CN1760889A - Method for sorting characters of ground object through interfusion of satellite carried microwave and infrared remote sensing - Google Patents

Abstract

Interfusion between radar image and infrared imagery in pixel level is carried out for complicated city zone by using principal component analysis method. Through combination of back propagation artificial neural network (BP-ANN) and genetic algorithm (GA), BP-ANN/GA hybrid algorithm is built up. By using ability of global optimization of GA algorithm, the method optimizes initial weight value of BP-ANN, overcomes disadvantages of slow convergence rate of traditional BP-ANN, and easy of getting into local minimum so as to increase speed for automatic sorting remote sensing images from multiple sources and accomplish sort for characters of ground objects.

Description

Satellite-borne microwave and infrared remote sensing merge the method that the city characters of ground object is classified

Invention field

The invention belongs to space remote sensing and computer image processing technology field, be specifically related to the method that a kind of fused images of utilizing satellite-borne microwave synthetic-aperture radar and infrared remote sensing is classified to complicated city multiclass characters of ground object.

Background technology

Modern Remote Sensing Technical provides the remote sensing image data of multiresolution, multiband, multidate for earth observation, as LandsatTM/ETM+, SPOT 5, Quickbird, ERS-1/2SAR and Radarsat etc., but the view data that single remote sensing means are obtained exists tangible limitation and otherness at aspects such as how much, spectrum and spatial resolutions.Multi-source Information Fusion produces than more accurate, more reliable estimation of single source information and judgement by multi-source data is synthetic.Utilizing the multi-source remote sensing fused images that face of land atural object is classified is one of focus of present Remote Sensing Study, but the diversity of city atural object and complicacy.The existence of especially a large amount of man-made features makes traditional sorting technique can not obtain desirable classifying quality.And back-propagation artificial neural network (Back Propagation Artificial Neural Network, BP-ANN) be the mathematical model of Artificial Neural System's first approximation, be widely used in solving all kinds of nonlinear problems, but be absorbed in local minimum easily in the training process.(Genetic Algorithm GA) is biological heredity and evolutionary process and a kind of adaptive global optimization probabilistic search algorithm that forms in physical environment of simulation to genetic algorithm.

Summary of the invention

The objective of the invention is to propose the method that a kind of fused images of utilizing satellite-borne microwave synthetic-aperture radar and infrared remote sensing is classified to complicated city multiclass characters of ground object.

At present the classification of urban surface atural object remotely-sensed data commonly used has high-definition remote sensing data, single multispectral data or radar data.Though high spectrum and multispectral data spectral information are abundant, there is the phenomenon of " jljl foreign matter " and " foreign matter is with spectrum " simultaneously; Though and radar data side-looking imaging is relatively more responsive to the geometry of atural object, the same difficult atural object of distinguishing the back scattering feature similarity.When the branch time-like that both is applied to complicated city atural object separately, nicety of grading is subjected to the restriction of characteristics of image.The method of using always during remote sensing image classification is handled is as minimum distance method and maximum likelihood method ^[1], all be based on the statistical nature of remotely-sensed data, depend on choosing of prior probability.Artificial neural network (ANN) ^[2-4](mainly being the BP-ANN of comparative maturity) not necessarily needs to suppose certain parametrization Density Distribution because data distribute to feature space in advance, overcome above-mentioned defective to a certain extent.But BP-ANN network using gradient descent method training weights, speed of convergence is slow, and is absorbed in local minimum easily.Genetic algorithm (GA) is the adaptive optimized Algorithm of a kind of global search, can remedy the deficiency of gradient descent method.

The present invention sets up the BP-ANN/GA hybrid algorithm that BP-ANN and GA combine, and realizes the tagsort of the face of land, complicated city atural object in radar and the infrared fused images.Concrete steps are as follows:

(1) pre-service of SAR image and infrared remote sensing view data.Comprise: the auto adapted filtering of radar image is handled; The normalized of the volume coordinate registration between radar image and the infrared image and gray scale resampling and data;

(2) artificial neural network training.At first from the infrared remote sensing image, select sample input BP-ANN network, it is trained, on the basis of analyzing and training sum of errors speed of convergence, the connection weight of utilization GA algorithm optimization network, to optimize the result and import the BP-ANN network, thereby make up the BP-ANN/GA algorithm.With same infrared remote sensing image input BP-ANN/GA algorithm, it is trained then.At last according to radar image to fluctuating atural object scattering signatures remarkable advantages, the information that fully utilizes infrared and radar image is to the city terrain classification.Specifically, it is the training of further importing the BP-ANN/GA of infrared image and radar fused images, promptly infrared image and the radar image utilization principal component analytical method to the same area carries out fusion treatment, account in the variance of analyzing each principal component fusion image on the basis of number percent of population variance, from the high major component of shared number percent, choose sample, as the input of BP-ANN/GA network, train, obtain more satisfactory training result;

(3) classification of remote sensing images.With the end product of network training fixedly BP-ANN, BP-ANN/GA network of network structure and network parameter, infrared image is imported BP-ANN and BP-ANN/GA network, the processing of classifying; And with fused images input BP-ANN/GA network infrared and radar, the processing of classifying.At last the precision and the speed of network class are carried out quantitative test respectively.

Content of the present invention further describes as follows:

1, data pre-service

1.1 the auto adapted filtering of radar image is handled (being noise reduction process)

Because radar wave and the interaction between the surface randomly, the spot in the radar image is intrinsic a kind of feature.The existence of the man-made features of different materials, differing heights and different azimuth such as urban area buildings, road makes that the scattering process between radar wave and the atural object is complicated more, also makes the blob features of radar image more outstanding.For reducing the influence that speckle noise is handled data in the radar image, adopt the Gamma method to carry out auto adapted filtering to it and handle ^[5], when reducing speckle noise, kept the detailed information of radar image.

1.2 volume coordinate registration between radar and the infrared image and gray scale resample

Because image-forming principle, spatial resolution and spectral resolution between SAR image and the infrared remote sensing image are different, before the fusion of carrying out both and classification processing, use polynomial method that it is carried out the volume coordinate registration respectively ^[6], utilization cube convolution method is carried out gray scale and is resampled ^[6]

1.3 the normalized of data

Condition for logarithm S type (Sigmoid) activation function that satisfies BP-ANN ^[7], data are done following normalized:

$x^{*}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}\·\·\·\left(1\right)$

X wherein ^*Be the data after handling, x is the sample data before handling, x _Min, x _MaxBe respectively the minimum value and the maximal value x of sample

2, artificial neural network training

2.1 the training of input infrared remote sensing image B P-ANN

BP-ANN is the mathematical model of Artificial Neural System's first approximation, is widely used in solving all kinds of nonlinear problems, and concrete network model and learning rules can be referring to documents ^[7]When being applied to the remote sensing images supervised classification, at first the input signal of the extraction feature of remote sensing images as network, network can be classified to output signal at output terminal after training by certain rule.The present invention adopts an input layer, hidden layer and output layer to form 3 layers of BP-ANN.Input layer number and output layer node number respectively with the wave band number of selected remote sensing images and sample treat that sub-category number is identical; According to the speed of convergence and the nicety of grading of network in the actual experiment, determine the number of hidden node.S type (Sigmoid) activation function with non-linear gain amplifier is all adopted in the output of hidden layer and output layer:

$f\left(x\right)=\frac{1}{1+e^{-x}}\·\·\·\left(2\right)$

To treat that the pairing target output of branch atural object is expressed as the vector T of N dimension (N is the neuron number of output layer), T=[t ₁, t ₂T _N], actual output is expressed as the vector R of same dimension, R=[r ₁, r ₂R _N].If do not obtain the output expected at output layer, the error sum of squares of then calculating the output of output layer target and actual output is as the error changing value:

$E=\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{(t_k-r_k)}^2\·\·\·\left(3\right)$

T wherein _kBe output layer k neuronic target output, r _kBe its actual output.

Turn to backpropagation then, error signal is returned along original connecting path anti-pass, revise the neuronic weights of each layer, up to the training error curve convergence by network.

The present invention at first chooses sample and as the input of BP-ANN it is trained from the infrared remote sensing image.But be subjected to the BP-ANN algorithm to be absorbed in local minimum limitation, the training speed of network and all not high easily to the precision of image classification.For training speed and the precision that improves network, overcome the shortcoming of random initializtion weight in the BP-ANN training, the present invention uses GA initialization network weight, dwindles its hunting zone, by the mechanism optimization weight of " survival of the fittest ", makes up the BP-ANN/GA algorithm.

2.2 the training of input infrared remote sensing image B P-ANN/GA

The GA algorithm is biological heredity and evolutionary process and a kind of adaptive global optimization probabilistic search algorithm that forms in physical environment of simulation ^[8-9]Adopt genetic algorithm optimization BP-ANN initial weight, be actually whole network training separated into two parts: the initial weight that at first adopts GA algorithm optimization network; To optimize the result then imports BP-ANN and finishes network training.

Connection weight according to BP-ANN generally is the characteristics of real number, and the present invention adopts real coding to chromosome, is different from general binary coding ^[8]The chromosomal number gene of each individuality of GA is identical with the connection weight number of BP-ANN, and population scale and termination evolutionary generation are set according to experiment experience, and initial population is got the random number between (0,1).Fitness function is defined as:

$f(t,p)=\frac{\max \_\mathrm{err}\left(t\right)-\mathrm{err}(t,p)}{\max \_\mathrm{err}\left(t\right)-\min \_\mathrm{err}\left(t\right)}\·\·\·\left(4\right)$

Wherein (t p) represents t generation p individual fitness to f; Max_err (t) and min_err (t) represent the minimum and maximum network output error of t for all individualities respectively; (t p) represents t generation p individual network output error to err.Adopt the probability that is directly proportional with fitness to carry out selection operation, floating number linear crossing, Gaussian mutation.

With the individuality with minimal network output error of the last iteration gained of the GA algorithm initial weight as BP-ANN, substitution BP-ANN trains, and makes up the BP-ANN/GA network.

Will with 2.1 same infrared remote sensing data input BP-ANN/GA networks, carry out network training, the speed of network training and the precision of terrain classification is improved simultaneously.

2.3 import infrared and training radar fused images BP-ANN/GA

Be the high and high advantage of radar image spatial resolution of spectral resolution of giving full play to the infrared remote sensing image, overcome the close deficiency of back scattering of water, smooth meadow and road in " the different spectrum of jljl; foreign matter is with spectrum " phenomenon and radar image of infrared image, two kinds of data are carried out principal component analysis (PCA) ^[10]Fusion treatment.

Principal component analysis (PCA) is a kind of orthogonal linear transformation of obtaining mutual uncorrelated proper vector, can reduce the dimension of data under the least possible situation of information loss.At first ask the eigenwert of multiband remote sensing data covariance matrix, and it is arranged by descending order, ask then and each eigenwert characteristic of correspondence vector, construct a transformation matrix.Utilize this transformation matrix that raw image data is carried out conversion, obtain each major component image.

The present invention at first forms multi-band image with infrared remote sensing image and radar image, and it is carried out principal component transform, obtains the principal component fusion image, and different principal component fusion images has comprised different terrestrial object information features.Account in the variance of analyzing each principal component fusion image on the basis of number percent of population variance, choose of the input of the high major component of shared number percent as the BP-ANN/GA network, network is trained, the training precision of network is further improved, obtain even more ideal result.

3, remote sensing image classification

At first, import the infrared remote sensing image respectively, the processing of classifying with the end product of network training in 2.1 and the 2.2 fixedly structure and the network parameter of BP-ANN and BP-ANN/GA network.The BP-ANN/GA sorting result is than the BP-ANN ideal, but because infrared image has the characteristics of " the different spectrum of jljl, foreign matter is with spectrum ", still exists tangible atural object to mix the branch phenomenon.With the end product of network training in the 2.3 fixedly structure and the network parameter of BP-ANN/GA network, import infrared and fused images radar then, it is classified, the classification results of gained relatively conforms to the distribution of actual atural object.

Effect for each method classification of quantitative comparison, result according to visual interpretation, other picture information and field study, from image, choose equally distributed sample data, set up confusion matrix, calculate the nicety of grading of different sorting techniques: overall nicety of grading and Kappa coefficient ^[6]In addition, also added up different sorting techniques and finished the required time of classification, to compare its classification speed.

Description of drawings

Landsat ETM+4 image behind Fig. 1 registration.

The original ERS-2SAR image of Fig. 2.

Fig. 3 imports Landsat ETM+4, the training error curve of 5,7 triband infrared image BP-ANN.

Fig. 4 imports LandsatETM+4, the training error curve of 5,7 triband infrared image BP-ANN/GA.

Fig. 5 imports Landsat ETM+4, the training error comparison diagram of 5,7 triband infrared image BP-ANN/GA and BP-ANN.

The principal component analysis (PCA) fused images of the infrared and radar image of Fig. 6, wherein Fig. 6 (a) is the 1st major component, and Fig. 6 (b) is the 2nd major component, and Fig. 6 (c) is the 3rd major component.

Fig. 7 imports infrared and graph of errors radar fused data BP-ANN/GA network.

Fig. 8 ETM+4,5, the 7BP-ANN classification results.

Fig. 9 ETM+4,5, the 7BP-ANN/GA classification results.

Figure 10 ETM+4,5,7 with SAR fused images BP-ANN/GA classification results.

Figure 11 study area communication chart.

Figure 12 radar and infrared image integrated classification process flow diagram.

Embodiment

The present invention chooses near the zone that land mass, river and buildings are arranged the Shanghai City Pudong Century Park, and infrared data is the 4th, 5,7 wave bands of the Landsat ETM+ on November 27th, 2002, and spatial resolution 30m, Fig. 1 have shown the 4th wave band wherein.Radar data is the ERS-2SAR on April 9th, 2002, and its frequency is 5.3GHz, the VV polarization, and spatial resolution is 12.5m, sees Fig. 2.

It is five classes that the present invention is divided into the atural object of study area: water, meadow, buildings, road and open ground.At first will treat the input of 4 samples of each class atural object selection of branch as BP-ANN from the 4th, 5,7 wave bands of LandsatETM+, it is trained that gained training error curve is seen Fig. 3, actual output of network and target output see Table 1.With sample input BP-ANN/GA network same on the infrared image, gained training error curve is seen Fig. 4 then, and actual output of network and target output see Table 2.Both training error contrasts, as shown in Figure 5.

The principal component analysis (PCA) result's of radar and infrared image the 1st, the 2nd and the 3rd major component is seen Fig. 6 (a), (b) and (c).Fig. 7 and table 3 has provided actual output and the target output of the sample of fused images being imported the training error curve and the network of BP-ANN/GA network gained respectively.

Fig. 8 and Fig. 9 are respectively that utilization BP-ANN and BP-ANN/GA carry out sorting result to infrared image, and Figure 10 utilizes BP-ANN/GA that radar and infrared fused images are carried out sorting result.The nicety of grading and the speed of distinct methods see Table 4.Gray scale among three figure has identical sign: black is represented water body, and Dark grey is represented the meadow, and the light grey buildings, canescence represented represented open ground and the white road of representing.Be the effect that shows classification directly perceived, Figure 11 has provided the communication chart of study area.Figure 12 has then shown the detail flowchart that utilization BP-ANN/GA classifies to radar and infrared fused images.

Table 1 input Landsat ETM+4, the actual output of 5,7 triband infrared image BP-ANN and target output.

Table 2 input Landsat ETM+4, the actual output of 5,7 triband infrared image BP-ANN/GA and target output.

Table 3 input actual output and target infrared and radar fused images BP-ANN/GA network are exported.

The nicety of grading of the different sorting techniques of table 4.

Table 1

Table 2

Table 3

Sorting technique	BP-ANN	BP-ANN/GA
				The input data	Landsat ETM+4,5,7 triband images	Landsat ETM+4,5,7 triband images	Landsat ETM+4, the fused images of 5,7 tribands and SAR
Overall nicety of grading	71.46％	72.75％	89.91％
				The Kappa coefficient	0.6339	0.6596	0.8698
The classification time (s)	4620.2	3830.3	3623.8

Table 4

Claims (3)

1, a kind of fused images of utilizing synthetic-aperture radar and infrared remote sensing method that complicated city characters of ground object is classified is characterized in that concrete steps are as follows:

(1) pre-service of SAR image and infrared remote sensing view data comprises: the auto adapted filtering of radar image is handled, the normalized of the volume coordinate registration between radar image and the infrared image and gray scale resampling and data;

(2) artificial neural network training is at first selected sample input BP-ANN network from the infrared remote sensing image, it is trained, on the basis of analyzing and training sum of errors speed of convergence, the connection weight of utilization GA algorithm optimization network, to optimize the result and import the BP-ANN network, thereby make up the BP-ANN/GA algorithm; With same infrared remote sensing image input BP-ANN/GA algorithm, it is trained then; The last training of further importing the BP-ANN/GA of infrared image and radar fused images;

(3) classification of remote sensing images is imported BP-ANN and BP-ANN/GA network with infrared image, the processing of classifying with the end product of network training fixedly BP-ANN, BP-ANN/GA network of network structure and network parameter; And with fused images input BP-ANN/GA network infrared and radar, the processing of classifying; At last the precision and the speed of network class are carried out quantitative test respectively.

2, the method for utilizing the fused images of synthetic-aperture radar and infrared remote sensing to the classification of complicated city characters of ground object according to claim 1, the auto adapted filtering that it is characterized in that described radar image is handled the Gamma method that adopts, volume coordinate registration between described SAR image and the infrared remote sensing image adopts polynomial method, and gray scale resamples and adopts the cube convolution method; Described data normalization is handled, and its formula is:

$x^{*}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}---\left(1\right)$

X wherein ^*Be the data after handling, x is the sample data before handling, x _Min, x _MaxBe respectively the minimum value and the maximal value of sample.

3, the method for utilizing the fused images of synthetic-aperture radar and infrared remote sensing to the classification of complicated city characters of ground object according to claim 1 is characterized in that:

The training of described input infrared remote sensing image B P-ANN, the steps include: to adopt an input layer, hidden layer and output layer to form 3 layers of BP-ANN, input layer number and output layer node number respectively with the wave band number of selected remote sensing images and sample treat that sub-category number is identical; According to the speed of convergence and the nicety of grading of network in the actual experiment, determine the number of hidden node; The S type activation function with non-linear gain amplifier is all adopted in the output of hidden layer and output layer:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(2\right)$

To treat that the pairing target output of branch atural object is expressed as the vector T of N dimension, N is the neuron number of output layer, T=[t ₁, t ₂T _N], actual output is expressed as the vector R of same dimension, R=[r ₁, r ₂R _N]; If do not obtain the output expected at output layer, the error sum of squares of then calculating the output of output layer target and actual output is as the error changing value:

$E=\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{(t_k-r_k)}^2---\left(3\right)$

T wherein _kBe output layer k neuronic target output, r _kBe its actual output;

Turn to backpropagation then, error signal is returned along original connecting path anti-pass, revise the neuronic weights of each layer, up to the training error curve convergence by network;

The training of described input infrared remote sensing image B P-ANN/GA the steps include: at first to adopt the initial weight of GA algorithm optimization network, will optimize the result then and import BP-ANN and finish network training.

The training of the infrared and radar fused images BP-ANN/GA of described input, the steps include: at first infrared remote sensing image and radar image to be formed multi-band image, it is carried out principal component transform, obtain the principal component fusion image, account in the variance of analyzing each principal component fusion image on the basis of number percent of population variance, choose of the input of the high major component of shared number percent, network is trained, the training precision of network is further improved as the BP-ANN/GA network.

Images