all the input bands, wherein the characteristic importance calculation expression is as follows:
wherein errOOB1 is the initial out-of-bag data error rate of the random forest; errOOB2 is the error rate of the data outside the bag after random noise is added to all input sample characteristics; n is the number of trees constructed in the random forest model, and if the error rate changes greatly after random noise is added, the importance degree of the characteristic wave band is high;
in step 5, further comprising: training and learning the characteristic wave band by using a self-attention mechanism model, wherein the model focuses more on characteristic information with high correlation to leaf area indexes; the expression of the self-attention mechanism model is as follows:
Q=X×W Q
K=X×W K
V=X×W V
wherein, self-Attention (Q, K, V) is the output result of the Self-Attention mechanism; q is a query vector; k is a key vector; v is a vector of values; q, K and V are multiplied by W by an embedded vector X through three matrixes without weights respectively Q 、W K 、W V Obtaining; k T Transposing the key vector K; d is a radical of k Is the dimension of the query vector Q; softmax is a normalized exponential function used for classifying results; inquiring and matching corresponding keys, and finishing model learning training by indexing corresponding values through the keys;
in step 6, further comprising: performing precision evaluation on a leaf area inversion model on ground actually-measured spectrum data by using a self-attention mechanism model, and verifying the suitability of the model on the field actually-measured spectrum;
in step 7, further comprising: performing radiometric calibration, atmospheric correction and geometric correction pretreatment on the hyperspectral image; meanwhile, leaf area index inversion is carried out on the pixel points on the hyperspectral image by the proposed self-attention mechanism model, and feasibility of the model for inverting the leaf area index on the regional scale hyperspectral image is verified;
inverting the measured spectrum and hyperspectral image of the leaf area index on the ground and adopting a decision coefficient R 2 And performing precision evaluation on the root mean square error RMSE, wherein the precision evaluation expression is as follows:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210338869.1A CN114755189B (en) | 2022-04-01 | 2022-04-01 | Feature-optimized self-attention-based hyperspectral satellite LAI inversion method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210338869.1A CN114755189B (en) | 2022-04-01 | 2022-04-01 | Feature-optimized self-attention-based hyperspectral satellite LAI inversion method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114755189A CN114755189A (en) | 2022-07-15 |
CN114755189B true CN114755189B (en) | 2022-12-13 |
Family
ID=82329808
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210338869.1A Active CN114755189B (en) | 2022-04-01 | 2022-04-01 | Feature-optimized self-attention-based hyperspectral satellite LAI inversion method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114755189B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116306322B (en) * | 2023-05-18 | 2023-08-25 | 天津中科谱光信息技术有限公司 | Water total phosphorus concentration inversion method and device based on hyperspectral data |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090237656A1 (en) * | 2008-03-05 | 2009-09-24 | Lin Ma | Tomographic Imaging Using Hyperspectral Absorption Spectroscopy |
CN102878957B (en) * | 2012-09-26 | 2015-05-27 | 安徽大学 | Leaf area index and chlorophyll content inversion method based on remote sensing image optimization PROSAIL model parameters |
CN109886962B (en) * | 2019-03-27 | 2022-08-09 | 南京林业大学 | Vegetation coverage and management measure factor remote sensing inversion method based on LAI and multi-angle data |
CN109977548B (en) * | 2019-03-27 | 2023-06-30 | 南京林业大学 | Leaf area index remote sensing inversion method for different ranges |
CN110544277B (en) * | 2019-08-12 | 2023-01-10 | 蔡建楠 | Method for inverting subtropical vegetation leaf area index by unmanned aerial vehicle-mounted hyperspectral imager |
US20210239606A1 (en) * | 2020-02-04 | 2021-08-05 | Andrea Gabrieli | Computationally efficient method for retrieving physical properties from 7-14 um hyperspectral imaging data under clear and cloudy background conditions |
CN111814707B (en) * | 2020-07-14 | 2022-09-23 | 中国科学院空天信息创新研究院 | Crop leaf area index inversion method and device |
-
2022
- 2022-04-01 CN CN202210338869.1A patent/CN114755189B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN114755189A (en) | 2022-07-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhou et al. | Maximum nighttime urban heat island (UHI) intensity simulation by integrating remotely sensed data and meteorological observations | |
CN110174359B (en) | Aviation hyperspectral image soil heavy metal concentration assessment method based on Gaussian process regression | |
CN111855589A (en) | Remote sensing inversion model and method for rice leaf nitrogen accumulation | |
CN111044516B (en) | Remote sensing estimation method for chlorophyll content of rice | |
CN113252583B (en) | Method for calculating alpine hay coverage based on hay vegetation index | |
CN110569605B (en) | NSGA 2-ELM-based non-glutinous rice leaf nitrogen content inversion model method | |
CN114460013B (en) | Coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method | |
CN115356249A (en) | Satellite polarization PM2.5 estimation method and system based on machine learning fusion model | |
CN114563353B (en) | Soil heat flux prediction method based on multi-source satellite remote sensing data | |
CN114755189B (en) | Feature-optimized self-attention-based hyperspectral satellite LAI inversion method | |
CN114372707A (en) | High-cold-wetland degradation degree monitoring method based on remote sensing data | |
CN114676636B (en) | Grassland area soil moisture rapid inversion method integrating vegetation and habitat characteristics | |
Hang et al. | Estimation of chlorophyll-a concentration in Lake Taihu from Gaofen-1 wide-field-of-view data through a machine learning trained algorithm | |
CN113887493A (en) | Black and odorous water body remote sensing image identification method based on ID3 algorithm | |
CN113159420A (en) | Wheat LAI estimation method combining GSA-MLs | |
CN116629453A (en) | Remote sensing yield estimation method suitable for whole crop growth period | |
GB2620469A (en) | Spatial prediction and evaluation method of soil organic matter content based on partition algorithm | |
CN117035066A (en) | Ground surface temperature downscaling method coupling geographic weighting and random forest | |
CN114136445B (en) | Method for improving airborne hyperspectral radiation correction precision of unmanned aerial vehicle | |
AU2021105233A4 (en) | Method of Retrieving Surface Temperature from Passive Microwave Remote Sensing Data AMSR E | |
CN115452167A (en) | Satellite remote sensor cross calibration method and device based on invariant pixel | |
CN115238967A (en) | Photovoltaic power prediction method and device combining cloud picture and adjacent power station cluster | |
CN114529838A (en) | Soil nitrogen content inversion model construction method and system based on convolutional neural network | |
CN115953694B (en) | Leaf age space distribution detection method based on satellite remote sensing image | |
CN116297239A (en) | Soil organic matter content hyperspectral inversion method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |