High-resolution image shadow information processing method

Abstract

The invention belongs to the field of shadow information processing. In order to solve the problems of large calculation amount and single analysis scale of the existing high-resolution image shadow information processing, the invention provides a high-resolution image shadow information processing method, which comprises the following steps of: step S01, dividing the response of an original image to a second-order Gaussian differential filter bank into seven different texture primitives; and S02, a basic image feature is formed by a set of seven different texture primitives in the step S01. And step S03, classifying the basic image features. According to the method, the texture information and the spectral information of the high-resolution image are combined, so that the separability of the sample can be effectively improved, and better classification precision is obtained. The method still has better performance in the extraction of most vegetation types in areas with lower reflectivity and lower DN value. And the overall precision is improved to 86.26% from the original 83.97%.

Description

Technical Field

[0001] The invention belongs to the field of shadow information processing and relates to a high-resolution image shadow information processing method. Background Art

[0002] The surface coverage plays a vital role in national or regional development decisions. The rapid development of remote sensing technology provides a strong guarantee for the rapid and accurate acquisition of surface coverage. The shadow problem has always been an important aspect of remote sensing science. It is caused by the partial or complete obstruction of direct light from the light source by the target. The elevation of objects in small areas within the city changes rapidly. The existence of a large number of buildings, bridges, towers, and street trees leads to a large proportion of shadows in urban images, which will affect the recognition of object information. The same problem also occurs in areas with large terrain undulations. Compared with the shadows inside the city, the shadows in areas with large terrain undulations still account for a large proportion of the image, but are more continuous.

[0003] The introduction of texture features to improve the quality of shadow processing in high-resolution images has been widely recognized. The texture features of ground objects will not change with the generation of shadows, that is, the texture features of ground objects maintain a certain continuity in the shadow area and the non-shadow area. Texture is an important method for identifying targets. It is a very important and difficult to describe feature in the image, and there is no generally accepted definition yet. In the process of classifying and identifying texture images, many textures are similar in appearance, but they are diverse in angle, direction, and light and dark changes. That is to say, when the same image is rotated while keeping the appearance unchanged, they should still belong to the same category. When extracting such images, the obtained eigenvalues ​​should be as close as possible. In response to this problem, many methods that can obtain rotation invariant features have been proposed, such as: geometric moments, orthogonal moments, gray-level co-occurrence matrix, etc. The first two methods are very computationally intensive. The gray-level co-occurrence matrix is ​​widely accepted and adopted. It is an important method for analyzing texture features, but it also has problems such as a single analysis scale and a large amount of calculation. Summary of the invention

[0004] 1. Technical problems to be solved:

[0005] The existing high-resolution image shadow information processing has large computational complexity and a single analysis scale.

[0006] 2. Technical solution:

[0007] In order to solve the above problems, the present invention provides a high-resolution image shadow information processing method, comprising the following steps:

[0008] Step S01: The response of the original image to the second-order Gaussian differential filter group is divided into 7 different texture primitives.

[0009] Step S02: The set of 7 different texture primitives in step S01 constitutes the basic image features.

[0010] Step S03: Classify basic image features.

[0011] In step S01, the fused image is orthorectified using the RPC file of the panchromatic image to obtain an orthorectified result of the fused image.

[0012] Use the Pansharpening fusion method in ENVI, which is specifically designed for high-resolution images.

[0013] In step S01, the expression of the one-dimensional Gaussian differential filter is:

[0014]

[0015]

[0016] In the above formula, σ is the scale parameter of the filter; n represents the order of the derivative; H n is the nth-order Hermite polynomial.

[0017] The one-dimensional filter is extended to two dimensions and expressed as:

[0018]

[0019] The spatial derivative of the image response to the Gaussian differential filter bank is:

[0020]

[0021] Where I(x, y) represents the dn value of the image point (x, y).

[0022] Compute using the image's response to a Gaussian derivative filter:

[0023] S mn =σ m+n j mn 0≤m+n≤2

[0024] Where m and n represent the order of the derivative respectively.

[0025] For S mn Perform scale normalization and then calculate the following two parameters:

[0026]

[0027]

[0028] The collection of texture primitives is represented as:

[0029]

[0030] Among them, ε is a fixed parameter, and its value is in (0,1).

[0031] The seven different texture primitives are: flat, slope, dark spot, bright spot, dark line, bright line, and saddle, and the corresponding formulas are:

[0032] εS 00 , λ, -λ, γ.

[0033] In step S03, the classification method is to select the optimal feature using the maximum value, specifically:

[0034] Step S21: For each pixel point, the number of the BIF feature with the largest value among the seven BIF values ​​is obtained.

[0035] Step S22: performing K-means clustering algorithm.

[0036] Step S23: Obtain classification results.

[0037] In step S03, the classification method is to select the optimal feature combination using PCA transformation, specifically:

[0038] Step S31: Perform principal component analysis on the seven bands of BIF.

[0039] Step S32: principal component analysis information on the first few bands whose cumulative contribution rate exceeds 95%.

[0040] Step S33: Perform SVM classification.

[0041] In step S33, the specific steps of SVM classification are:

[0042] Step S71: Select the band. Use the fused near-infrared band image to extract texture features.

[0043] Step S72: extracting BIF features.

[0044] Step S73: Establish a training area.

[0045] Step S74: Check the separability of the training area: After selecting the training sample area and the optimal band, the class separability index JM distance method is used to calculate the class separability. The value range of the JM distance is [0-2]. The larger the value, the easier the classification between classes, and vice versa.

[0046] Step S75: Classification is performed using a SVM classifier with the radial basis function as the kernel function of the SVM.

[0047] In step S72, the specific method is: select the best classification band. Extract the 7 BIF features of the study area image, then perform principal component analysis on the 7 BIF features, select the first three principal components of the analysis results, and select the classification bands of the fused red, green, blue, near-infrared bands and the first three bands of the principal component analysis.

[0048] In step S73, specifically: establish classification areas for fir in non-shadow area, pine in non-shadow area, broad-leaved plants in non-shadow area, fir in shadow area, pine in shadow area, broad-leaved plants in shadow area, bamboo forest, and shrubs.

[0049] 3. Beneficial effects:

[0050] The present invention combines the texture information and spectral information of high-resolution images, which can effectively improve the separability of samples and obtain better classification accuracy. In the extraction of most vegetation types, it still has good performance in areas with low reflectivity and low DN values. The overall accuracy is increased from the original 83.97% to 86.26%. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the present invention.

[0052] Figure 2 is a schematic diagram of the seven texture primitives. DETAILED DESCRIPTION

[0053] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 As shown, a high-resolution image shadow information processing method comprises the following steps:

[0055] Step S01: The response of the original image to the second-order Gaussian differential filter group is divided into 7 different texture primitives.

[0056] Step S02: The set of 7 different texture primitives in step S01 constitutes the basic image features.

[0057] Step S03: Classify basic image features.

[0058] In one embodiment, in step S01, the image is orthorectified and atmospherically corrected. The Quick Atmospheric Correction Tool (QUAC) is used to remove some of the atmospheric effects, and then the fused image is orthorectified using the RPC file of the panchromatic image. The accuracy of the orthorectified result of the fused image is consistent with that of the orthorectified result of the panchromatic image under the same conditions. Such a sequence can reduce the process and improve efficiency, and when the panchromatic and multispectral images are fused, accurate spatial registration between them can be guaranteed.

[0059] In one embodiment, the panchromatic and multispectral images are fused using the Pansharpening fusion method in ENVI that is specifically designed for high-resolution images.

[0060] In one embodiment, the response of the image to the second-order Gaussian differential filter bank is divided into 7 different texture primitives, and the set of these 7 different texture primitives is called basic image features. Most textures are composed of these 7 texture primitives.

[0061] In one embodiment, the expression of the one-dimensional Gaussian differential filter is:

[0062]

[0063]

[0064] In the above formula, σ is the scale parameter of the filter; n represents the order of the derivative; H n is the nth-order Hermite polynomial.

[0065] The one-dimensional filter is extended to two dimensions and expressed as:

[0066]

[0067] The spatial derivative of the image response to the Gaussian differential filter bank is:

[0068]

[0069] Where I(x, y) represents the dn value of the image point (x, y).

[0070] Compute using the image's response to a Gaussian derivative filter:

[0071] S mn =σ m+n j mn 0≤m+n≤2

[0072] Where m and n represent the order of the derivative respectively.

[0073] For S mn Perform scale normalization and then calculate the following two parameters:

[0074]

[0075]

[0076] The collection of texture primitives is represented as:

[0077]

[0078] Among them, ε is a fixed parameter, which generally takes a value in (0, 1). The corresponding images are as follows Figure 2 As shown, they represent: flat, slope, dark spot, bright spot, dark line, bright line, saddle type. The corresponding formulas are:

[0079] εS 00 , λ, -λ, γ.

[0080] In one embodiment, in step S03, the classification method is to select the optimal feature using the maximum value, specifically:

[0081] Step S21: For each pixel point, the number of the BIF feature with the largest value among the seven BIF values ​​is obtained.

[0082] Step S22: performing K-means clustering algorithm.

[0083] Step S23: Obtain classification results.

[0084] In one embodiment, in step S03, the classification method is to select the optimal feature combination using PCA transformation, specifically:

[0085] Step S31: Perform principal component analysis on the seven bands of BIF.

[0086] Step S32: principal component analysis information on the first few bands whose cumulative contribution rate exceeds 95%.

[0087] The principal component analysis method is used for feature selection of BIF. As an effective way of information gain and dimensionality reduction, principal component analysis can be applied to the selection of BIF to transform multiple BIF texture features into several comprehensive texture features, which can solve the problem of "mixed" texture primitives. We generally believe that the first few bands with a cumulative information contribution rate of more than 95% after principal component analysis are principal component bands. Generally speaking, the cumulative contribution rate of the first principal component is generally close to 80%. The first few bands after principal component analysis have a large signal-to-noise ratio and relatively small noise, which highlights the main information. In addition, the first few bands after principal component analysis are composed of 7 bands of BIF.

[0088] Step S33: Perform SVM classification.

[0089] The present invention takes the vegetation classification of high-resolution remote sensing images as an example, and selects Lushan, Jiangxi Province as the research area to verify the proposed method. Lushan is located in the southeast of Jiujiang City, Jiangxi Province, and its geographical location is between 115°50' and 116°10' east longitude and 29°28' and 29°45' north latitude. Lushan belongs to the subtropical evergreen broad-leaved forest area and the eastern mid-subtropical evergreen broad-leaved forest belt in the "Chinese Vegetation Zoning". The vegetation of Lushan can be divided into two plant subclasses: coniferous forest and broad-leaved forest. According to the plant subclasses of Lushan, the classification system established in the article is: pine, fir, broad-leaved, shrub, bamboo. Among them, because they are generally distributed on both sides of the mountain, the vegetation types that are most affected by shadows are: pine, fir and broad-leaved. The high-resolution remote sensing image used in this article is Geoeye image, with a spatial resolution of 0.5 meters, the image shooting time is March 25, 2011, and the scope of the image information processing area is: 6 kilometers * 3 kilometers. The verification data is from 2009 to 2010. Lushan is a national nature reserve. The vegetation in the Lushan area is less disturbed by human activities and changes more slowly, although there is a certain lag compared with satellite data.

[0090] First: image orthorectification and atmospheric correction. Use the Quick Atmospheric Correction Tool (QUAC) to remove some of the atmospheric effects, and then use the RPC file of the panchromatic image to orthorectify the fused image. The orthorectified result of the fused image is consistent with the orthorectified result of the panchromatic image under the same conditions. This sequence can reduce the process and improve efficiency, and when fusing panchromatic and multispectral images, it can ensure accurate spatial registration between them.

[0091] : Inter-band fusion. First, the panchromatic and multispectral images are fused using the Pansharpening fusion method in ENVI, which is specially designed for high-resolution images.

[0092] Secondly: select the band. Texture feature extraction is performed on a single band, so it is necessary to select a suitable band for extraction. The vegetation area of ​​the near-infrared image has high brightness, and the texture of the vegetation in this band is more obvious, so it is more appropriate to use the fused near-infrared band image to extract texture features.

[0093] : Extract BIF features and select the best classification band. Extract 7 BIF features of the study area image, then perform principal component analysis on the 7 BIF features and select the first three principal components of the analysis results. The classification bands select the fused red, green, blue, near-infrared bands and the first three bands of the principal component analysis.

[0094] Establish training areas: Establish classification areas for fir in non-shaded areas, pine in non-shaded areas, broad-leaved plants in non-shaded areas, fir in shaded areas, pine in shaded areas, broad-leaved plants in shaded areas, bamboo forests, and shrubs.

[0095] Test the separability of the training area. After selecting the training sample area and the best band (the 7 bands selected above), the class separability index JM (Jeffries-Matusita) distance method is used to calculate the class separability. The value range of JM distance is [0-2]. The larger the value, the easier the classification between classes, and vice versa. Generally, studies believe that if it is greater than 1.8, it means that the samples are separable and belong to qualified samples; if it is less than 1.8, it is necessary to reselect samples; if it is less than 1, consider combining the two types of samples into one type of sample. It can be concluded from Table 1 that the sample area belongs to qualified samples. Among them, shrubs and artificial buildings have a high separability compared to other categories. The JM distance values ​​of pine forests and bamboo forests in the non-shaded area are the smallest, but they are still qualified samples.

[0096] Finally, the SVM classifier is used, with the radial basis function as the kernel function of the SVM.

[0097] Table 1 JM distance of training samples

[0098]

[0099] The experimental results show that the use of BIF combined with PCA to extract texture features can effectively improve the classification accuracy. The classification accuracy of the SVM-based method is 86.26%, which is better than the 83.97% of extracting texture features using gray-level co-occurrence matrix combined with spectral features.

[0100] The experimental results show that for the classification of Geoeye images, the improved BIF can maintain the same accuracy as the gray-level co-occurrence matrix in the extraction of most vegetation types, and in the shadow area where the DN value is overall low, BIF still has a good performance relative to the low accuracy of the gray-level co-occurrence matrix. The accuracy of the producers of fir in the shadow area increased from 54.31% to 68.15%, and the user accuracy increased from 51.36% to 56.49%. The accuracy of the producers of pine in the shadow area increased from 78.70% to 81.07%, and the user accuracy increased from 42.09% to 45.67%. Although the accuracy of the producers of broad-leaved trees in the shadow area decreased by 7%, the user accuracy increased by about 20%.

Claims

1. A high-resolution image shadow information processing method, characterized by: The following steps are involved: Step S01: the response of the original image to the second-order Gaussian differential filter group is divided into 7 different texture primitives; Step S02: The set of 7 different texture primitives in step S01 constitutes the basic image feature; Step S03: Classify basic image features.

2. The high-resolution image shadow information processing method according to claim 1, characterized in that: In step S01, the fused image is orthorectified using the RPC file of the panchromatic image to obtain an orthorectified result of the fused image.

3. The high-resolution image shadow information processing method according to claim 2, characterized in that: Use the Pansharpening fusion method in ENVI, which is specifically designed for high-resolution images.

4. The high-resolution image shadow information processing method according to claim 1, characterized in that: In step S01, the expression of the one-dimensional Gaussian differential filter is: In the above formula, σ is the scale parameter of the filter; n represents the order of the derivative; H n is the n-order Hermite polynomial, which generalizes the one-dimensional filter to two dimensions and is expressed as: The spatial derivative of the image response to the Gaussian differential filter bank is: Where I(x,y) represents the dn value of the image point (x,y). Compute using the image's response to a Gaussian derivative filter: S mn =s m+n j mn 0≤m+n≤2 In the formula, m and n represent the order of the derivative, For S mn Perform scale normalization and then calculate the following two parameters: The collection of texture primitives is represented as: 其中 , ε 为固定参数 , the value is ( 0 , 1 ) 内 .

5. The high-resolution image shadow information processing method according to claim 1, characterized in that: The seven different texture primitives are: flat, slope, dark spot, bright spot, dark line, bright line, and saddle, and the corresponding formulas are: εS 00 , λ, -λ, γ.

6. The high-resolution image shadow information processing method according to any one of claims 1 to 5, characterized in that: In step S03, the classification method is to select the optimal feature using the maximum value, specifically: Step S21: For each pixel point, the number of the BIF feature with the largest value among the seven BIF values ​​is taken; Step S22: Perform K-means clustering algorithm Step S23: Obtain classification results.

7. The high-resolution image shadow information processing method according to any one of claims 1 to 5, characterized in that: In step S03, the classification method is to select the optimal feature combination using PCA transformation, specifically: Step S31: performing principal component analysis on the seven bands of BIF; Step S32: analyzing the first few bands whose cumulative contribution rate of principal component information exceeds 95%; Step S33: Perform SVM classification.

8. The high-resolution image shadow information processing method according to claim 7, characterized in that: In step S33, the specific steps of SVM classification are: Step S71: Select the band: use the fused near-infrared band image to extract texture features; Step S72: extracting BIF features; Step S73: Establishing a training area: Step S74: Check the separability of the training area: After selecting the training sample area and the optimal band, the class separability index JM distance method is used to calculate the class separability. The value range of the JM distance is [0-2]. The larger the value, the easier the classification between classes, and vice versa. Step S75: Classification is performed using a SVM classifier with the radial basis function as the kernel function of the SVM.

9. The high-resolution image shadow information processing method according to claim 8, characterized in that: In step S72, the specific method is: select the best classification band, extract the 7 BIF features of the study area image, then perform principal component analysis on the 7 BIF features, select the first three principal components of the analysis results, and select the fused red, green, blue, and near-infrared bands and the first three bands of the principal component analysis as the classification band.

10. The high-resolution image shadow information processing method according to claim 8, characterized in that: In step S73, specifically: establish classification areas for fir in non-shadow area, pine in non-shadow area, broad-leaved plants in non-shadow area, fir in shadow area, pine in shadow area, broad-leaved plants in shadow area, bamboo forest, and shrubs.

Images