CN114460013A - Coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method - Google Patents

Coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method Download PDF

Info

Publication number
CN114460013A
CN114460013A CN202210107317.XA CN202210107317A CN114460013A CN 114460013 A CN114460013 A CN 114460013A CN 202210107317 A CN202210107317 A CN 202210107317A CN 114460013 A CN114460013 A CN 114460013A
Authority
CN
China
Prior art keywords
vegetation
biomass
sample
model
samples
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.)
Granted
Application number
CN202210107317.XA
Other languages
Chinese (zh)
Other versions
CN114460013B (en
Inventor
马毅
陈琛
任广波
王建步
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
First Institute of Oceanography MNR
Original Assignee
First Institute of Oceanography MNR
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by First Institute of Oceanography MNR filed Critical First Institute of Oceanography MNR
Priority to CN202210107317.XA priority Critical patent/CN114460013B/en
Publication of CN114460013A publication Critical patent/CN114460013A/en
Application granted granted Critical
Publication of CN114460013B publication Critical patent/CN114460013B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • 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
    • G01N21/01—Arrangements or apparatus for facilitating the optical investigation
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S19/00—Satellite radio beacon positioning systems; Determining position, velocity or attitude using signals transmitted by such systems
    • G01S19/38—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system
    • G01S19/39—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system the satellite radio beacon positioning system transmitting time-stamped messages, e.g. GPS [Global Positioning System], GLONASS [Global Orbiting Navigation Satellite System] or GALILEO
    • G01S19/42—Determining position
    • 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
    • G01N21/01—Arrangements or apparatus for facilitating the optical investigation
    • G01N2021/0106—General arrangement of respective parts
    • G01N2021/0112—Apparatus in one mechanical, optical or electronic block
    • 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
    • G01N2021/1797—Remote sensing in landscape, e.g. crops

Landscapes

  • Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Remote Sensing (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Physics & Mathematics (AREA)
  • Analytical Chemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Chemical & Material Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

A coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method relates to the technical field of satellite remote sensing, and comprises the following steps: step 1, optimizing feature constraint factors of a GAN network aiming at the vegetation of the coastal wetland, and designing a generation confrontation network model constrained by the feature constraint factors; and 2, constructing a linear regression model by using biomass sample data generated by the generated confrontation network model to invert the biomass on the vegetation ground. The method can improve the inversion accuracy of the biomass of the vegetation of the coastal wetland, and solves the problems that an inversion model of the biomass of the vegetation of the coastal wetland is difficult to construct and the inversion accuracy is low due to the fact that field biomass data are few.

Description

Coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method
Technical Field
The invention relates to the technical field of satellite remote sensing, in particular to a coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method.
Background
With the global increase of the attention on the blue carbon, the coastal wetland is taken as the main body of the coastal zone blue carbon ecosystem, and the importance of the coastal wetland on the carbon cycle is also highlighted.
The hygrophyte vegetation is an important component of the coastal wetland ecosystem, the biomass of the hygrophyte vegetation is one of main indexes for measuring the primary productivity of the coastal wetland ecosystem, and the hygrophyte vegetation is not only a foundation for researching the structure and the function of the coastal wetland ecosystem, but also a scientific basis for researching the carbon fixation capacity and the carbon cycle of the coastal wetland ecosystem and even predicting global change.
In the prior art, there are 2 main means for large-area estimation of coastal wetland biomass:
1. the method of estimation by manual measurement: the method belongs to the traditional method, has obvious defects and is embodied in that: the method needs to calculate the average dry weight of the aboveground part of each type of sample based on biomass sample data of each type of sample which is collected on site and distributed in large quantity, so as to obtain the biomass of the aboveground part with a unit area in each vegetation type, and further calculate the total aboveground biomass in the region. According to the method, a large amount of field investigation is needed to obtain distribution information and biomass samples of all vegetation types, a large amount of manpower and material resources are consumed, and the accuracy of a mode of calculating the biomass in a research area through the biomass in unit area is not high.
2. The method for inversion by satellite remote sensing comprises the following steps: the method is based on the aboveground biomass sample and the spectral characteristics of satellite ground features to establish a model for inverting the vegetation biomass, can realize accurate control on the distribution of vegetation information in a research area, and has great advantages compared with the traditional estimation mode, but has the following defects: the construction of the inversion model requires a large amount of field samples for training, the coastal wetland is located in an unmanned area, the natural conditions are severe, and an investigator is difficult to collect a large amount of biomass samples. Under the condition of small field sample number, the constructed inversion model has the problems of overfitting and low generalization performance, and the inversion result has no reference and application value.
In recent years, the method is successfully applied to many fields based on a neural network and deep learning combined with a satellite remote sensing technology, but an effective solution is not provided in the aspect of coastal wetland biomass inversion under the condition of insufficient samples, so that the method can accurately and effectively measure and calculate the coastal wetland biomass and has important significance in monitoring and protecting the ecological environment of the coastal wetland, and therefore, a technical solution is needed to be provided for the existing problems.
Disclosure of Invention
The invention provides a coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method, which is characterized in that a biomass sample augmentation model based on a generation countermeasure network is built by taking an optimal characteristic factor as a constraint condition, so that the coastal wetland vegetation overground biomass sample augmentation based on joint satellite multispectral data and on-site sampling data is realized, and 4 types of coastal wetland vegetation biomass multiple regression models of reed, suaeda salsa, tamarix chinensis and Spartina alterniflora are respectively built, so that a good biomass inversion result is obtained.
In order to achieve the purpose, the technical scheme of the invention is as follows:
the coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method comprises the following steps:
step 1, optimizing feature constraint factors of a GAN network aiming at the vegetation of the coastal wetland, and designing a generation confrontation network model constrained by the feature constraint factors;
and 2, constructing a linear regression model by using the biomass sample data generated by the designed generation confrontation network model to invert the biomass on the vegetation ground.
Preferably, the step 1 includes the following specific steps:
s1, acquiring satellite multispectral remote sensing image data;
s2, measuring the spectral data of the vegetation canopy on site, manually collecting samples, analyzing and processing the measured spectral data of the vegetation canopy by using spectrum processing software View Spec Pro, thereby obtaining the on-site hyperspectral data of the vegetation, drying and weighing in a laboratory, and calculating the terrestrial biomass value of a unit area;
s3, extracting features, screening feature variables based on correlation analysis, and determining feature constraint factors;
s4, performing equivalent wave band conversion on the remote sensing multispectral data and the field hyperspectral data;
and S5, obtaining a generation countermeasure network model based on the characteristic constraint factor constraint.
Preferably, in S2, the field sample is collected when the weather condition is good and the spectrum collection time is appropriate, and the collection method includes:
recording the vegetation type of a field data acquisition site, selecting a proper sample, recording the longitude and latitude coordinates of the sample by a handheld GPS, recording the vegetation growth parameters in the sample, measuring the spectrum of a vegetation canopy, and acquiring a vegetation sample in the field sample;
for herbaceous vegetation samples, collecting plant samples by taking 0.5m multiplied by 0.5m as a sample in an area with uniform vegetation growth, and randomly obtaining vegetation samples in 5 0.5m multiplied by 0.5m sub-samples within 10m multiplied by 10m in the area with non-uniform vegetation growth; for shrub vegetation, acquiring the crown width and the base diameter of the vegetation, and collecting partial upper trunk and branch samples;
cutting off plants in representative vegetation samples in the condition in a field, and marking site information;
bagging the collected vegetation samples and taking the vegetation samples back to a laboratory, drying the vegetation samples for 24 hours at a constant temperature of 80 ℃ by using an oven to constant weight, then weighing the dry weight of the vegetation samples, and calculating the biomass value of unit area in kg/m by combining with on-site survey record information2。
Preferably, the S3 includes the following steps:
and S31, taking the spectral characteristics and vegetation index characteristics of the multispectral data simulated by the on-site hyperspectral data as variables to be selected, screening the characteristics by a correlation analysis method, and forming a characteristic constraint factor of the GAN model by the screened preferred variables, the spectral characteristics and the biomass for constraining the biomass sample generated by the model.
Preferably, in step S31, the 10 spectral bands in the spectral features are combined pairwise and substituted into 6 vegetation indexes to be calculated to obtain 270 vegetation index feature variables, 280 feature variables of the 10 spectral features and the 270 vegetation index feature variables are used as candidate variables, and the 6 vegetation indexes refer to NDVI, DVI, RVI, SAVI, MSAVI, and OSAVI; calculating variables with strong correlation of biomass characteristics and weak correlation among characteristic interiors in variables to be selected based on a calculation formula (1) of a Pearson correlation coefficient R as preferred variables; wherein the closer the absolute value of R is to 1, the stronger the correlation; the closer the absolute value of R is to 0, the weaker the correlation is; and (3) taking the preferred variable as a part of characteristic constraint factor of the GAN generation sample, and using the preferred variable for constructing a vegetation biomass inversion model:
Figure BDA0003493830120000031
wherein R is Pearson's correlation coefficient, xiAnd yiThe values of the independent variable and the dependent variable at each sample point i are shown, and n is the number of sample points.
Preferably, the S4 includes the following steps:
making a spectrum-biomass sample data set of a simulation satellite through a spectrum conversion model between on-site hyperspectral data of vegetation and satellite multispectral remote sensing image data, and using the spectrum-biomass sample data set as basic data for vegetation biomass data expansion of the coastal wetland; the conversion operation is as follows:
(1) satellite band conversion: establishing conversion between a narrow wave band in the on-site hyperspectral data and a wide wave band in the satellite multispectral remote sensing image data, and calculating by adopting a formula (2):
Figure BDA0003493830120000032
where ρ is the reflectance of a broadband of the simulated satellite, λmin、λmaxThe initial wavelength and the final wavelength of the spectrum detection of the satellite sensor are respectively, S (lambda) is a spectral response function value of the satellite sensor at the lambda wavelength, and rho (lambda) is the reflectivity of the spectrum of the wetland vegetation canopy at the lambda wavelength;
(2) linear conversion among wide bands: and establishing a linear conversion relation between the simulated satellite spectrum and the satellite spectrum, and constructing a linear regression model by using the simulated satellite data and the satellite data to perform reflectivity conversion on each corresponding waveband of the two data sources.
Preferably, the S5 includes the following steps:
designing a GAN-FC model aiming at a spectrum-biomass sample data set based on a simulation satellite, and adding a characteristic constraint factor in the process of generating a sample to ensure that the model controls the reasonability of the characteristic of the model while generating a pseudo sample set; the method for generating the sample set of each vegetation type based on the GAN-FC model comprises the following steps:
preprocessing the real sample data:
dividing real samples into S groups according to a rule (1), and respectively inputting the S groups of samples into a GAN-FC model; continuously training a generator and a discriminator of the GAN model based on input real samples and random noise to generate corresponding S groups of new samples; screening a new sample based on the characteristic constraint factors;
selecting a new sample meeting the constraint factor condition of the rule (2) as the finally generated sample data, and randomly selecting a certain number of samples from the finally generated sample data set as training samples for constructing a biomass inversion model;
the rule (1) refers to: based on the existing biomass data range of 0-5.29kg/m2(ii) a When the biomass is less than 1kg/m2When it is used, the amount of the additive is 0.1kg/m2The interval groups of (1); when the biomass is more than 1kg/m2When it is used, the amount of the additive is 0.5kg/m2The interval groups of (1); groups of less than 3 samples are merged into a group of the adjacent two groups having a smaller number of samples;
the rule (2) refers to: and constructing a constraint condition according to the characteristic constraint factors of the real samples: the spectral reflectance threshold increase Ps is 0.001, and the vegetation index threshold increase Pf is 0.003; and reserving the generated sample with the spectrum and vegetation index characteristics meeting the 2 characteristic constraints as an output sample, namely the sample data generated by the GAN-FC.
Preferably, the step 2 comprises the following specific steps:
establishing a multiple linear regression model by taking the preferred variable as an independent variable and the biomass on the vegetation ground as a dependent variable:
yi＝b0+b1xi1+b2xi2+…+bkxi3+εi(i＝1，2，3，...，n) (3)
wherein y is the aboveground biomass of the vegetation, x is the preferred variable for the vegetation, b is the regression parameter,. epsilon.iIs a random error, and n is the number of samples;
performing regression modeling by using statistical software SPSS19.0 based on the generated biomass sample set to obtain an aboveground biomass estimation model of each vegetation;
when the accuracy of the inversion model is measured, the on-site real biomass is compared with the biomass obtained by the inversion of the multiple linear regression model, and the accuracy of the model is evaluated by calculating the R, the mean absolute error (MAE, formula 4) and the root mean square error (RMSE, formula 5) of the two:
Figure BDA0003493830120000041
Figure BDA0003493830120000042
wherein, yiAnd y is the inverted biomass and the true biomass, respectively, of the ith sample, and n is the number of samples.
The coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method has the beneficial effects that:
the specific spectral characteristics and vegetation index characteristics of the coastal wetland vegetation are extracted, a plurality of characteristics are preferably selected as constraint factors for resisting generation of a network model, a GAN-FC model expanded aiming at the coastal wetland vegetation biomass sample is designed, and an inversion model is constructed based on the generated biomass sample to estimate the overground biomass of the coastal wetland vegetation. The biomass inversion result based on the generated sample is verified by using part of field data and is compared with the inversion result based on the real sample, and the experimental result proves that the sample generated based on the model can improve the inversion accuracy of the vegetation biomass of the coastal wetland, and the problems that the coastal wetland vegetation biomass inversion model is difficult to construct and the inversion accuracy is low due to the fact that the field biomass data are few are solved.
Drawings
FIG. 1 is a schematic diagram of a research area and site survey site distribution;
FIG. 2, a coastal wetland vegetation biomass sample generation model based on a GAN-FC model;
FIG. 3 is a result diagram of the biomass of the salt marsh vegetation in the wetland at the yellow river mouth;
FIG. 4 is a diagram showing the results of saline marsh vegetation biomass in Bay gulf of Jiaozhou.
Detailed Description
In the following, embodiments of the present invention are described in detail in a stepwise manner, which is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention, and any modifications, equivalents, improvements, etc. made within the spirit and principle of the present invention should be included in the scope of the present invention.
In the description of the present invention, it should be noted that the terms "upper", "lower", "left", "right", "top", "bottom", "inner", "outer", and the like indicate orientations and positional relationships based on the orientations and positional relationships shown in the drawings, and are only used for describing the present invention and simplifying the description, but do not indicate or imply that the referred device or element must have a specific orientation and a specific orientation configuration and operation, and thus, the present invention is not to be construed as being limited thereto.
Step 1, optimizing Feature constraint factors of a GAN network aiming at coastal wetland vegetation, and designing a generation countermeasure network model (GAN-FC) constrained by the Feature constraint factors:
s1, acquiring satellite multispectral remote sensing image data: the method utilizes multispectral remote sensing image data of a Sentinel-2 satellite covering a Shandong province Yingcity yellow river mouth coastal wetland in 2019, 24 months and 10 months and 14 days in 2019, covering a Shandong province Qingdao city Jiaozhou Bay wetland;
s2, measuring the spectral data of the vegetation canopy on site, manually collecting samples, calculating the biomass value of a unit area, and analyzing and processing the measured spectral data of the vegetation canopy by using spectrum processing software View Spec Pro, thereby obtaining the on-site hyperspectral data of the vegetation;
the method takes the field sampling data of the coastal wetland of the yellow river estuary in 2013 and the gulf wetland in 2019 as the manual sampling samples. The geographical location of the study area, the color composite image and the site are shown in fig. 1. The area of the yellow river mouth wetland research area is 2753km multiplied by 3072km, four types of swamp vegetations including reed, suaeda salsa, tamarix chinensis and Spartina alterniflora are grown, and the statistical result of field biomass sampling data is 246 reeds, 37 suaeda salsa, 39 tamarix chinensis mixed and 13 Spartina alterniflora, and the total number is 335 biomass samples. Three saline marsh vegetations of reed, suaeda salsa and spartina alterniflora are grown in a large area in the gulf wetland research area of the Jiaozhou, 3 areas where the spartina alterniflora is intensively distributed are selected for monitoring, and field biomass sampling data are all 15 samples of the spartina alterniflora. It can be seen that, except for the reeds, the sample data of the vegetation biomass is few, and most of the sample data are sampled in the edge zone of the vegetation community.
And on-site data acquisition is carried out under the conditions of good weather conditions and proper spectrum acquisition time. Recording the types of the stations and the ground objects nearby the stations, selecting a proper sample, recording the longitude and latitude coordinates of the sample by a handheld GPS, recording the vegetation growth parameters in the sample, measuring the spectral data of the vegetation canopy, and collecting the vegetation sample in the field sample. For herbaceous vegetation samples, the vegetation samples are collected in a 0.5m multiplied by 0.5m sample square in a uniform vegetation growing area, and the vegetation samples in 5 0.5m multiplied by 0.5m sample squares are randomly obtained in a 10m multiplied by 10m sample square in the non-uniform vegetation growing area. For brush vegetation, the crown width and the base diameter of the vegetation are obtained, and partial upper trunk and branch samples are collected. And cutting off the plants in the representative vegetation sample under the conditions in a flush manner, and marking the site information. The collected vegetation samples are bagged and brought back to the laboratory. Drying the plant sample for 24 hours at the constant temperature of 80 ℃ by using an oven to constant weight, weighing the dry weight, and calculating the biomass value of unit area in kg/m by combining with on-site survey record information2. And analyzing and processing the measured vegetation canopy spectrum data by using spectrum processing software View Spec Pro so as to obtain the field hyperspectral data of the vegetation.
TABLE 1 field data statistics of aboveground biomass of saline marsh plants
Figure BDA0003493830120000061
S3, extracting features, taking the extracted spectral features and vegetation index features as candidate variables, and screening the candidate variables based on correlation analysis:
extracting spectral characteristics and vegetation index characteristics of multispectral remote sensing image data as variables to be selected, wherein the screened characteristic variables serve as part of characteristic constraint factors of the GAN and are used for the biomass inversion of the vegetation in the coastal wetland, the inversion is reverse calculation, and in the specific implementation, 280 characteristic variables serve as the variables to be selected, wherein the 280 vegetation index characteristics comprise 270 vegetation index characteristics which are calculated based on 6 planting vegetation indexes (NDVI, DVI, RVI, SAVI, MSAVI and OSAVI) and comprise 10 spectral bands; based on the Pearson correlation coefficient R (formula (1)), candidate variables which are strongly correlated with biomass (the more the absolute value of R is close to 1) and weakly correlated between feature interiors (the more the absolute value of R is close to 0) are calculated as preferred variables (about 5-7), and then the preferred variables are used as partial feature constraint factors of the GAN generation sample and are used for constructing a vegetation biomass inversion model, wherein the formula (1) is as follows:
Figure BDA0003493830120000071
wherein R is Pearson's correlation coefficient, xi、yiThe values of independent variable and dependent variable at each sample point i are shown, and n is the number of sample points.
S4, performing equivalent wave band conversion on the remote sensing multispectral data and the field hyperspectral data;
establishing a spectrum conversion model between wetland vegetation field canopy hyperspectral data and Sentinel-2 satellite multispectral remote sensing image data, and making a spectrum-biomass sample data set simulating a Sentinel-2 satellite as basic data for coastal wetland vegetation biomass data expansion; the conversion operation is as follows:
(1) satellite band conversion: establishing conversion between a narrow wave band in on-site hyperspectral data and a wide wave band in satellite multispectral remote sensing image data, and calculating by adopting a formula (2):
Figure BDA0003493830120000072
where ρ is the reflectance of a satellite-simulated broadband, λmin、λmaxThe initial wavelength and the final wavelength of the spectral detection of the satellite sensor are respectively, S (lambda) is the spectral response function value of the Sentinel-2 sensor at the lambda wavelength, and rho (lambda) is the reflectivity of the wetland vegetation canopy spectrum at the lambda wavelength. Since the field measurement spectrum coverage (350-1050nm) corresponds to the coverage (400-1000nm) of the first 10 bands (band1-band9) of the sentinel 2A satellite, the simulated satellite data obtained by the method is 10 bands.
(2) Linear conversion among wide bands: and establishing a linear conversion relation between the simulated satellite spectrum and the satellite spectrum. Ideally, the reflectivity of the simulated satellite data obtained in the step (1) is highly similar to that of the real satellite data. Under the influence of uncertain factors such as a data acquisition mode, acquisition time, measurement conditions and the like, the reflectivity of the simulated satellite data and the satellite data may have certain deviation. In order to reduce the influence of deviation on subsequent biomass inversion work, a linear regression model is constructed by using simulated satellite data and Sentinel-2 data to perform reflectivity conversion on each corresponding waveband of two data sources.
S5, obtaining a generation countermeasure network model based on the feature constraint factor constraint:
aiming at a spectrum-biomass sample data set of the seaside wetland based on Sentinel-2, a GAN-FC model shown in figure 2 is designed, and a characteristic constraint factor is added in the process of generating the sample, so that the model controls the reasonability of the characteristic of the model while generating a pseudo sample set.
Specifically, the method firstly preprocesses real sample data, divides the samples into S groups according to a rule (1), and respectively inputs the S groups of samples into a GAN-FC model. Based on the input real samples and random noise, the generator and the discriminator of the GAN model are continuously trained, and corresponding S groups of new samples are generated. The new sample has certain similarity with the real sample, but in order to enable the new sample to be better used for biomass estimation, the invention screens the new sample based on the characteristic constraint factors.
In the implementation of the invention, 16-18 GAN characteristic constraint factors are used, including 10 spectral characteristics, 1 biomass characteristic and 5-7 vegetation index characteristics (preferred variables).
The invention extracts features from real samples as feature constraint factors to limit the generated samples. And (3) selecting a new sample meeting the constraint condition in the rule (2) as the finally generated sample data. A certain number of samples are randomly selected from the generated data set to be used as training samples for constructing a biomass inversion model, wherein 300 reed samples, suaeda salsa samples and tamarix chinensis samples are selected, 200 spartina alterniflora samples are selected, and the statistics of the generated samples participating in the inversion are described in a table 2.
Rule (1): the biomass data used according to the invention are in the range from 0 to 5.29kg/m2. When the biomass is less than 1kg/m2When it is used, the amount of the additive is 0.1kg/m2The interval groups of (1); when the biomass is more than 1kg/m2When it is used, the amount of the additive is 0.5kg/m2Are grouped. Groups of less than 3 samples are merged into a group of the two adjacent groups having a smaller number of samples.
Rule (2): according to 16-18 features of the real sample, the spectral reflectance threshold increase Ps is 0.001, the vegetation index threshold increase Pf is 0.003, and a generated sample meeting the feature constraint condition is taken as an output sample, namely sample data generated by GAN-FC.
TABLE 2 statistical results of the biomass sample data on the saline marsh vegetation land generated based on the generated countermeasure network
Figure BDA0003493830120000081
And 2, constructing a linear regression model by using biomass sample data generated by the GAN-FC model to invert the biomass on the vegetation ground.
And (3) establishing a multiple linear regression model (MLRM, formula (3)) by taking the preferred variable as an independent variable and the biomass on the vegetation ground as a dependent variable to estimate the biomass on the coastal wetland salt marsh vegetation ground.
yi＝b0+b1xi1+b2xi2+…+bkxi3+εi(i＝1，2，3，...，n) (3)
Wherein y is the biomass on the vegetation ground, x is the preferred variable of the vegetation, b is the regression parameter, epsiloniAnd n is the number of samples.
And performing regression modeling by using statistical software SPSS19.0 based on the generated biomass sample set to obtain an aboveground biomass estimation model of each vegetation. When the estimation accuracy of each model is measured, the field real data is compared with the biomass obtained by inversion, and the accuracy of the model is evaluated by calculating the Pearson correlation coefficient R, mean absolute error (MAE, formula 4) and root mean square error (RMSE, formula 5):
Figure BDA0003493830120000082
Figure BDA0003493830120000083
in the formula, yiAnd y are the inverted biomass and the true biomass, respectively, for the ith sample, and n is the number of samples.
To summarize: and (3) evaluation of biomass inversion results and precision of the coastal wetland:
according to the invention, under the same experimental conditions, the coastal wetland biomass inversion effects before and after the sample amplification are compared. The biomass inversion model before sample amplification is modeled based on a sample set constructed by field data, and biomass inversion result verification is carried out by utilizing part of field samples. And the biomass inversion model after sample amplification is modeled based on a sample set generated by the GAN-FC model, and biomass inversion result verification is performed by using the same field sample as that before sample amplification. The saline marsh vegetation inverted in the gulf wetland research area is spartina alterniflora. The saltwater vegetation inverted in the yellow river estuary wetland research area comprises 4 types of vegetation including reed, suaeda salsa, tamarix chinensis mixed growth and spartina alterniflora, and the precision evaluation results of biomass inversion model comparison of the saltwater vegetation are shown in a table 3. The various vegetation in the table comprises two rows, wherein the upper row is used for evaluating the precision of the biomass inversion model before sample expansion, and the lower row is used for evaluating the precision of the biomass inversion model after sample expansion. It can be seen that the biomass inversion error based on each type of sample generated is lower than the biomass inversion error based on the field samples. The estimation result of the biomass of the salt marsh vegetation of the wetland at the yellow estuary is shown in figure 3.
TABLE 3 Biomass inversion model accuracy evaluation before and after sample expansion based on GAN
Figure BDA0003493830120000091
The results of the evaluation of the inversion result precision of the gulf wetland are shown in table 4. It can be seen that the biomass inversion error based on the 200 spartina alterniflora samples generated is lower than that estimated based on the 15 samples limited in the field. And when the field samples are less, the biomass inversion result obtained based on the generated samples is more objective (R is 0.643). The estimation results of spartina alterniflora biomass in gulf of glue are shown in fig. 4.
TABLE 4 Biomass inversion model accuracy evaluation before and after sample expansion based on GAN
Figure BDA0003493830120000092
In conclusion, when the GAN-FC model is applied to the onshore biomass inversion of the saline marsh vegetation of the coastal wetland and the gulf wetland of the yellow estuary, the inversion accuracy is improved compared with that obtained by limited onsite biomass samples. The accuracy evaluation of the inversion result based on the limited field biomass samples is as follows: reed RMSE is 0.53kg/m2Suaeda salsa RMSE 0.15kg/m2The RMSE of tamarisk mixed growth is 0.35kg/m2The RMSE of the spartina alterniflora is 0.46kg/m2(ii) a Wet sensation in gulf of jiao ZhouGround spartina alterniflora RMSE 0.11kg/m2. The accuracy evaluation of the inversion result obtained based on the sample generated by the model of the invention is as follows: reed RMSE of the yellow estuary wetland is 0.52kg/m2Suaeda salsa RMSE is 0.16kg/m2The RMSE of tamarisk mixed growth is 0.29kg/m2The Mesempervirens RMSE is 0.37kg/m2(ii) a The RMSE of the spartina alterniflora of Bay wetland in Jiaozhou is 0.05kg/m2. Therefore, the effectiveness and feasibility of the coastal wetland vegetation overground biomass inversion of the samples generated by the method under the condition of few samples are verified.

Claims (8)

1. The coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method is characterized by comprising the following steps: the method comprises the following steps:
step 1, optimizing feature constraint factors of a GAN network aiming at the vegetation of the coastal wetland, and designing a generation confrontation network model constrained by the feature constraint factors;
and 2, constructing a linear regression model by using the biomass sample data generated by the designed generation confrontation network model to invert the biomass on the vegetation ground.
2. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 1, which is characterized in that: the step 1 comprises the following specific steps:
s1, acquiring satellite multispectral remote sensing image data;
s2, measuring the spectral data of the vegetation canopy on site, manually collecting samples, analyzing and processing the measured spectral data of the vegetation canopy by using spectrum processing software View Spec Pro, thereby obtaining the on-site hyperspectral data of the vegetation, drying and weighing in a laboratory, and calculating the terrestrial biomass value of a unit area;
s3, extracting features, screening feature variables based on correlation analysis, and determining feature constraint factors;
s4, performing equivalent wave band conversion on the remote sensing multispectral data and the field hyperspectral data;
and S5, obtaining a generation countermeasure network model based on the characteristic constraint factor constraint.
3. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 2, which is characterized in that: in S2, the field sample is collected when the weather condition is good and the spectrum collection time is appropriate, and the collection method includes:
recording the vegetation type of a field data acquisition site, selecting a proper sample, recording the longitude and latitude coordinates of the sample by a handheld GPS, recording the vegetation growth parameters in the sample, measuring the spectrum of a vegetation canopy, and acquiring a vegetation sample in the field sample;
for the herbaceous vegetation sample, collecting the vegetation sample by taking 0.5m multiplied by 0.5m as a sample in an area with uniform vegetation growth, and randomly obtaining 5 vegetation samples in 0.5m multiplied by 0.5m sub-sample in 10m multiplied by 10m in the area with non-uniform vegetation growth; for shrub vegetation, acquiring the crown width and the base diameter of the vegetation, and collecting partial upper trunk and branch samples;
cutting off plants in representative vegetation samples in the condition in a field, and marking site information;
bagging the collected vegetation samples and taking the vegetation samples back to a laboratory, drying the vegetation samples for 24 hours at a constant temperature of 80 ℃ by using an oven to constant weight, then weighing the dry weight of the vegetation samples, and calculating the biomass value of unit area in kg/m by combining with on-site survey record information2。
4. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 3, which is characterized in that: the S3 includes the following steps:
and S31, taking the spectral characteristics and vegetation index characteristics of the multispectral data simulated by the on-site hyperspectral data as variables to be selected, screening the characteristics by a correlation analysis method, and forming a characteristic constraint factor of the GAN model by the screened preferred variables, the spectral characteristics and the biomass for constraining the biomass sample generated by the model.
5. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 4, which is characterized in that: in the step S31, 10 spectral bands in the spectral features are combined pairwise to be brought into 6 vegetation indexes, 270 vegetation index feature variables are obtained through calculation, 280 feature variables of 10 spectral features and 270 vegetation index feature variables are used as variables to be selected, and 6 vegetation indexes refer to NDVI, DVI, RVI, SAVI, msaiv and OSAVI; calculating variables with strong correlation of biomass characteristics and weak correlation among the characteristics in the variables to be selected as preferred variables based on a calculation formula (1) of a Pearson correlation coefficient R; wherein the closer the absolute value of R is to 1, the stronger the correlation; the closer the absolute value of R is to 0, the weaker the correlation is; and (3) taking the preferred variable as a part of characteristic constraint factor of the GAN generation sample, and using the preferred variable for constructing a vegetation biomass inversion model:
Figure FDA0003493830110000021
wherein R is Pearson's correlation coefficient, xiAnd yiThe values of the independent variable and the dependent variable at each sampling point i are shown, and n is the number of the sampling points.
6. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 5, which is characterized in that: the S4 includes the following steps:
making a spectrum-biomass sample data set of a simulation satellite through a spectrum conversion model between on-site hyperspectral data of vegetation and satellite multispectral remote sensing image data, and using the spectrum-biomass sample data set as basic data for vegetation biomass data expansion of the coastal wetland; the conversion operation is as follows:
(1) satellite band conversion: establishing conversion between a narrow wave band in the on-site hyperspectral data and a wide wave band in the satellite multispectral remote sensing image data, and calculating by adopting a formula (2):
Figure FDA0003493830110000022
where p is the reflectivity of the broadband of the simulated satellite,λmin、λmaxthe initial wavelength and the final wavelength of the spectrum detection of the satellite sensor are respectively, S (lambda) is a spectral response function value of the satellite sensor at the lambda wavelength, and rho (lambda) is the reflectivity of the spectrum of the wetland vegetation canopy at the lambda wavelength;
(2) linear conversion among wide bands: and establishing a linear conversion relation between the simulated satellite spectrum and the satellite spectrum, and constructing a linear regression model by using the simulated satellite data and the satellite data to perform reflectivity conversion on each corresponding waveband of the two data sources.
7. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 6, which is characterized in that: the S5 includes the following steps:
designing a GAN-FC model aiming at a spectrum-biomass sample data set based on a simulation satellite, and adding a characteristic constraint factor in the process of generating a sample to ensure that the model controls the reasonability of the characteristic of the model while generating a pseudo sample set; the method for generating the sample set of each vegetation type based on the GAN-FC model comprises the following steps:
preprocessing the real sample data:
dividing real samples into S groups according to a rule (1), and respectively inputting the S groups of samples into a GAN-FC model; continuously training a generator and a discriminator of the GAN model based on input real samples and random noise to generate corresponding S groups of new samples; screening a new sample based on the characteristic constraint factors;
selecting a new sample meeting the constraint factor condition of the rule (2) as the finally generated sample data, and randomly selecting a certain number of samples from the finally generated sample data set as training samples for constructing a biomass inversion model;
the rule (1) refers to: based on the existing biomass data range of 0-5.29kg/m2(ii) a When the biomass is less than 1kg/m2When it is used, the amount of the additive is 0.1kg/m2The interval groups of (1); when the biomass is more than 1kg/m2When it is used, the amount of the additive is 0.5kg/m2The interval of (2) is grouped; groups of less than 3 samples are merged into a group of the adjacent two groups having a smaller number of samples;
the rule (2) refers to: and constructing a constraint condition according to the characteristic constraint factors of the real samples: the spectral reflectance threshold increase Ps is 0.001, and the vegetation index threshold increase Pf is 0.003; and reserving the generated sample with the spectrum and vegetation index characteristics meeting the 2 characteristic constraints as an output sample, namely the sample data generated by the GAN-FC.
8. The coastal wetland vegetation above-ground biomass GAN model self-learning remote sensing inversion method of claim 7, which is characterized in that: the step 2 comprises the following specific steps:
establishing a multiple linear regression model by taking the preferred variable as an independent variable and the biomass on the vegetation ground as a dependent variable:
yi＝b0+b1xi1+b2xi2+…+bkxi3+εi(i＝1，2，3，…，n) (3)
wherein y is the aboveground biomass of the vegetation, x is the preferred variable for the vegetation, b is the regression parameter,. epsilon.iIs a random error, and n is the number of samples;
performing regression modeling by using statistical software SPSS19.0 based on the generated biomass sample set to obtain an aboveground biomass estimation model of each vegetation;
when the accuracy of the inversion model is measured, the on-site real biomass is compared with the biomass obtained by the inversion of the multiple linear regression model, and the accuracy of the model is evaluated by calculating the R, the mean absolute error (MAE, formula 4) and the root mean square error (RMSE, formula 5) of the two:
Figure FDA0003493830110000031
Figure FDA0003493830110000032
wherein, yiAnd y is the biomass and true of the inversion of the ith sample, respectivelyAnd n is the number of samples.
CN202210107317.XA 2022-01-28 2022-01-28 Coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method Active CN114460013B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210107317.XA CN114460013B (en) 2022-01-28 2022-01-28 Coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210107317.XA CN114460013B (en) 2022-01-28 2022-01-28 Coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method

Publications (2)

Publication Number Publication Date
CN114460013A true CN114460013A (en) 2022-05-10
CN114460013B CN114460013B (en) 2023-10-17

Family

ID=81411680

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210107317.XA Active CN114460013B (en) 2022-01-28 2022-01-28 Coastal wetland vegetation overground biomass GAN model self-learning remote sensing inversion method

Country Status (1)

Country Link
CN (1) CN114460013B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116842351A (en) * 2023-09-01 2023-10-03 山东省海洋资源与环境研究院(山东省海洋环境监测中心、山东省水产品质量检验中心) Coastal wetland carbon sink assessment model construction method, assessment method and electronic equipment
CN117456351A (en) * 2023-10-08 2024-01-26 宁波大学 Method for estimating carbon reserves of salt biogas vegetation of coastal wetland by cooperation of starry sky and land
CN117494066A (en) * 2023-11-15 2024-02-02 黑龙江省网络空间研究中心(黑龙江省信息安全测评中心、黑龙江省国防科学技术研究院) Long-time sequence overground biomass inversion quantitative evaluation method
CN117671521A (en) * 2024-02-02 2024-03-08 中交四航工程研究院有限公司 Invasive species biomass inversion method and device based on multi-source remote sensing data

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948693A (en) * 2019-03-18 2019-06-28 西安电子科技大学 Expand and generate confrontation network hyperspectral image classification method based on super-pixel sample
CN110991430A (en) * 2020-03-02 2020-04-10 中科星图股份有限公司 Ground feature identification and coverage rate calculation method and system based on remote sensing image
CN111428678A (en) * 2020-04-02 2020-07-17 山东卓元数据技术有限公司 Method for generating confrontation network remote sensing image sample expansion under space constraint condition for ground object change detection
CN111488902A (en) * 2020-01-14 2020-08-04 沈阳农业大学 Method and system for quantitatively estimating carbon reserves of ecosystem of primary coastal wetland
CN111597751A (en) * 2020-03-24 2020-08-28 自然资源部第一海洋研究所 Crude oil film absolute thickness inversion method based on self-expansion depth confidence network
CN111626947A (en) * 2020-04-27 2020-09-04 国家电网有限公司 Map vectorization sample enhancement method and system based on generation of countermeasure network
CN111814707A (en) * 2020-07-14 2020-10-23 中国科学院空天信息创新研究院 Crop leaf area index inversion method and device
CN111860640A (en) * 2020-07-17 2020-10-30 大连海事大学 Specific sea area data set augmentation method based on GAN
WO2021226976A1 (en) * 2020-05-15 2021-11-18 安徽中科智能感知产业技术研究院有限责任公司 Soil available nutrient inversion method based on deep neural network

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948693A (en) * 2019-03-18 2019-06-28 西安电子科技大学 Expand and generate confrontation network hyperspectral image classification method based on super-pixel sample
CN111488902A (en) * 2020-01-14 2020-08-04 沈阳农业大学 Method and system for quantitatively estimating carbon reserves of ecosystem of primary coastal wetland
CN110991430A (en) * 2020-03-02 2020-04-10 中科星图股份有限公司 Ground feature identification and coverage rate calculation method and system based on remote sensing image
CN111597751A (en) * 2020-03-24 2020-08-28 自然资源部第一海洋研究所 Crude oil film absolute thickness inversion method based on self-expansion depth confidence network
CN111428678A (en) * 2020-04-02 2020-07-17 山东卓元数据技术有限公司 Method for generating confrontation network remote sensing image sample expansion under space constraint condition for ground object change detection
CN111626947A (en) * 2020-04-27 2020-09-04 国家电网有限公司 Map vectorization sample enhancement method and system based on generation of countermeasure network
WO2021226976A1 (en) * 2020-05-15 2021-11-18 安徽中科智能感知产业技术研究院有限责任公司 Soil available nutrient inversion method based on deep neural network
CN111814707A (en) * 2020-07-14 2020-10-23 中国科学院空天信息创新研究院 Crop leaf area index inversion method and device
CN111860640A (en) * 2020-07-17 2020-10-30 大连海事大学 Specific sea area data set augmentation method based on GAN

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
ONISIMO MUTANGA, ET AL: "High density biomass estimation for wetland vegetation using WorldView-2 imagery and random forest regression algorithm", INTERNATIONAL JOURNAL OF APPLIED EARTH OBSERVATION AND GEOINFORMATION, pages 399 - 406 *
张国飞;岳彩荣;章皖秋: "基于Sentinel数据的滇池湖滨湿地地上生物量反演", 太赫兹科学与电子信息学报, vol. 18, no. 001, pages 142 - 149 *
牟蒙: "基于高光谱遥感的翅碱蓬生物量反演模型研究", 中国优秀硕士学位论文全文数据库基础科学辑, pages 12 - 13 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116842351A (en) * 2023-09-01 2023-10-03 山东省海洋资源与环境研究院(山东省海洋环境监测中心、山东省水产品质量检验中心) Coastal wetland carbon sink assessment model construction method, assessment method and electronic equipment
CN116842351B (en) * 2023-09-01 2023-11-10 山东省海洋资源与环境研究院(山东省海洋环境监测中心、山东省水产品质量检验中心) Coastal wetland carbon sink assessment model construction method, assessment method and electronic equipment
CN117456351A (en) * 2023-10-08 2024-01-26 宁波大学 Method for estimating carbon reserves of salt biogas vegetation of coastal wetland by cooperation of starry sky and land
CN117456351B (en) * 2023-10-08 2024-05-17 宁波大学 Method for estimating carbon reserves of salt biogas vegetation of coastal wetland by cooperation of starry sky and land
CN117494066A (en) * 2023-11-15 2024-02-02 黑龙江省网络空间研究中心(黑龙江省信息安全测评中心、黑龙江省国防科学技术研究院) Long-time sequence overground biomass inversion quantitative evaluation method
CN117671521A (en) * 2024-02-02 2024-03-08 中交四航工程研究院有限公司 Invasive species biomass inversion method and device based on multi-source remote sensing data

Also Published As

Publication number Publication date
CN114460013B (en) 2023-10-17

Similar Documents

Publication Publication Date Title
CN109884664B (en) Optical microwave collaborative inversion method and system for urban overground biomass
CN114460013A (en) Coastal wetland vegetation ground biomass GAN model self-learning remote sensing inversion method
Yuan et al. Estimating surface soil moisture from satellite observations using a generalized regression neural network trained on sparse ground-based measurements in the continental US
Haboudane et al. Remote estimation of crop chlorophyll content using spectral indices derived from hyperspectral data
CN112395808A (en) Biomass remote sensing mapping method combining random forest and collaborative kriging
Puente et al. A genetic programming approach to estimate vegetation cover in the context of soil erosion assessment
CN113205014B (en) Time sequence data farmland extraction method based on image sharpening
GB2620469A (en) Spatial prediction and evaluation method of soil organic matter content based on partition algorithm
Cao et al. Estimation of forest structural parameters using UAV-LiDAR data and a process-based model in ginkgo planted forests
CN114005048A (en) Multi-temporal data-based land cover change and thermal environment influence research method
CN116665073A (en) Corn yield remote sensing estimation method based on multi-source data
Xu et al. A 21-year time series of global leaf chlorophyll content maps from MODIS imagery
CN113011372A (en) Automatic monitoring and identifying method for saline-alkali soil
Hu et al. Integrating CART algorithm and multi-source remote sensing data to estimate sub-pixel impervious surface coverage: a case study from Beijing Municipality, China
CN112800827A (en) Hyperspectral image classification experimental method
CN117437538A (en) Tropical rainforest ecosystem space-time pattern feature extraction and prediction method
CN116562161A (en) Inversion model construction method for black soil corrosiveness
Guo et al. Estimating aboveground biomass using Pléiades satellite image in a karst watershed of Guizhou Province, Southwestern China
Bao et al. A fine digital soil mapping by integrating remote sensing-based process model and deep learning method in Northeast China
CN115830464A (en) Plateau mountain agricultural greenhouse automatic extraction method based on multi-source data
Sui et al. Disentangling error structures of precipitation datasets using decision trees
Wang et al. LAI inversion from MODIS data using deep belief network (DBN)
Atzberger et al. Estimation of inter-annual winter crop area variation and spatial distribution with low resolution NDVI data by using neural networks trained on high resolution images
Zakeri et al. High-Resolution Snow Water Equivalent Estimation: A Data-Driven Method for Localized Downscaling of Climate Data
SINCHANA et al. A Review on Land Suitability Mapping Using Geospatial Technology.

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