CN112287871B - Remote sensing image extraction method for nearshore aquaculture area based on multi-feature and spectral fusion - Google Patents

Abstract

The invention discloses a near-shore aquaculture area remote sensing image extraction method based on multi-feature and spectrum fusion. Secondly, the improved constrained energy minimization algorithm is used for enhancing the target ground objects in the culture area, background spectrum information is weakened, then the Otsu method is used for calculating a threshold value, and the single-waveband threshold value is combined for carrying out primary extraction on the target ground objects. And finally, according to the ground feature characteristics, the primarily extracted result is subjected to customized elimination of ground feature interfering objects by using a gray level co-occurrence texture matrix or an object-oriented method, and the final extracted result of the culture area is output. Compared with the traditional target detection method, the method can effectively overcome the interference of background ground objects even in the culture area with obvious foreign matter co-spectrum phenomenon, obtain the extraction result of the raft culture area with higher precision, and better meet the high-precision extraction requirement of the raft culture area.

Description

Remote sensing image extraction method for nearshore aquaculture area based on multi-feature and spectral fusion

technical field

The invention belongs to the technical fields of remote sensing technology and image processing, and in particular relates to a method for extracting remote sensing images of coastal aquaculture areas based on multi-feature and spectrum fusion.

Background technique

Since the second industrial revolution, more and more countries and regions have paid more and more attention to fishery resources as a supplement to terrestrial food resources. As an important part of fishery resources, aquaculture activities have higher self-selectivity and economic benefits, but have a deeper impact on the natural environment. With the continuous deposition of excrement in water and the increase of residual bait, the ammonia nitrogen content in aquaculture water continues to increase. Accumulation, the phenomenon of eutrophication in water body intensifies, and the natural environment of aquaculture area deteriorates. Therefore, how to carry out reasonable breeding and scientific planning of aquaculture areas is the main problem of sustainable development of fishery resources, and effectively mastering the spatial layout of aquaculture areas and improving the dynamic monitoring technology of aquaculture areas is the key to rational planning and scientific management of fishery resources. important link.

Compared with traditional monitoring technology, remote sensing technology can macroscopically, continuously and automatically observe surface objects by formulating satellite orbit path and operation cycle, which overcomes the shortcomings of long traditional monitoring cycle, time-consuming and labor-consuming, and large human interference factors. An important means of dynamic monitoring in breeding areas. The extraction methods of aquaculture areas based on remote sensing technology mainly include visual interpretation method, object-oriented method, methods based on pixel spectral characteristics and texture analysis. Some scholars have achieved relatively ideal extraction results in the corresponding aquaculture area extraction experiments. and application prospects, greatly promoting the development of remote sensing technology. Among them, the visual interpretation method is the most commonly used, but its accuracy largely depends on the interpretation experience of the visual interpretation personnel, which has low objectivity, heavy workload and time-consuming, which is not conducive to long-term and dynamic monitoring in breeding areas need. The object-oriented extraction method comprehensively considers the spatial, spectral, texture, and shape features of the classification objects in remote sensing images, and reduces the influence of "salt and pepper" noise that is difficult to solve in traditional image extraction methods. However, due to the subjectivity of the segmentation scale in the extraction process The problem of "different objects with the same spectrum" in some pixels may lead to a decrease in the extraction accuracy of the breeding area. The pixel-based extraction method can make better use of the spectral reflection characteristics of aquaculture areas, highlight the differences between aquaculture areas and non-aquaculture areas, and use thresholds to automatically extract aquaculture areas. However, due to the differences between sensors and their own parameters, there is no aquatic characteristic index that is available for most satellite image data, and the reflection characteristics of aquaculture areas are differentiated due to the influence of different water quality factors in the same aquaculture area Phenomenon. This also adds difficulties to the pixel-based extraction method, and it also makes it difficult for the pixel extraction method to independently and accurately extract the aquaculture area, and the combination with other methods can make better use of the spectral characteristics of the aquaculture area. Improve the accuracy of the aquaculture area extraction algorithm.

Aiming at the deficiencies of the existing algorithms, the present invention proposes a remote sensing image extraction method based on multi-feature and spectral fusion in nearshore aquaculture areas, and constructs a method that combines the spectral features of remote sensing images, threshold methods, and gray-level symbiotic texture matrices. Combined with the multi-feature analysis method, the precise extraction of breeding areas can be realized.

Contents of the invention

The purpose of the present invention is to provide a method for extracting remote sensing images of coastal aquaculture areas based on multi-feature and spectral fusion, so as to solve the above-mentioned problems.

The technical solution that realizes the object of the present invention is:

The method for extracting remote sensing images of coastal aquaculture areas based on multi-feature and spectral fusion is characterized in that it includes the following steps:

Step 1: Input an original remote sensing image;

Step 2: Carry out image preprocessing on the input remote sensing image, and use the feature index method to extract the characteristic spectrum for the object features in the processed image, and construct the target object feature set according to the obtained characteristic spectrum;

Step 3: Construct a finite impulse response linear filter, based on the constrained energy minimization algorithm of the gradient integral recurrent neural network, select the spectral data of the target surface object pixel, and use the finite impulse response linear filter to enhance the target surface object spectral data, and obtain Enhanced remote sensing images;

Step 4: For the enhanced remote sensing image, use the Otsu algorithm and single-band threshold to perform preliminary extraction of the target feature, and obtain the remote sensing image after the preliminary extraction of the target feature;

Step 5: Use the object-oriented method or the gray-level co-occurrence texture matrix to remove the disturbing objects of the ground features in the breeding area based on the texture features and geometric features of the ground features after the preliminary extraction;

Step 6: Output the final extracted remote sensing image of the breeding area.

Further, it is characterized in that the specific operation steps of step 3 include:

Step 31: Construct a finite impulse response linear filter according to the known prior information of the target pixel spectrum;

Step 32: Express the constrained energy minimization algorithm as a linear constrained optimization mathematical model, the expression is:

d表示约束条件向量，并且d满足条件：Among them, w represents the filtering coefficient, R is the autocorrelation matrix of the remote sensing image, and

d represents a vector of constraints, and d satisfies the condition:

Step 33: Under the constraint condition vector, when the output y _i of the filter for the input _ri satisfies the following formula:

Then the average output energy corresponding to the remote sensing image {r ₁ ,r ₂ ,...,r _N } is:

Among them, {r ₁ ,r ₂ ,...,r _N } is the pixel vector in the image, representing the spectral information in the remote sensing image, N is the total pixel value in the image, and each pixel r _i =[ r _i1 ,r _i2 ,...,r _il ] ^T is an l-dimensional column vector, and l is the number of image bands, 1≤i≤N;

Step 34: Use the Lagrange multiplier method to convert formula (1) into an unconstrained optimization mathematical model, the formula is:

F(w)= ^wT Rw+λ( ^dTw -1) (5),

Among them, λ is the Lagrangian multiplier;

Step 35: Convert formula (5) into a linear equation equation mathematical model, the formula is:

Gs(t)=b(6),

Among them, the autocorrelation coefficient matrix G=[2R,d ^T ; d,0]∈R ^(l+1)×(l+1) ; b is the coefficient vector and b=[0,1] ^T ∈ R ^l+1 ;s(t)=[w(t),λ(t)] ^T ∈R ^l+1 is the vector to be sought; w(t)={w ₁ (t),w ₂ (t),…,w _l (t)} ^T is an l-dimensional vector composed of filter coefficients, and λ(t)∈R is a Lagrange function multiplier;

Step 36: Define the error function of formula (6) as:

e(t)=Gs(t)-b (7),

Step 37: According to the formula (7), the recursive formula for constructing the integral enhanced gradient is:

Step 38: Carry out recursive calculation according to the formula (8), until the calculation ends when the calculated error is less than the allowable error, and the filter output coefficient w(t) is obtained;

Step 39: Perform inversion according to the obtained filter output coefficients, and output the enhanced remote sensing image.

Further, the specific operation steps of step 4 include:

Step 41: For the enhanced remote sensing image, use the Otsu algorithm to calculate the optimal threshold for feature extraction, and perform threshold segmentation on the remote sensing image according to the obtained threshold, extract part of the feature, and obtain the Otsu threshold extraction result;

Step 42: Extract the spectral value at the same position as the Otsu threshold value extraction result on the single band of the remote sensing image, perform threshold segmentation on the single band gray value, remove some interfering objects, and obtain the single band threshold value extraction result.

Further, the specific operation steps of step 5 include:

Step 51: Based on the texture characteristics of the target object, establish the ratio bands between the bands of the remote sensing image, select the sensitive bands by using the Bhattacharyian distance method, and establish a gray-level co-occurrence texture matrix on the basis of the sensitive bands, and delineate the threshold value to carry out Target extraction, removing interfering objects;

Step 52: Based on the spatial attributes of the target ground objects, perform target extraction from the area, extension, perimeter, compactness, solidity, shape elements, and roundness features of the ground objects, and remove interfering objects.

Compared with the prior art, this method has the following beneficial effects:

First, the method proposed in the present invention uses the improved constrained energy minimization algorithm to enhance the target features in the breeding area, weaken the background spectral information, and after obtaining the preliminary extraction results, use the gray-scale co-occurrence texture matrix or object-oriented method to Customized elimination of ground object interference, even in the breeding area where the phenomenon of "different objects with the same spectrum" is more obvious, this method can still effectively overcome the background ground object interference, and obtain higher-precision breeding area extraction results;

Second, this application can use the spectral features of the ground objects to extract the ground features in the breeding area, use the threshold value calculated by the Otsu algorithm and the single-band threshold to perform threshold segmentation, extract some target feature areas, and can extract some target feature areas according to the distance between the ground features in the breeding area. The texture features and geometric features of the ground objects are customized to eliminate the interfering objects, which further improves the accuracy of extraction;

Third, the method of the present invention converts the constrained energy minimization algorithm into an unconstrained optimization mathematical model using the Lagrangian multiplier method, uses the error function and the evolution formula to converge the calculation error of each spectrum to zero, and the classification accuracy is improved.

In summary, the present invention combines the actual application effects of current satellite remote sensing images and the actual needs of offshore aquaculture areas, and constructs a multi-feature method that combines the spectral features of remote sensing images, threshold methods, and gray-scale co-occurrence texture matrices. Analytical method, using this method to achieve accurate extraction of breeding areas.

Description of drawings

Fig. 1 is a schematic flow sheet of the inventive method;

Figure 2 is a remote sensing image of the raft breeding area in Zhanjiang Harbor;

Fig. 3 is the result figure extracted by this method in Zhanjiang harbor raft type culture area;

Figure 4 is a remote sensing image of the Zhanjiang harbor culture pond area;

Fig. 5 is the result map of the Zhanjiang harbor culture pond area extracted by this method.

detailed description

In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Referring to accompanying drawings 1-5, it can be seen that the method for extracting remote sensing images of coastal aquaculture areas based on multi-feature and spectral fusion includes the following steps:

Step 1: Input an original remote sensing image;

Step 2: Carry out image preprocessing on the input remote sensing image, separate the sea and land from the remote sensing image of the farming area to be extracted, cut and retain the farming area to be extracted, and construct the characteristic spectrum according to the features of the ground objects in the processed image , use the Bhattacharyian distance method to screen the sensitive bands in the characteristic spectral bands and the original bands of remote sensing images, and construct the target feature set based on the obtained characteristic spectra;

Step 3: Construct a finite impulse response linear filter, based on the constrained energy minimization algorithm of the gradient integral recurrent neural network, select the spectral data of the target object pixel, and use the finite impulse response linear filter to suppress the background information of the image and at the same time Enhanced object spectral data to obtain enhanced remote sensing images;

Step 4: For the enhanced remote sensing image, use the Otsu algorithm and single-band threshold to perform preliminary extraction of the target feature, and obtain the remote sensing image after the preliminary extraction of the target feature;

Step 5: Use the object-oriented method or the gray-level co-occurrence texture matrix to remove the disturbing objects of the ground features in the breeding area based on the texture features and geometric features of the ground features after the preliminary extraction;

Step 6: Output the final extracted remote sensing image of the breeding area.

Further, the specific operation steps of step 3 include:

Step 31: Construct a finite impulse response linear filter according to the known prior information of the target pixel spectrum;

Step 32: Express the constrained energy minimization algorithm as a linear constrained optimization mathematical model, the expression is:

d表示约束条件向量，并且d满足条件：Among them, w represents the filtering coefficient, R is the autocorrelation matrix of the remote sensing image, and

d represents a vector of constraints, and d satisfies the condition:

Step 33: Under the constraint condition vector, when the output y _i of the filter for the input _ri satisfies the following formula:

Then the average output energy corresponding to the remote sensing image {r ₁ ,r ₂ ,...,r _N } is:

Among them, {r ₁ ,r ₂ ,...,r _N } is the pixel vector in the image, representing the spectral information in the remote sensing image, N is the total pixel value in the image, and each pixel r _i =[ r _i1 ,r _i2 ,...,r _il ] ^T is an l-dimensional column vector, and l is the number of image bands, 1≤i≤N;

Step 34: Use the Lagrange multiplier method to convert formula (1) into an unconstrained optimization mathematical model, the formula is:

F(w)= ^wT Rw+λ( ^dTw -1) (5),

Among them, λ is the Lagrangian multiplier;

Step 35: Convert formula (5) into a linear equation equation mathematical model, the formula is:

Gs(t)=b(6),

Among them, the autocorrelation coefficient matrix G=[2R,d ^T ; d,0]∈R ^(l+1)×(l+1) ; b is the coefficient vector and b=[0,1] ^T ∈ R ^l+1 ;s(t)=[w(t),λ(t)] ^T ∈R ^l+1 is the vector to be sought; w(t)={w ₁ (t),w ₂ (t),…,w _l (t)} ^T is an l-dimensional vector composed of filter coefficients, and λ(t)∈R is a Lagrange function multiplier;

Step 36: Define the error function of formula (6) as:

e(t)=Gs(t)-b (7),

Step 37: According to the formula (7), the recursive formula for constructing the integral enhanced gradient is:

Step 38: Carry out recursive calculation according to the formula (8), until the calculation ends when the calculated error is less than the allowable error, and the filter output coefficient w(t) is obtained;

Step 39: Perform inversion according to the obtained filter output coefficients, and output the enhanced remote sensing image.

Further, the specific operation steps of step 4 include:

Step 41: For the enhanced remote sensing image, use the Otsu algorithm to calculate the optimal threshold for feature extraction, and perform threshold segmentation on the remote sensing image according to the obtained threshold, extract part of the feature, and obtain the Otsu threshold extraction result;

Step 42: Using the characteristic that the gray value of some interfering objects in a single band is smaller than the target feature, extract the spectral value at the same position as the Otsu threshold extraction result on the single band of the remote sensing image, and threshold the single band gray value Segmentation, remove some interfering objects, and obtain single-band threshold extraction results;

Further, the specific operation steps of step 5 include:

Step 51: Based on the texture characteristics of the target object, establish the ratio bands between the bands of the remote sensing image, select the sensitive bands by using the Bhattacharyian distance method, and establish a gray-level co-occurrence texture matrix on the basis of the sensitive bands, and delineate the threshold value to carry out Target extraction, removing interfering objects;

Step 52: Based on the spatial attributes of the target ground objects, perform target extraction from the area, extension, perimeter, compactness, solidity, shape elements, and roundness features of the ground objects, and remove interfering objects.

Example

1. Test process

Firstly, use the sensor to obtain the original remote sensing image as shown in Figure 2 and Figure 4, preprocess the original remote sensing image, retain the breeding area to be extracted, and then use the NDAI characteristic index, suspended sediment difference index TSM and chlorophyll a The concentration index CHL constructs the characteristic index, and combines the 4 original bands of the remote sensing image, and uses the Bhattachary distance method to screen the 7 bands to construct the spectral feature set of the ground objects in the breeding area. Among them, the spectral feature set of the raft culture area is mainly composed of red light band, near red wave band and CHL wave band, while the culture pond area is composed of blue light wave band, CHL wave band and red light wave band;

Secondly, input the pixel values of the target surface objects in the raft culture area and the culture pond area respectively, and use the constrained energy minimization algorithm based on the gradient integral recursive neural network to perform step 31-step 39 to perform image enhancement processing, and obtain the enhanced image Remote Sensing Image;

Among them, after the image enhancement processing in the raft farming area, there is a clear difference between the gray scale of the raft frame pixel and the nearby water body, the raft frame and the surrounding water body are effectively distinguished, and the edge profile of the raft frame is effectively enhanced. There are obvious differences between the water body of the aquaculture pond and the sea water in the aquaculture pond area after the image enhancement, and it is easy to distinguish.

Thirdly, the preliminary extraction of the target features is carried out. First, the threshold value calculated by the Otsu algorithm is used to perform threshold segmentation on the enhanced remote sensing image, and part of the features are extracted as the roughly extracted target feature area. After the Otsu algorithm threshold segmentation in the raft farming area, a lot of seawater remained in the raft frame area, and a large number of small artificial ground objects remained in the farming pond area. Then, using the characteristic that the gray value of some interfering objects in a single band is smaller than the target object, the spectral value at the same position as the Otsu algorithm result is extracted on a single band of the image, and the single band gray value is thresholded. , to extract some interfering objects. Among them, the threshold segmentation is performed in the blue light band in the raft farming area, and the threshold segmentation is performed in the green light band in the culture pool area.

Finally, customized elimination of interfering objects is performed according to the texture and geometric features of the ground objects in the breeding area. For example, there is a relatively obvious difference between the texture characteristics of the target ground objects in the raft farming area and the texture of the water body, and the mean value feature in the gray-level symbiotic texture matrix is used to remove the residual water body in the raft farming area. Most of the aquaculture ponds in the aquaculture pond area are rectangular or square, and the curvature of the river is more obvious, so the geometric characteristics of the river are used to eliminate them. Among them, the raft farming area is mainly based on the ratio of the green light band to the red light wave, and the ratio of the near-red band to the blue light band to generate the mean value feature of the gray-scale co-occurrence texture matrix, while the breeding pool area uses the Rect_Fit attribute and the Elongation attribute in the object-oriented method. Extract the breeding area; wherein, there are many sporadic tiny spots in the final extraction result of the breeding area. The present invention uses the existing spot grouping method to filter the classification results to remove the sporadic spots.

Finally, referring to accompanying drawings 3 and 5, it can be seen that the result map extracted according to this method.

2. Result analysis

Because the present invention adopts the method of stratified random sampling, establishes verification samples on the field sampling or visual interpretation results of higher spatial resolution images, and applies the confusion matrix method to verify the accuracy of the extraction results, therefore, this method is used in Google Earth The visual interpretation results of the software refer to the assignment of the attributes of each sample point in the ENVI software, and the construction of a confusion matrix and accuracy evaluation.

As can be seen from the various precision tables of the breeding area shown in Table 1, for the extraction results of the GF-1 image raft farming area, the method of the present invention effectively overcomes the high accuracy on the basis of making full use of the spectral features and texture features of the ground objects. Influenced by turbidity water body, the outline of the raft frame is clear, and the overall accuracy reaches 98.74%. According to the extraction results of the ZY-3 image cultivation pond area, the multi-feature analysis method uses the geometric difference between the river and the cultivation pond to achieve high-precision extraction of the cultivation pond , with an overall accuracy of 96.74%.

Table 1 Various types of accuracy tables in breeding areas

In summary, no matter which type of farming is mentioned above, the method of the present invention has significant comparative advantages in overcoming "same spectrum and different spectrum" and "same spectrum and different spectrum" of ground features. After the method is processed, the extraction results are clear and the accuracy rate is high, so it can be used as the optimal algorithm for the extraction of breeding areas in high-resolution remote sensing images.

The content not described in detail in this specification belongs to the prior art known to those skilled in the art. Although the patent of the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments, or perform equivalent replacements for some of the technical features. Within the spirit and principles of the present invention, any modifications, equivalent replacements, improvements, etc., shall be included within the protection scope of the present invention.

Claims (3)

1. The method for extracting the remote sensing image of the near-shore aquaculture area based on multi-feature and spectrum fusion is characterized by comprising the following steps of:

step 1: inputting an original remote sensing image;

step 2: carrying out image preprocessing on an input remote sensing image, extracting a characteristic spectrum by using a characteristic index method aiming at the feature of the ground object in the processed image, and constructing a target ground object feature set according to the obtained characteristic spectrum;

and step 3: constructing a finite impulse response linear filter, inputting target ground object pixel spectral data based on a constrained energy minimization algorithm of a gradient integral recurrent neural network, and enhancing the target ground object spectral data by using the finite impulse response linear filter to obtain an enhanced remote sensing image;

and 4, step 4: performing preliminary extraction on the target ground object by using an Otsu algorithm and a single-band threshold value on the enhanced remote sensing image to obtain a preliminarily extracted remote sensing image of the target ground object;

and 5: the preliminarily extracted remote sensing image is subjected to customized elimination on ground feature interference objects in the culture area by adopting an object-oriented method or a gray level symbiotic texture matrix based on the texture features and the geometric features of the ground features;

and 6: outputting the final remote sensing image extracted from the culture area;

the specific operation steps of the step 3 comprise:

step 31: constructing a finite impulse response linear filter according to the prior information of the known target pixel spectrum;

step 32: the constrained energy minimization algorithm is expressed as a linear constrained optimization mathematical model, and the expression is as follows:

wherein w represents the substitution filter coefficient, R is the autocorrelation matrix of the remote sensing image, and

d represents a constraint condition vector, and d satisfies the condition:

step 33: under the constraint condition vector, when the filter is corresponding to the input r _i Output y of _i Satisfies the following formula:

then the remote sensing image r ₁ ,r ₂ ,...,r _N The corresponding average output energy is:

wherein, { r ₁ ,r ₂ ,...,r _N The pixel vector in the image represents the spectral information in the remote sensing image, N is the total pixel value in the image, and each pixel r _i ＝[r _i1 ,r _i2 ,...,r _il ] ^T Is a l-dimensional column vector, wherein l is the wave band number of the image, and i is more than or equal to 1 and less than or equal to N;

step 34: converting the formula (1) into an unconstrained optimization mathematical model by using a Lagrange multiplier method, wherein the formula is as follows:

F(w)＝w ^T Rw+λ(d ^T w-1) (5)，

wherein λ is a Lagrange multiplier;

step 35: converting equation (5) into a mathematical model of a linear equation, wherein the equation is as follows:

Gs(t)＝b (6)，

wherein the autocorrelation coefficient matrix G = [2R ^T ；d,0]∈R ^(l+1)×(l+1) (ii) a b is a coefficient vector and b = [0,1 =] ^T ∈R ^l+1 ；s(t)＝[w(t),λ(t)] ^T ∈R ^l+1 The vector is to be solved; w (t) = { w ₁ (t),w ₂ (t),…,w _l (t)} ^T Is a vector of dimension l formed by filter coefficients, and lambda (t) belongs to R and is a Lagrange function multiplier;

step 36: the error function defining equation (6) is:

e(t)＝Gs(t)-b (7)，

step 37: according to equation (7), the integrated enhanced gradient recurrence equation is constructed as:

step 38: performing recursive calculation according to the formula (8) until the calculated error is smaller than the allowable error, and obtaining a filter output coefficient w (t);

step 39: and performing inversion according to the obtained filter output coefficient, and outputting the enhanced remote sensing image.

2. The near-shore aquaculture area remote sensing image extraction method based on multi-feature and spectrum fusion of claim 1, characterized in that the specific operation steps of step 4 comprise:

step 41: aiming at the enhanced remote sensing image, calculating an optimal threshold value for ground feature extraction by adopting an Otsu algorithm, carrying out threshold segmentation on the remote sensing image according to the obtained threshold value, extracting partial ground features and obtaining an Otsu threshold value extraction result;

step 42: and extracting a spectral value with the same position as the Otsu threshold extraction result on a single waveband of the remote sensing image, performing threshold segmentation on the single waveband gray value, and removing part of interference objects to obtain a single waveband threshold extraction result.

3. The near-shore aquaculture area remote sensing image extraction method based on multi-feature and spectrum fusion of claim 1, characterized in that the specific operation steps of step 5 comprise:

step 51: establishing specific wave bands among wave bands of the remote sensing image based on the texture features of the target ground object, selecting sensitive wave bands by adopting a Babbitt distance method, establishing a gray level co-occurrence texture matrix on the basis of the sensitive wave bands, defining a threshold value for target extraction, and eliminating interference objects;

step 52: based on the spatial attributes of the target ground objects, the target is extracted from the characteristics of the area, extensibility, perimeter, compactness, firmness, shape elements and roundness among the ground objects, and the interference objects are removed.

Images