CN1677085A - Agricultural Application Integrated System and Method of Earth Observation Technology - Google Patents

Abstract

The invention discloses an integrated system and method for agricultural application of earth observation technology. The system includes: a data storage unit, including: a spectral database, a remote sensing database, and a basic database; a model unit, including a parameter selection model library and an agricultural situation feedback a model library; a control and operation processing unit, used to use the model of the model unit to process the image data; and an information output unit, used to display or output relevant information in other ways according to the extracted agricultural parameters. Agricultural information. The invention can utilize advanced earth observation technology to carry out precise modern management of agriculture. The method includes: obtaining hyperspectral data; processing the obtained hyperspectral data; performing data band selection on the hyperspectral data; performing feature extraction on the hyperspectral data after band selection to obtain agricultural parameters; The acquired agricultural parameters are used to output relevant agricultural information in the hyperspectral data.

Description

Agricultural Application Integrated System and Method of Earth Observation Technology

technical field

The present invention relates to the application technology of information resources, in particular to the application technology of providing auxiliary information for agricultural applications such as agricultural experiments, crop type identification and agricultural condition diagnosis by using information obtained by imaging spectrum technology.

Background technique

Imaging spectroscopy technology is currently the cutting-edge technology in the field of earth observation. Since it can acquire continuous spectra of surface vegetation, soil, water bodies and other ground objects for analyzing their physical and chemical processes, it has great application potential in agricultural applications.

The civil aerospace remote sensors that have been in operation, such as Landsat TM, SPOT and NOAA, generally only have 5-6 bands, and the spectral resolution is above 50nm. It is difficult to identify multiple crop types. The peak width of the main factors of vegetation is about 20nm, and the redshift component of vegetation damage stress is 5-17nm. These phenomena are difficult to be detected by low spectral resolution remote sensors. The superiority of aerospace or aeronautical imaging spectral remote sensors is that they can be subdivided into dozens or hundreds of bands within the 0.4-14μm spectral range, and the spectral resolution is 5-10nm. In this way, the ability to identify crops can be improved, and the growth and change information of crops can be monitored at the same time, which is conducive to the fine management of crops.

my country is a large agricultural country in the world, and it is also one of the few countries that has mastered imaging spectroscopy technology. At present, how to select the spectral parameters that are useful for agriculture in the hyperspectral data measured by aerial and satellite imaging spectrometers, how to process imaging spectral data, and how to extract useful agricultural information are key technical issues that need to be solved urgently.

Contents of the invention

Therefore, the object of the present invention is to provide an integrated system and method for the agricultural application of earth observation technology, which can process the imaging data by selecting appropriate spectral parameters from the data obtained from the imaging spectrum and other aerial remote sensing data. to extract useful agricultural information.

In order to achieve the above object, the present invention provides an integrated method for the agricultural application of earth observation technology, including: obtaining earth observation data; performing data band selection on the obtained earth observation data; Feature extraction is performed on the earth observation data to obtain agricultural condition parameters; and feature analysis is performed using the acquired agricultural condition parameters to obtain required information in the earth observation data.

In a preferred embodiment of the present invention, the earth observation data is hyperspectral data, and the above method further includes a step of preprocessing the hyperspectral data.

On the other hand, the present invention also provides an integrated system for agricultural applications of earth observation technology, including: a data storage unit, including:

Spectral database, which stores the spectral curves of different crops or ground objects in the form of vector graphics data;

The remote sensing database stores the multi-band remote sensing images acquired by the airborne imaging spectrometer in the form of raster image data;

The basic database is used to store other auxiliary geospatial data (graphics and image data in raster or vector form) and attribute data (statistical data in text or table form) that match the remote sensing images of the study area. Such as vectorized regional administrative boundaries, meteorological elements, land cover type maps, etc.;

The model unit includes a parameter selection model library and an agricultural situation inversion model library. The parameter selection model library is used for spectral reconstruction of images, data composite feature analysis and band selection. The agricultural situation inversion model library is used to provide multiple Various types of agricultural information models;

a control and arithmetic processing unit, configured to use the model of the model unit to process the image data accordingly; and

The information output unit is used to display or output relevant agricultural information according to the extracted agricultural condition parameters.

The invention can utilize advanced earth observation technology to provide abundant and accurate information for agricultural production, experiment and management.

Description of drawings

Fig. 1 is the schematic flowchart of the agricultural application method of the earth observation technology of an embodiment of the present invention;

Fig. 2 is a schematic flowchart of packet conversion and feature selection of an embodiment of the method of the present invention;

Fig. 3 is the schematic flow diagram of the application of the method of the present invention in wheat growth classification evaluation;

Fig. 4 is the schematic flow diagram of the application of the method of the present invention in wheat leaf area coefficient extraction;

Fig. 5 is the schematic block diagram of the structure of the agricultural application integrated system of the earth observation technology of the present invention;

Fig. 6 is a schematic diagram of the process of establishing a model library in the system of the present invention

Fig. 7 shows the process of land use classification in the system of the present invention.

.

Detailed ways

As shown in Figure 1, the present invention provides a method for agricultural application of earth observation technology, including: acquiring hyperspectral data; processing the obtained hyperspectral data; performing data band selection on the hyperspectral data; Extract the agricultural conditions parameters in the hyperspectral data, and display relevant agricultural information.

In the preferred embodiment of the present invention, spectral data, basic data and remote sensing data are required to be used.

Spectral data refers to the spectral curves of different crops or ground objects, that is, the curves formed by the reflectivity of crops or ground objects to electromagnetic waves in different bands, which are obtained through spectral measurement and remote sensing images of airborne imaging spectrometers.

It should be noted that the real spectrum of ground features is obtained by using the spectrometer to measure the reflectance on the ground; the remote sensing data directly obtains the radiance value of the ground features (that is, how much the ground features reflect and radiate energy), not the reflectance . Spectral reconstruction is the process of converting the radiance value of remote sensing images into reflectance, and then obtaining the spectral curve of ground objects.

Hyperspectral data (image) is a type of remote sensing data (image). Its main feature is that it can distinguish spectra with very fine wavelength intervals (such as 10nm). Because it covers many spectral segments, it is called hyperspectral. Imaging spectroscopy is a hyperspectral technique. The remote sensing data used in the present invention is imaging spectrum data.

Remote sensing data: refers to the multi-band remote sensing images acquired by the airborne imaging spectrometer. Stored as raster image data.

Basic data: Refers to other auxiliary geospatial data (graphics and image data in raster or vector form) and attribute data (statistical data in text or table form) that match the remote sensing images of the study area. Such as vectorized regional administrative boundaries, meteorological elements, land cover type maps, etc.

The basic data provide background data and information related to the research area; the remote sensing database provides multi-band image data of ground objects; the spectral database contains the characteristic spectral curves of specific ground objects selected on the basis of the former two; usually , different ground objects have their own specific spectral curves.

Hyperspectral data can be obtained using airborne aerial imaging spectrometers, airborne (shipborne) aerospace imaging spectrometers, etc. In some embodiments, in order to better realize the present invention, devices such as a visible-near-infrared intelligent spectrometer, a visible-near-infrared spectrometer, an infrared radiometer, and a GPS portable camera should also be used to obtain ground information. Of course, such information can also be provided through other sources, for example, obtained from other information providers through the Internet.

After obtaining the hyperspectral data, the image needs to be preprocessed. Such preprocessing may include radiometric corrections, geometric corrections, etc., as desired. In addition, spectral enhancement, spectral identification, grouping KL conversion\image classification and other processing can also be performed.

Further, spectral reconstruction, multi-source data compounding, and band selection processing can also be performed on the image. Among them, the role of spectral reconstruction is to obtain the spectral curves of ground objects through remote sensing images, and these spectral curves can be used for image classification and ground object recognition. Multi-source data composition is used to extract parameter features. These processes will be described in more detail later.

According to an embodiment of the present invention, image transformation may also be performed after spectral reconstruction.

Image transformation refers to the process of converting an image from the spatial domain to the transform domain. The purpose of image transformation is to simplify the image processing process. Through transformation, the vector has some better properties in the new space, which is more conducive to the analysis and solution of the problem.

Feature selection is generally done by linear transformation. The expression of the linear transformation is:

Y＝AX

In the formula, X is the n-dimensional random vector before transformation, Y is the n-dimensional random vector after transformation, A is an n×n transformation matrix, and the difference of A determines the different nature of the transformation.

The basic idea of feature selection through image transformation is: select a subset of m components of Y in the transform domain (1≤m<n), when deleting the remaining n-m components and only use the retained When m components represent X, the error caused is the smallest. In this way, the compression from n-dimension to m-dimension is realized. The size of the error is generally measured by the mean square error criterion. The mean square error criterion is: retain m component subsets with the largest variance, and delete the remaining n-m components.

The present invention adopts the method of grouping KL transformation. In order to explain the grouped KL transform, the KL transform will be explained first.

KL transform is an orthogonal linear transform, and it is one of the most commonly used and useful transform algorithms in remote sensing digital image processing. It is an effective method for decorrelation, feature extraction, and data compression.

For the convenience of derivation, the expression of linear transformation is written as:

$Y=\mathrm{AX}=\underset{i=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{A}_i{x}_i$ (Formula 2)

It is hoped that after the KL transformation, the original n-dimensional information can be reflected in the new space with only the first m-dimensional vector under the condition of minimum error. Now divide Y into two parts, the first m items are the first part, and the latter n-m items are the second part, then:

$Y=\mathrm{AX}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{A}_i{x}_i+\underset{i=m+1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{A}_i{x}_i$ (Formula 3)

Record x _i in the second part of the formula as b _i and record

$Y\left(m\right)=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{A}_i{x}_i+\underset{i=m+1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{A}_i{b}_i$ (Formula 4)

Let the error between Y and Y(m) be ε, then:

$\mathrm{\&epsiv;}=\underset{i=m+1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{A}_i(x_i-b_i)$ (Formula 5)

Its mean square error is:

$\stackrel{\&OverBar;}{{\mathrm{\&epsiv;}}^2}\left(k\right)=\underset{i=m+1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}\mathrm{E.}{\left[A_i(x_i-b_i)\right]}^2$ (Formula 6)

In order to minimize the mean square error, the derivative of _bi is:

b _i =E{ _xi }=m _x (Equation 7)

For a new random variable Y, to take the first m items and omit the following n-m items, it should be expressed as

y _i ＝A _i {xi _{-m x} ), that is, Y＝A(Xm _x ) (Formula 8)

The covariance of the new random variable Y ∑ _y =E{(Ym _y )(Ym _y ) ^t }, where m _y is the mean value of Y, and the transformed vector is a random vector with zero mean value, so my _y =0, Hence:

∑ _y ＝E{YY ^t }＝E{[A(Xm _x )][A(Xm _x )] ^t }＝A∑ _x A ^t

This shows that the covariance matrix of the random variable X is diagonalized, and the resulting diagonal matrix is the covariance moment Σ _y of the new random variable Y. Each element in the diagonal matrix is an eigenvalue of ∑ _x . Arrange the eigenvalues in order of magnitude, λ ₁ >λ ₂ >λ ₃ ...λ _n . In this way, when we only take the first m items and omit the items from m+1 to n, we can make

$\stackrel{\&OverBar;}{{\mathrm{\&epsiv;}}^2}\left(k\right)=\underset{i=m+1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i$ Min (Equation 9)

It can be seen that if the transformation matrix A is an orthogonal matrix and it is composed of the eigenvectors of the covariance matrix Σ _x of the original image data matrix X, the transformation is KL transformation.

To perform KL transformation, firstly, calculate its covariance matrix ∑ _x according to the original image matrix X; secondly, by the characteristic equation

(λI-∑ _x )u＝0 (Formula 10)

Wherein λ is an eigenvalue, I is an identity matrix, and u is an eigenvector, and each eigenvalue λ _i =(i=1, 2, ..., n) of the covariance matrix Σ _x is obtained, and it is pressed by λ ₁ ≥ λ ₂ ≥... _≥λn arrangement, find the eigenvector u corresponding to each eigenvalue ₁

u _i =[u _1i , u _2i ..., u _ni ] ^t (Formula 11)

Then, take the transformation matrix A= ^UT to obtain the concrete expression of KL transformation:

$\mathrm{the\; y}=\left[\begin{array}{cccc}{u}_{11}& u_{12}& \&Center\; Dot;\&CenterDot;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& u_{1\mathrm{no}}\\ u_{\mathrm{twenty\; one}}& u_{\mathrm{twenty\; two}}& \&Center\; Dot;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;& u_{2\mathrm{no}}\\ & & \&Center\; Dot;& \\ & & \&CenterDot;& \\ & & \&Center\; Dot;& \\ u_{\mathrm{no}1}& u_{\mathrm{no}2}& \&Center\; Dot;\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;& u_n\end{array}\right]x=u^{t}x$ (Formula 12)

After KL transformation, a new set of variables (that is, each row vector of Y) is obtained, which are called the first principal component, the second principal component, ..., the nth principal component in turn. Because the rows of the transformation matrix are the eigenvectors of ∑ _x , the i-th component of Y is actually the weighted sum of each component of X with the weight of each component of the i-th eigenvector, and each component of the entire Y is the weighted sum of each component of X. The linear combination of the information of the components, which synthesizes the information of the original features instead of simply choosing, makes the new n-dimensional random vector reflect the characteristics of the original things well.

From the principle of KL transform, it can be seen that it is the best orthogonal transform in the sense of minimum mean square error, and it has the following characteristics:

·Since the KL transformation is an orthogonal linear transformation, the sum of the variance before and after the transformation remains unchanged, but the original variance is redistributed to the new principal component image in an unequal amount.

Geometrically, the KL transformation is equivalent to the rotation of the spatial coordinates. The first principal component takes the direction where the data is most concentrated in the spectral space, and the second principal component takes the direction that is orthogonal to the first principal component and takes the direction where the data is scattered sub-concentrated. , and so on. Therefore, the first principal component contains most of the total variance (generally more than 80%), and the variance is consistent with the amount of information, so the result of KL transformation makes the first principal component almost contain the absolute part of the original image information of each band. Most of the information contained in the remaining principal components decreases rapidly in turn.

· KL transform is an effective method to decorrelate and eliminate data redundancy. In the original space, the components are oblique to each other and have a large correlation. After the KL transformation, the components in the new space are orthogonal and independent, and the correlation coefficient is zero, and because the information is concentrated in the first few In terms of components, under the premise of minimum information loss, fewer components can be used to replace the original high-dimensional data, achieving the effect of dimensionality reduction, which greatly reduces the time and cost of data processing. On the other hand, since the principal components are perpendicular to each other, the class distance is increased, the intra-class difference is reduced, and the classification accuracy can be improved.

It should be pointed out that although the first few principal components often contain more than 98% of the information, it cannot be simply considered that the latter principal components are useless, and sometimes the information contained in the principal components is just the required information. , so the choice of principal components should be based on specific application goals and specific analysis should be done.

The packet KL transformation adopted by the present invention is introduced below.

For conventional remote sensing data, feature selection is relatively easy due to its small number of bands. But for imaging spectral data, due to the sharp increase in the number of bands, feature extraction is difficult to achieve with conventional methods.

It is still difficult to apply KL transform directly to imaging spectral data. The KL transformation mainly includes two steps. The first step is to generate a transformation matrix, and the second step is to use the transformation matrix to transform the image. The first step does not require a large amount of calculation, but the second step is to use the transformation matrix to operate on each pixel, which is a very time-consuming process, because the calculation amount for each pixel is N×N multiplications and N x (N-1) additions. Therefore, if the transformation is applied to all bands of imaging spectral data, the efficiency can be imagined to be inefficient. Due to the high spectral resolution of imaging spectral data, it is more obviously affected by the solar absorption spectrum, and this effect is different for different wavelength ranges, and decreases with the increase of wavelength. Therefore, if the imaging spectral data is not radiometrically corrected, or the radiometric correction is not ideal (this is often the case), then the radiation distortion will affect the variance of the band, and the variance of the band affected by the radiation distortion is often higher than that of the band affected by the radiation distortion. It affects the variance of large bands, and the size of the variance will affect the result of KL transformation, so it is not suitable to directly use KL transformation for imaging spectral data.

Through the correlation analysis of imaging spectral data, we understand the characteristics of its band grouping, that is to say, the correlation between adjacent bands is very strong. With a certain spectral range as the limit, several bands are highly correlated within a group and relatively independent between groups. band group. The grouped KL transformation is based on the recognition of the characteristic of imaging spectral data. Change the method of traditional KL transformation to transform all data, but focus on each relatively independent band group. The process is shown in Figure 2. First, the correlation analysis is performed on the original data, a threshold value is set, and they are divided into K band groups according to the correlation coefficient between the bands, and the number of bands in each group is n ₁ , n ₂ ,...n _k ; secondly, perform KL transformation on each band group separately to generate principal component images of each group; then, perform feature selection. If there are many characteristic parameters selected, the above steps can be repeated to perform principal component transformation and feature selection on the selected characteristic parameters until the specific requirements are met.

Grouped KL transformation has obvious advantages over conventional KL transformation. As mentioned earlier, for conventional KL transformation, the amount of multiplication of each pixel point is N×N. Using grouped KL transformation, each point in each group The multiplication operation of is n _k ×n _k ,

The total multiplication operation amount of K groups is The ratio of the multiplication operations of the two different methods is $\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{no}}_k^2/(N\&times;N)=\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{\left(\frac{{\mathrm{no}}_k}{N}\right)}^2,$ That is, it is proportional to the square of the ratio between the number of bands in the group and the number of bands in the original image. The smaller the number of bands _nk in the group, the smaller the multiplication ratio above, and the more time saved.

Assuming that it is divided into three groups, and each group has the same size (ie K=3, n ₁ −n ₂ =n ₃ ), then only considering the amount of multiplication, 2/3 of the time can be saved. At the same time, the results of grouping KL transformation processing are more reasonable. The reason why bands can form highly correlated groups is that they are inherently related. They are located in specific spectral segments and have some common similar properties, that is, consistency. The influence of the solar spectrum on the bands of different spectral ranges is different, which can cause abnormal variance and thus affect the results of KL transformation. However, in groups with relatively narrow spectral ranges, this effect is relatively consistent, so grouping KL transformation can be avoided. This error yields more reasonable results.

Preferably, the method of the present invention further includes the steps of hyperspectral image processing, including: image enhancement, spectrum enhancement, spectrum identification and image classification. These methods can all adopt well-known technologies in the art, for example, refer to "Earth Observation Technology and Precision Agriculture" edited by Wang Changyao et al. (Spatial Information Acquisition and Processing Series Monographs, 2001, Beijing: Science Press). No more details in this article.

According to an embodiment of the present invention, the image can be processed in the following order: radiation correction, geometric correction, spectral reconstruction, band selection, and multi-source data compounding.

Band selection is to select a small number of bands that contain enough relevant information of ground objects among many bands, and find a small number of bands that can best reflect the information of ground objects.

The band selection problem can be attributed to assuming that the original data has n bands, and it is necessary to select an m-dimensional data subset (m<n) in this n-dimensional data set without losing important information. People hope to select as few bands as possible to reflect as much information as possible in the original band. When selecting the best band for conventional remote sensing data, the common method is to evaluate all possible combinations according to certain algorithms and criteria, so as to obtain the best band combination. For conventional remote sensing data, the number of bands is relatively small. For example, TM data has 7 bands. This method is feasible. But for hyperspectral data, due to the sharp increase in the number of bands, it is obviously inappropriate to copy this method. All combinations of selecting m-dimensional subsets in an n-dimensional data set may be $C_{\mathrm{no}}^m=\mathrm{no}!/(\mathrm{no}-m)!m!,$ This is a huge number. Taking the imaging spectrum data of 30 bands as an example, if 10 bands are selected among the 30 bands, there are 30,045,015 possible combinations. The actual imaging spectral data ranges from dozens of bands to hundreds or even hundreds of bands. Therefore, for hyperspectral data with many bands, it is unrealistic and unnecessary to consider all possible situations of selecting m bands from n bands one by one. Due to spectral subdivision, different from conventional remote sensing data, the number of bands owned by hyperspectral data in different spectral bands is no longer one or two, but more than a dozen or even dozens. They constitute several spectral groups, divided into It belongs to the spectral range with different properties in the electromagnetic wave, and the bands within the group have greater similarity, while different groups have greater differences. This is shown very clearly on the correlation coefficient matrix. Hyperspectral data can be clearly divided into several groups, and the correlation of each band within a group is strong, while the correlation between different groups is weak. The focus of spectral band selection should be on the evaluation of each band within a group, not between groups. Therefore, when selecting hyperspectral data bands, the hyperspectral data should be divided into several groups according to their properties, and then the quality of the spectral bands should be evaluated in units of groups, which not only avoids a lot of unnecessary calculations, but also can Improve the accuracy and efficiency of band selection.

Based on the above considerations, the method for calculating the band index for band evaluation of imaging spectral data is given below. Let _ρij be the correlation coefficient between band i and j, the imaging spectral data is divided into k groups, and the number of bands in each group is n ₁ , n ₂ ... n _k , and the band index is defined as:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}$

in

R _i =R _w +R _a

${\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&rho;</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>$

In the formula, σ _i is the mean square error of the i-th band, R _w is the average value of the sum of the absolute values of the correlation coefficients between the i-th band and other bands in the group, and Ra is the correlation between the i-th band and other bands outside the group. The sum of the absolute values of the correlation coefficients. That is to say, the band index is the ratio of the mean square error of the band to the average correlation coefficient of the band within the group and the sum of the absolute values of the correlation coefficients of the band and the bands outside the group.

Since the correlation of each band within a group is strong but the correlation between bands between groups is weak, the overall correlation of a band is mainly determined by its correlation with each band within a group, and the size of each group is different, that is, the composition The number of bands in a group is different, so using the average of the sum of the absolute values of the correlation coefficients between a band and other bands in a group as an item in the denominator of the band index can more reasonably reflect the overall quality of the band.

The meaning of the band index is very clear, it indicates the degree of dispersion between the band and other bands. The larger the band index, the less duplicate information it contains with the information contained in other bands. In other words, the larger the mean square error, the greater the discreteness of the band and the richer the information contained, while the smaller the absolute value of the overall correlation coefficient of the band, the stronger the independence of the band data and the greater the information redundancy. Small. Therefore, the band index Pi can comprehensively reflect the two factors of band information content and correlation, and can be used as one of the important parameters for band selection. Obviously, the band with larger Pi should be selected.

After band selection, various application processing can be performed on the data, such as multi-source data compounding, feature extraction, feature analysis, etc.

Multi-source data composition refers to superimposing or merging graphic or image data layers from different sources into one layer to obtain richer and more complete information. It is the process of combining multiple layers into one layer, that is, combining multiple data into one data. The purpose is to increase the amount of information.

Feature extraction is to identify and extract some information with specific features (including geometric features, spectral features, etc.) in the image, such as linear rivers, roads, water bodies with low reflectivity, etc. This is the process of selectively extracting some desired information from the numerous information in the image. The well-known genetic algorithm can be used to perform feature extraction and neuron network classification on remote sensing data such as imaging spectrometers and radars. In feature extraction, when necessary, multi-source data compounding is required.

Feature analysis refers to the analysis of the spatial distribution of various features and the relationship between each other, using statistical, mathematical or text methods to measure and describe them, and using the background knowledge (basic data) of the target to make data for the target. Analyze, understand, identify. Feature analysis techniques have been used in image recognition, and these techniques include detection of edges and corners reflecting geographical features in feature images, separation of foreground and background, texture analysis and spectrum, etc., which are commonly used at present.

In most cases, the processing sequence is multi-source data compounding, feature extraction, and feature analysis. However, there is no strict order among the three. For example, the data used for feature extraction may or may not need to be compounded; the data layer obtained by feature extraction may also be compounded with other data.

For agricultural monitoring, the parameters provided by the reconstructed spectra are not enough. That is to say, the agricultural monitoring model not only needs spectral parameters, but also needs some auxiliary information obtained from the ground to match the spectral information. This requires the use of the above-mentioned multi-source data composition. This process is realized through well-known steps such as layer overlay and attribute information table merging.

In the present invention, richer comprehensive information can be obtained through graphic compounding of remote sensing data and basic data.

Through the superposition of different layers between basic data and remote sensing data, different types of ground information can be obtained. These data from different sources form composite information, which provides auxiliary reference and support for the judgment and analysis of the spectrum of ground objects. Applying characteristic spectral curves to remote sensing images can distinguish different ground objects in a large area and further mine the deep surface object information contained in the image.

Since different data sources may contain repeated attribute information, attribute information tables can be merged so that these information can be consolidated to form unified and complete attribute information.

This combination of attribute information tables can use known GIS technology to edit and arrange graphic attribute tables from different sources.

The informative characteristics of the data can be analyzed in terms of relevant statistical parameters. Entropy and variance can be used to measure the evaluation basis of information size. The larger their value is, the richer the information is; If the contrast is large, the separability of the ground objects is strong; the correlation coefficient between the bands reflects the strength of the correlation of the bands. The strong correlation indicates that the data redundancy is large, and the weak correlation indicates that the independence of the bands is strong. The time phase of the remote sensing image is different and the condition of the underlying surface is different, the entropy, variance, mean value and dynamic range of the image will change, but the correlation coefficient between the bands changes very little.

A small amount of information-rich bands obtained by band selection are data sources for feature extraction. Both classification and feature selection are carried out on the basis of band selection, and the information obtained by feature extraction can be used for classification.

In a preferred embodiment of the present invention, a step of displaying agricultural information is provided. The following describes the method for analyzing agricultural conditions using data (spectrum, remote sensing data, etc.) obtained by earth observation technology in the present invention in conjunction with specific embodiments.

As shown in FIG. 3 , in one embodiment of the present invention, the agricultural situation is wheat growth. The steps are:

Utilizing the characteristics of hyperspectrum, that is, the characteristics of hyperspectrum are composed of rich spectral bands with small wavelength intervals, that is, the spectrum is very finely divided; for example, 10nm intervals, that is, the wavelength range covered by each band is only 10nm. These subdivided spectra can provide rich options for band selection and combination. By subdividing the spectrum, based on the ground observation data and conducting principal component analysis with the spectral data, the three main factors (chlorophyll, protein, water) reflecting the growth of wheat are obtained. Band information, and normalize them respectively. According to the contribution of these three main factors (chlorophyll C, protein P, water W) to the growth of wheat, the method of linear combination is used to calculate the growth function G of wheat,

G＝aC+bP+cW

Among them, a=2, b=1, c=2; they are the contribution rates of chlorophyll, protein and water factors to the growth of wheat respectively.

Chlorophyll information: the second band (0.46-0.48um) can well reflect chlorophyll information.

Protein information: The 15th band (1.00-1.02um) can well reflect protein information.

Moisture information: the 12th band (0.94-0.96um) can well reflect the moisture information.

The 8th band (0.64-0.66um) is a low reflection area for chlorophyll, protein, and water, and the 2nd and 8th bands, the 15th and 8th bands, and the 12th and 8th bands are used for normalization. Synthetic treatment eliminates the influence of bands, and at the same time highlights chlorophyll, protein, and water information to reflect the growth of wheat.

Then graded and evaluated the wheat growth function G to obtain the wheat growth results.

·Using basic data, superimposing thematic information on crop distribution, and removing non-wheat areas.

• Enter the database.

As shown in Fig. 4, in one embodiment of the present invention, the leaf area coefficient can be calculated.

First, radiometric correction and geometric correction are performed on the hyperspectral data, and then classification is performed, and the greenness value NI of wheat is obtained through normalization processing.

$\mathrm{<span\; class="notranslate">\; NI</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; CHANNEL</span>}\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; CHANNEL</span>}\mathrm{<span\; class="notranslate">\; 8</span>}}{\mathrm{<span\; class="notranslate">\; CHANNEL</span>}\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; CHANNEL</span>}\mathrm{<span\; class="notranslate">\; 8</span>}}$

Further obtain the wheat coverage fv, where CHANNEL8 and CHANNEL12 represent the values in bands 8 and 12, respectively. Use LAI＝K ^-1 ·In(1-fv) ^-1

where K is the extinction coefficient of wheat,

Finally, the distribution of wheat leaf area coefficient was obtained.

Figure 5 shows a schematic block diagram of a system for implementing the above method according to an embodiment of the present invention, the system comprising:

1. Data storage unit, including:

1) Spectral database, which stores the spectral curves of different crops or ground objects in the data format of vector graphics;

2) The remote sensing database stores the multi-band remote sensing images acquired by the airborne imaging spectrometer in the data format of raster images.

3) The basic database is used to store other auxiliary geospatial data (graphics and image data in raster or vector form) and attribute data (statistical data in text or table form) that match the remote sensing images of the research area. Such as vectorized regional administrative boundaries, meteorological elements, land cover type maps, etc.

Specifically, for vector data, the x and y coordinates can be used as the identification of the entity position, and the topology structure can also be used to reflect the relationship between entities. mainly include:

Point features, such as wells, observation stations, etc.

·Line elements, such as public (railway) roads, rivers, etc.

· Polygon elements, such as divisions, fields, etc.

·Note elements.

For vector data files

Adopt the map-oriented method, proceed one by one according to themes, and form corresponding thematic data files, including:

Coordinate file, storing spatial geometry data

attribute file, storage space attribute data

Various thematic data files in the same application field can form a file group.

In the system of the present invention, the vector data format can adopt ARC/INFO (E00) data format or/and Auto CAD (DXF) data format

For raster data, the map or image is divided into grids composed of several rows and columns, and each grid is used as a point to scan and sample the whole image to obtain the attribute data of each point. The whole picture after the grid is compiled is a regular array, so the coordinate position of the entity is hidden in the storage address of the grid.

This system needs to store a variety of thematic spatial data, and it is necessary to select areas of the same size and the same scale. At this time, each grid contains two or more attributes, which can be recorded in the following two ways:

File group mode. Each topic is independently documented, and all related documents form a document group.

multiband approach. Put each different theme in a different band to form a file.

The raster data format can adopt row-column matrix data format, PCI data format, TIFF data format, BMP data format.

For multi-band raster data, the following methods can be used to store them.

In BIP format, the file first arranges the data of each band of the first point in order, then arranges the data of each band of the second point, and so on to the last point.

In BIL format, the file first arranges the first band data in the first line, and then the second band data in the first line, until all the band data in the first line are arranged, and then arranges the band data in the second line in this order, until the last line.

In BSQ format, the file first arranges the data of all points in the first band, and then arranges the data of all points in the second band, and so on until the last band.

In one embodiment of the present invention, the above-mentioned database can be designed by using ARC/INFO geographic information system and FOXPRO database processing software.

2. The control and operation processing unit, including the image preprocessing unit, includes conventional image processing tools, such as known GIS functions. It is used to perform radiometric correction, geometric correction, noise removal, mosaic stitching, projective transformation, image enhancement, spectral enhancement, spectral identification, and image classification on remote sensing data images; it can also only perform some of the above-mentioned processing. All these processing methods can be implemented by using existing technologies, so details are not repeated here.

In one embodiment of the present invention, can adopt the PCI system Version6.2 each function module that Canadian PCI company develops as the core, and combine above-mentioned ARC/INFO geographical information system, FOXPRO database processing software, and the C/C++ of Microsoft Corporation Software, etc. form a systematic tool library, which is used to provide system software tools for data collection and entry, data management and maintenance, data routine processing, using models in the model units described later to process and calculate data in the database, and information map output, etc. support.

3. The model unit is used in the system of the present invention to store and modify the model algorithm, perform purposeful and specialized calculations on various data, and obtain the central processing module of the required information. The model in the model unit is not collected by the general geographic information system and remote sensing image processing system, and it is an algorithm module specially developed for the completion of various goals and tasks of this subject.

The model unit in the system of the present invention includes a parameter selection model library and an agricultural situation inversion model library.

The parameter selection model is used for spectral reconstruction, data composite feature analysis, and band selection for the processed image. These processing procedures have been described in detail above and are omitted here.

In the model unit, a farm situation inversion model library should be established. To this end, it is first necessary to collect models that have been researched and adopted by other sub-topics and have been proven to have good effects in practice. On the basis of analyzing the algorithm structure, select the necessary functional modules from the tool library, and adopt the algorithm according to the model. Linking in sequence, supplemented by program development when necessary, forms a complete set of model electronic processes from data input, model calculation to information output. When using the model, the user only needs to input the information of the data file to be processed and the name of the result file, without knowing the intermediate model operation process.

Referring to Figure 6, the following is a simple NDVI vegetation index calculation model as an example to illustrate the development process of each model in the agricultural inversion model library.

As mentioned above, when building a model, it is first necessary to calculate the formula according to the model provided by other sub-topics, in this case it is $\mathrm{NDVI}=\frac{\mathrm{Ch}2-\mathrm{Ch}1}{\mathrm{Ch}2+\mathrm{Ch}1},$ To analyze the algorithm structure, that is, to analyze the operation steps contained in the formula (band subtraction, band addition, and division operations, etc.).

Then, according to the obtained specific algorithm structure, the electronic construction of the model is carried out. When the user inputs an image to be processed, the system automatically invokes the band calculation module in the control and operation processing unit to subtract the two bands of the image and output the result to the first band of the new image.

The system calls the band calculation module in the tool library to add the two bands of the image and output the result to the second band of the new image.

The system calls the band operation module in the tool library to perform a division operation on the two bands of the new image to obtain the final result map.

Then, the system automatically deletes the images in the middle process and outputs the resulting image to the user.

The agricultural situation information models in the agricultural situation inversion model library of this system mainly include:

Hyperspectral as a physical and chemical property extraction model

Crop Growth Dynamic Monitoring Model

Composite model of hyperspectral flight data and existing spaceborne data

Parameter Selection Model for Imaging Spectrometer

The specific development methods of these models are basically similar and will not be repeated in this paper.

4. An information output unit, used to display or output relevant agricultural information according to the extracted agricultural conditions parameters.

The system of the present invention can provide a variety of information for agricultural applications, including land use maps, crop distribution maps, crop growth data, drought monitoring data, multiple information composite results, and agricultural parameter selection schemes. The method of the system of the present invention will be described below by taking land use information as an example.

Land use information is used to study the spectral characteristics of ground objects on hyperspectral images, and to explore the methods and potentials of using hyperspectral data for land use classification.

Land use maps are the basis of crop distribution maps, wheat growth maps and other maps, and the quality of land use maps directly affects the quality of other thematic maps. Therefore, when performing remote sensing mapping of land use maps, the system of the present invention can adopt various known methods such as visual interpretation, computer automatic identification and classification, and comprehensive classification, thereby ensuring the accuracy of this basic map. Figure 7 shows the specific process of taking the current land use map of Shunyi County, Beijing as an example.

Hyperspectral data: Imaging time was April 24, 1998, the season of jointing and heading of wheat. The spectral range is 0.44～2.40 μm, a total of 32 bands. The band ranges of each band are shown in Table 5. Due to instrument problems, seven bands (1, 9, 17, 22, 23, 25, and 32 bands) are of poor quality, so There are actually 25 bands that can be used.

Non-remote sensing data: 1:50,000 land color maps, which are thematic maps such as land use maps completed by predecessors.

Geometric correction and image mosaic: The Shunyi test area is composed of seven east-west flight strips, and each flight strip has a certain width of lateral overlap. On the same airway, the geometric distortion gradually increases from the middle to the west. In order to make full use of the better-quality images in the middle of the airway, the images with larger geometric distortion in the lateral overlapping range are cut off first, and then compared with the hyperspectral image and 1:50,000 topographic map, uniformly select control points on the image, establish a polynomial conversion equation, and perform geometric correction on the image after re-sampling. Finally, after geometric correction, the seven navigation belts with common map projection (Gauss-Krüger projection) and in the same coordinate system are mosaiced into an image.

Establish interpretation signs: conduct field investigations, compare with hyperspectral images, analyze and study the image characteristics of various objects in the experimental area, and establish image interpretation signs. On this basis, a cartographic specification and classification system are developed.

Classification of land use status: including visual interpretation, computer automatic classification and other methods.

(1) Visual interpretation method

Select the 12th band, 10th band and 6th band of hyperspectral flight data for RGB color synthesis;

·Using CorelDraw software as the operating platform, according to the established interpretation marks, referring to topographic maps, land use maps made by predecessors and other relevant materials, combined with comprehensive analysis of geoscience professional knowledge, human-computer interaction interpretation is carried out.

·Export the interpretation results from CorelDraw and save them in DXF format acceptable to Arc/Info.

·In Arc/Info, modify, edit and assign attribute codes to the interpretation results, and perform coordinate conversion.

·According to the difficult problems encountered in the interpretation, go to the field to investigate and solve them, and verify and modify the interpretation results.

• Enter the database.

(2) Computer automatic classification

A supervised classification

·Spectral band selection

For classification for a certain purpose, the correct selection of the best band and its combination is very important. Compared with other remote sensing data, hyperspectral data is characterized by a narrow band range, extremely rich spectral information, more optional bands, and a stronger ability to identify objects. But this does not mean that the more bands used in classification, the better. This is because, first, the number of spectral bands is not simply equal to the number of information bands. There are some bands with strong correlations between the data (Table 4.1). If all bands are used for classification, due to the mutual interference of related data, not only will the classification accuracy not be improved, but the classification results will be affected instead; second, too many bands selected will affect the calculation speed and put forward higher requirements for computer hardware ; Third, existing image processing software cannot meet this requirement. Various software have certain restrictions on the number of bands. For example, when PCI software performs supervised classification, the number of image bands cannot be greater than 16.

When selecting spectral bands, we mainly considered two factors, one is the correlation of each band, and the other is the spectral response characteristics of the objects in the experimental area.

Table 1 shows the correlation coefficients between each band of hyperspectral data. From Table 1, we can see the following characteristics:

a. In the range of visible light from bands 2 to 11, the correlation between each band is very strong, and the correlation between band 2 and other bands is the weakest, followed by band 11.

b. The infrared bands of bands 12-31 are not very strongly correlated with visible light bands, especially bands 12-16 have little correlation with visible light (correlation coefficient is less than 0.2), while the correlation coefficient between bands 18-31 is less than 0.2.

c. The correlation coefficients between bands 12-16 are all above 0.94, while the correlation coefficients between them and bands 18-31 are less than 0.2.

d. The correlation coefficient between bands 18-31 is 0.7-0.85, the strongest correlation is between bands 28 and 31, the weakest is between bands 19 and 29, and the correlation between band 19 and other bands is weak.

Judging from the light change response of ground objects, the spectral characteristics of wheat fields and orchards, orchards and vegetable fields and woodlands, reservoirs and rivers and ponds in the experimental area are similar in some bands, but if all bands To observe their spectral curves, you will find that they have differences in other bands (see Table 2).

Table 1 Correlation coefficient matrix of each band of hyperspectral data band 2 3 4 5 6 7 8 10 11 12 13 14 2 1.00 0.84 0.86 0.85 0.85 0.84 0.85 0.82 0.75 -0.04 -0.06 -0.11 3 0.84 1.00 0.92 0.95 0.95 0.95 0.95 0.93 0.86 -0.08 -0.08 -0.11 4 0.86 0.92 1.00 0.92 0.94 0.93 0.94 0.91 0.86 -0.02 -0.03 -0.06 5 0.85 0.95 0.92 1.00 0.95 0.96 0.96 0.94 0.86 -0.11 -0.12 -0.14 6 0.85 0.95 0.94 0.95 1.00 0.96 0.98 0.96 0.86 -0.16 -0.16 -0.19 7 0.84 0.95 0.93 0.96 0.96 1.00 0.99 0.97 0.87 -0.17 -0.17 -0.20 8 0.85 0.95 0.94 0.96 0.98 0.99 1.00 0.98 0.87 -0.18 -0.18 -0.20 10 0.82 0.93 0.91 0.94 0.96 0.97 0.98 1.00 0.86 -0.16 -0.15 -0.17 11 0.75 0.86 0.86 0.86 0.86 0.87 0.87 0.86 1.00 0.15 0.18 0.17 12 -0.04 -0.08 -0.02 -0.11 -0.16 -0.17 -0.18 -0.16 0.15 1.00 0.94 0.95 13 -0.06 -0.08 -0.03 -0.12 -0.16 -0.17 -0.18 -0.15 0.18 0.94 1.00 0.97 14 -0.11 -0.11 -0.06 -0.14 -0.19 -0.20 -0.20 -0.17 0.17 0.95 0.97 1.00 15 -0.12 -0.12 -0.07 -0.16 -0.20 -0.21 -0.22 -0.19 0.15 0.95 0.97 0.97 16 -0.09 -0.11 -0.05 -0.14 -0.19 -0.20 -0.21 -0.18 0.16 0.97 0.98 0.98 18 0.51 0.61 0.59 0.61 0.63 0.65 0.66 0.68 0.63 -0.02 -0.01 0.00 19 0.49 0.58 0.56 0.58 0.59 0.60 0.61 0.60 0.53 -0.07 -0.06 -0.07 20 0.55 0.61 0.60 0.60 0.61 0.63 0.63 0.64 0.65 0.15 0.16 0.15 twenty one 0.53 0.60 0.59 0.59 0.60 0.62 0.63 0.64 0.67 0.19 0.21 0.21 twenty four 0.67 0.76 0.74 0.76 0.79 0.81 0.82 0.81 0.71 -0.16 -0.15 -0.16 26 0.65 0.72 0.70 0.71 0.73 0.75 0.76 0.76 0.68 -0.07 -0.08 -0.09 27 0.66 0.72 0.71 0.72 0.74 0.76 0.77 0.76 0.69 -0.07 -0.08 -0.09 28 0.65 0.70 0.69 0.70 0.71 0.73 0.73 0.73 0.65 -0.07 -0.08 -0.10 29 0.65 0.69 0.69 0.68 0.70 0.71 0.72 0.71 0.65 -0.03 -0.04 -0.07 30 0.58 0.64 0.64 0.64 0.65 0.67 0.67 0.67 0.63 0.03 0.03 0.01 31 0.65 0.70 0.69 0.69 0.71 0.72 0.73 0.72 0.65 -0.06 -0.07 -0.09

Continued Table 1 15 16 18 19 20 twenty one twenty four 26 27 28 29 30 31 2 -0.12 -0.09 0.51 0.49 0.55 0.53 0.67 0.65 0.66 0.65 0.65 0.58 0.65 3 -0.12 -0.11 0.61 0.58 0.61 0.60 0.76 0.72 0.72 0.70 0.69 0.64 0.70 4 -0.07 -0.05 0.59 0.56 0.60 0.59 0.74 0.70 0.71 0.69 0.69 0.64 0.69 5 -0.16 -0.14 0.61 0.58 0.60 0.59 0.76 0.71 0.72 0.70 0.68 0.64 0.69 6 -0.20 -0.19 0.63 0.59 0.61 0.60 0.79 0.73 0.74 0.71 0.70 0.65 0.71 7 -0.21 -0.20 0.65 0.60 0.63 0.62 0.81 0.75 0.76 0.73 0.71 0.67 0.72 8 -0.22 -0.21 0.66 0.61 0.63 0.63 0.82 0.76 0.77 0.73 0.72 0.67 0.73 10 -0.19 -0.18 0.68 0.60 0.64 0.64 0.81 0.76 0.76 0.73 0.71 0.67 0.72 11 0.15 0.16 0.63 0.53 0.65 0.67 0.71 0.68 0.69 0.65 0.65 0.63 0.65 12 0.95 0.97 -0.02 -0.07 0.15 0.19 -0.16 -0.07 -0.07 -0.07 -0.03 0.03 -0.06 13 0.97 0.98 -0.01 -0.06 0.16 0.21 -0.15 -0.08 -0.08 -0.08 -0.04 0.03 -0.07 14 0.97 0.98 0.00 -0.07 0.15 0.21 -0.16 -0.09 -0.09 -0.10 -0.07 0.01 -0.09 15 1.00 0.98 -0.02 -0.08 0.14 0.19 -0.18 -0.11 -0.11 -0.11 -0.08 0.00 -0.10 16 0.98 1.00 -0.03 -0.08 0.14 0.19 -0.17 -0.10 -0.10 -0.10 -0.06 0.01 -0.09 18 -0.02 -0.03 1.00 0.74 0.79 0.82 0.80 0.77 0.79 0.76 0.75 0.74 0.76 19 -0.08 -0.08 0.74 1.00 0.78 0.75 0.68 0.67 0.68 0.69 0.65 0.69 0.69 20 0.14 0.14 0.79 0.78 1.00 0.91 0.70 0.72 0.74 0.73 0.72 0.76 0.74 twenty one 0.19 0.19 0.82 0.75 0.91 1.00 0.71 0.72 0.74 0.73 0.71 0.76 0.73 twenty four -0.18 -0.17 0.80 0.68 0.70 0.71 1.00 0.86 0.87 0.84 0.84 0.76 0.84 26 -0.11 -0.10 0.77 0.67 0.72 0.72 0.86 1.00 0.88 0.84 0.85 0.76 0.84 27 -0.11 -0.10 0.79 0.68 0.74 0.74 0.87 0.88 1.00 0.86 0.88 0.78 0.88 28 -0.11 -0.10 0.76 0.69 0.73 0.73 0.84 0.84 0.86 1.00 0.86 0.79 0.97 29 -0.08 -0.06 0.75 0.65 0.72 0.71 0.84 0.85 0.88 0.86 1.00 0.80 0.87 30 0.00 0.01 0.74 0.69 0.76 0.76 0.76 0.76 0.78 0.79 0.80 1.00 0.78 31 -0.10 -0.09 0.76 0.69 0.74 0.73 0.84 0.84 0.88 0.97 0.87 0.78 1.00

Table 2 Spectral separability of some ground objects in the experimental area

Note: √ indicates that the ground objects can be separated in this band

Considering the band characteristics of ground objects and the correlation of data in each band, select bands 2, 6, 11, 12, 19, 24 and 30 to participate in the classification.

· Select training area

According to the spectral characteristics of each type of surface object, a typical sample area that can represent its characteristics is selected as the training area. Considering the existence of the same objects and different spectrums, more than one training area is selected for each type of ground object.

·Classification

Classification with Maximum Likelihood

B. Based on standard spectral database taxonomy

·Establish the standard spectral database of the main ground features in the test area

Convert the standard spectral database to the database file format of the PCI software

Convert the hyperspectral image from gray value to reflectance value according to the calibration coefficient

·Comparing the reflectance curve of the pixel on the hyperspectral image with the standard spectral curve, judging the belonging class of each pixel point by point, and obtaining the classification map.

C. Comprehensive taxonomy

Each of the above classification methods has its own advantages, but at the same time there are disadvantages. The classification accuracy of the visual interpretation method is high, but the interpretation results are greatly affected by the professional quality of the interpreters, and it takes a lot of time. Although the supervised classification method is fast, it cannot solve the problem of the same objects because it is a pure spectral classification method. Due to the problem of different spectra and same spectrum and foreign objects, the classification accuracy is limited; the classification method based on the standard spectral database must be based on a sufficient spectral database, and at the same time has high requirements for image preprocessing and image quality. Based on the above considerations, we propose a comprehensive classification method. The so-called comprehensive classification method is to organically combine various classification methods to achieve the best overall effect in order to save time, effort and practicality under the premise of satisfying the accuracy.

The specific steps are:

·For key features, it can be classified under the support of standard spectral database

·Classify features that can be separated only by spectrum using supervised classification

·On the basis of the above two classifications, through human-computer interaction, interpret the features that are not easy to separate by other methods.

Doing so can not only ensure the mapping accuracy, but also greatly reduce the classification time.

The preferred embodiments of the present invention have been described in detail above for the purpose of illustration, but those of ordinary skill in the art should realize that various improvements, additions and substitutions are possible within the scope and spirit of the present invention, and All are within the scope of protection defined by the claims of the present invention.

Claims (10)

1. the agriculture application integration method of an earth observation technology comprises:

Obtain the earth observation data;

The earth observation data that obtained are carried out the data band selection;

Earth observation data after described band selection are carried out feature extraction, obtain agricultural feelings parameter;

Utilize the agricultural feelings parameter of being obtained to carry out signature analysis, obtain the information needed in the described earth observation data.

2. method according to claim 1, it is characterized in that, described earth observation data are high-spectral data, described method further comprises carries out pretreated step to high-spectral data, and described pre-treatment step comprises one or more following processing: radiant correction, geometry correction, rebuilding spectrum, image transformation.

3. method according to claim 2, it is characterized in that, the step of described feature extraction comprises the compound step of obtaining the variety classes terrestrial information by multi-source data, and described multi-source data is compound to carry out figure stacked Calais realization by described high-spectral data and terrestrial information.

4. method according to claim 2 is characterized in that, described image transformation is grouping KL conversion.

5. method according to claim 2 is characterized in that, described band selection comprises:

Wave band to described high-spectral data carries out band grouping; With

Utilize institute to divide each batch total calculation band index;

Wherein, establish ρ _IjBe the related coefficient between wave band i and the j, imaging spectrometer data is divided into the k group, and every group wave band number is respectively n ₁, n ₂N _k, the definition band index is:

$P_i=\frac{{\mathrm{\&sigma;}}_i}{R_i}$

R _i＝R _w+R _a

$R_w=\frac{1}{n_k}\underset{j=1}{\overset{n_k}{\mathrm{\&Sigma;}}}{\mathrm{\&rho;}}_{\mathrm{ij}},$ I ≠ j wherein

σ in the formula _iBe the mean square deviation of i wave band, R _wBe the mean value of the absolute value sum of other wave band related coefficient in i wave band and the place group, Ra is the absolute value sum of the related coefficient between i wave band and place group other wave band in addition.

6. method according to claim 1 is characterized in that, described agricultural feelings parameter comprises chlorophyll information, protein information and wheat moisture content information, and described relevant Agricultural Information comprises the wheat growth information; Perhaps

Described agricultural feelings parameter information is a vegetation index, and described relevant Agricultural Information is a crop yield trend.

7. the agriculture application integrating system of an earth observation technology comprises:

Data storage cell comprises:

Spectra database, it is with the data mode storage different crops of vector graphics or the curve of spectrum of ground object target;

The Multi-Band Remote Sensing Images that airborne imaging spectrometer obtains is stored with the data mode of grating image in the remotely-sensed data storehouse;

Model unit, comprise selection model storehouse and agricultural feelings inverse model storehouse, described selection model storehouse is used for image is carried out rebuilding spectrum, the analysis of data compound characteristics and band selection, and described agricultural feelings inverse model storehouse is used to provide polytype agricultural feelings information model;

Control and operation processing unit are used to utilize the model of described model unit that view data is handled accordingly; With

Information output unit is used for according to the agricultural feelings parameter of being extracted, the Agricultural Information that demonstration or output otherwise are correlated with.

8. the agriculture application integrating system of earth observation technology according to claim 7 is characterized in that, further comprises basic database, is used for storing complementary geographical spatial data and the attribute data that is complementary with the survey region remote sensing images.

9. the agriculture application integrating system of earth observation technology according to claim 7, it is characterized in that, figure, view data that described complementary geographical spatial data is grid or vector form, described attribute data is the statistics of text or form, the regional Administrative boundaries that comprise vector quantization, meteorological element, soil cover type map.

10. the agriculture application integrating system of earth observation technology according to claim 7, it is characterized in that described agricultural feelings information model comprises: high spectrum crop physicochemical property extraction model, crop growing state dynamic monitoring model, high spectrum flying quality and existing spaceborne data composite model and imaging spectrometer selection model.

Images