Method for Extracting Ground Object Spectral Reflectance Image Based on Multi-Temporal Intrinsic Image Decomposition

Abstract

Method for extracting ground object spectral reflectance image based on multi-temporal intrinsic image decomposition. The present invention relates to the extraction of reflectance information from multi / hyperspectral remote sensing images. The purpose of the present invention is to solve the problems of low extraction accuracy and poor robustness of the existing multi / hyperspectral remote sensing image intrinsic reflectance information. The process is as follows: First, establish an expression model for the intrinsic reflectance information of multi-temporal remote sensing images; Second, construct a decomposition constraint model under local spatio-temporal energy constraints; Third, transform the optimal solution of the decomposition constraint model into an iterative optimization solution model for the common reflectance and two shadow components; Fourth, given multi-temporal remote sensing images and initial parameters, based on the iterative optimization solution model in step three, obtain the final intrinsic reflectance. The present invention is used in the field of digital image processing.

Description

Technical Field

[0001] The present invention relates to the extraction of reflectance information from multi / hyperspectral remote sensing images. Background Art

[0002] The spectral reflectance information of ground objects is the main means for quantitative inversion of remote sensing images (such as the normalized difference vegetation index, the normalized difference water index), and is also an important way for identifying and interpreting ground objects using remote sensing images. However, the imaging process of multi / hyperspectral remote sensing images is very complex. In addition to the spectral reflectance information of ground objects, the image also contains information such as the shading component and the illuminance component caused by the spatial undulation distribution of ground objects. Therefore, inverting the spectral reflectance information from the image is a typical ill-posed decomposition problem, and the reliability of the extracted reflectance is not high. In addition, due to different imaging factors for different images (solar altitude angle, atmospheric environment), the extraction of reflectance under different images is not stable enough, seriously affecting the remote sensing applications of multi / hyperspectral.

[0003] Currently, the methods for inverting the spectral reflectance information of ground objects from multi / hyperspectral remote sensing images mainly include relative spectral correction and reflectance information inversion based on intrinsic image decomposition. Since relative spectral correction requires that the reference image must be the reflectance information extracted after strict correction, its application range is small. Intrinsic image decomposition is a typical ill-posed decomposition problem, and its robustness and reflectance extraction accuracy are insufficient.

[0004] In order to improve the accuracy of ground object reflectance extraction from multi / hyperspectral remote sensing images based on intrinsic image decomposition, it can be solved by starting from increasing the prior parameters input in the intrinsic decomposition. Increasing the number of input remote sensing images, that is, constructing a multi-temporal intrinsic image decomposition method can theoretically significantly improve the accuracy and robustness of reflectance extraction. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of low accuracy and poor robustness in the extraction of intrinsic reflectance information from existing multi / hyperspectral remote sensing images, and to propose a method for extracting the spectral reflectance image of ground objects based on multi-temporal intrinsic image decomposition.

[0006] The specific process of the method for extracting the spectral reflectance image of ground objects based on multi-temporal intrinsic image decomposition is as follows:

[0007] Step 1: Establish an expression model for the intrinsic reflectance information of multi-temporal remote sensing images, where the multi-temporal remote sensing images are composed of the same reflectance component and independent shading components;

[0008] Step 2: Based on the intrinsic reflectance information expression model of multi-temporal remote sensing images, construct a decomposition constraint model under local spatio-temporal energy constraints. By minimizing the decomposition constraint model under local spatio-temporal energy constraints, obtain the optimal solution of the decomposition constraint model, and invert the common reflectance R and the shading components of each image;

[0009] Step 3: Convert the optimal solution of the decomposition constraint model into an iterative optimization solution model for the common reflectance and two shading components;

[0010] Step 4: Given multi-temporal remote sensing images and initial parameters, based on the iterative optimization solution model in Step 3, calculate the final intrinsic reflectance;

[0011] The initial parameters are the local neighborhood scale range, the neighborhood similarity regulation parameter, the adjacent iteration parameter error threshold, and the initial values of the reflectance R and the shading components of each image;

[0012] The initial values of the shading components of each image are set as all-zero matrices.

[0013] The beneficial effects of the present invention are as follows:

[0014] The purpose of the present invention is to improve the extraction accuracy and stability of the intrinsic reflectance information of existing multi- / hyperspectral remote sensing images, and a precise extraction method for the ground object spectral reflectance image based on multi-temporal intrinsic image decomposition is proposed. This method constructs an intrinsic information expression, spatio-temporal joint optimization and inversion mechanism under multi-temporal conditions by adding more image prior information, thereby improving the accuracy and robustness of information extraction.

[0015] The method of the present invention extends the traditional single multi- / hyperspectral image intrinsic reflectance extraction model to a multi-temporal (i.e., multi-image) mode. Starting from adding prior information during the decomposition of the intrinsic reflectance, a local spatio-temporal energy constraint model and a corresponding iterative solution method are constructed to achieve high-precision and robust extraction of the ground object reflectance.

[0016] To verify the performance of the present invention, it was verified on a set of real GF-2 multi-temporal remote sensing images. The experimental results show that, compared with the reflectance extraction results of a single image, the reflectance image extracted by multi-temporal intrinsic decomposition is closer to the color of the original image, without color difference, and the spectral reflectance differences of the same type of ground objects are very small. The experimental results verify the effectiveness of the precise extraction method for the ground object spectral reflectance image based on multi-temporal intrinsic image decomposition proposed by the present invention. Description of the Drawings

[0017] Figure 1 It is the flow chart of the present invention;

[0018] Figure 2a It is the GF-2 Phase 1 remote sensing image (only showing the RGB bands);

[0019] Figure 2b For the GF-2 Phase 2 remote sensing image (only the RGB bands are shown);

[0020] Figure 3a For the result of the reflectance extraction of the single image of GF-2 Phase 1 (only the RGB bands are shown);

[0021] Figure 3b For the result of the reflectance extraction of the single image of GF-2 Phase 2 (only the RGB bands are shown);

[0022] Figure 3c For the result of the reflectance extraction under multi-temporal eigen decomposition (only the RGB bands are shown). Detailed implementation manners

[0023] Detailed implementation manner 1: The specific process of the method for extracting the ground object spectral reflectance image based on multi-temporal eigen image decomposition in this implementation manner is as follows:

[0024] Step 1: Establish an eigen reflectance information expression model for multi-temporal remote sensing images, where the multi-temporal remote sensing images are composed of the same reflectance components and independent shading components;

[0025] Step 2: Based on the eigen reflectance information expression model for multi-temporal remote sensing images, construct a decomposition constraint model under local spatio-temporal energy constraints. By minimizing the decomposition constraint model under local spatio-temporal energy constraints, obtain the optimal solution of the decomposition constraint model (generally, the solution at the extreme point of the constraint model when solving the optimization function is usually the optimal solution), and invert the common reflectance R and the shading components of each image to realize the inversion of the reflectance parameters in the decomposition constraint model;

[0026] Step 3: For the minimization of the local spatio-temporal energy constraint decomposition model proposed in Step 2, based on the idea that the zero point of the first-order differential reciprocal is the best extreme point, transform the optimal solution of the decomposition constraint model into an iterative optimization solution model for the common reflectance and two shading components (the reflectance R of the multi-temporal remote sensing image, the shading component S of the Phase 1 remote sensing image I 1 ), the shading component S of the Phase 2 remote sensing image I 1 ), the shading component S of the Phase 2 remote sensing image I 2 ); 2 )

[0027] Step 4: Given multi-temporal remote sensing images (Image 1 and Image 2) and initialization parameters, based on the iterative optimization solution model in Step 3, find the final eigen reflectance (the reflectance R of the multi-temporal remote sensing image, the shading component S of the Phase 1 remote sensing image I 1 ), the shading component S of the Phase 2 remote sensing image I 1 ), the shading component S of the Phase 2 remote sensing image I 2The shading component S 2 );

[0028] The initialization parameters are the local neighborhood scale range (manually set), the neighborhood similarity regulation parameter, the adjacent iteration parameter error threshold, the reflectance R, and the initial values of the shading components of each image;

[0029] The initial values of the shading components of each image are set to all-zero matrices.

[0030] Specific Embodiment 2: The difference between this embodiment and Specific Embodiment 1 is that in step 1, an intrinsic reflectance information expression model for multi-temporal remote sensing images is established; the specific process is as follows:

[0031] Collect multi-temporal multi- / hyperspectral remote sensing images at the same location at different times (here only the case of two-temporal remote sensing images is shown). The intrinsic reflectance information expression model for multi-temporal remote sensing images is:

[0032] I 1 = S 1 R

[0033] I 2 = S 2 R

[0034] where I 1 is the remote sensing image of phase 1, I 2 is the remote sensing image of phase 2, R is the common reflectance of the multi-temporal remote sensing images, and the sizes of I 1 , I 2 and R are all m rows and n columns in the B band. S 1 is the shading component of the remote sensing image I 1 of phase 1, S 2 is the shading component of the remote sensing image I 2 of phase 2, and the sizes of S 1 and S 2 are both m rows and n columns (the size is different from the image size because all bands share one shading component); B is the number of spectral bands of the multi- / hyperspectral remote sensing image, generally ranging from 4 to several thousand.

[0035] Other steps and parameters are the same as those in Specific Embodiment 1.

[0036] Specific Embodiment 3: The difference between this embodiment and Specific Embodiment 1 or 2 is that in step 2, the decomposition constraint model under local spatio-temporal energy constraint is expressed as:

[0037] To achieve the inversion and extraction of multi-temporal intrinsic image reflectance, the above model is still an ill-posed problem model (given I 1 and I 2, find R and S 1 and S 2 ), To achieve the decomposition of the above ill-posed problem, we present two prior conditions to construct a constraint model. One is the similarity of the energy of the reflectance within a local range, and the other is that the proposed reflectance and the shading components of each image must conform to the information representation model described in Embodiment 2; the decomposition constraint model under the above local spatio-temporal energy constraint can be expressed in the following form:

[0038]

[0039] Among them, E represents the local spatio-temporal energy; P is all the pixels in the image, the number of which is m×n, and N(i) represents the pixels within the local neighborhood (the local neighborhood refers to 3×3, 5×5, 7×7 or 9×9) around the pixel point i;

[0040] The first term on the right side of the above equation is the local energy similarity constraint, and the second and third terms are the signal model representation relationship constraints on each image. In the formula, w ij represents the neighborhood similarity between pixel i and its neighboring point j (this only takes the neighboring points within the local neighborhood range, usually the local neighborhood is taken as 3×3, 5×5, 7×7 or 9×9). The formula for calculating the neighborhood similarity is as follows:

[0041]

[0042] Among them, I ib represents the gray value of the b-th band at the i-th point of image I, and I jb represents the gray value of the b-th band at the j-th point of image I, and σ is the neighborhood similarity regulation parameter;

[0043] By minimizing the decomposition constraint model under the above local spatio-temporal energy constraint, the optimal solution of the decomposition constraint model is obtained (generally, the solution at the extreme point of the constraint model when solving the optimization function is usually the optimal solution), and the common reflectance R and the shading components of each image are inverted to realize the inversion of the reflectance parameters in the decomposition constraint model;

[0044] That is:

[0045]

[0046]

[0047] Among them, R b represents the gray value of the b-th band of the reflectance R.

[0048] Other steps and parameters are the same as those in Embodiment 1 or 2.

[0049] Embodiment 4: The difference between this embodiment and any one of Embodiments 1 to 3 is that in Step 3, the optimal solution of the decomposed constraint model is transformed into an iterative optimization solution model of the common reflectance and two shading components (the reflectance R of the multi-temporal remote sensing image, the shading component S of the remote sensing image I at Phase 1 1 of 1 , the remote sensing image I at Phase 2 2 of the shading component S 2 ); The specific process is as follows:

[0050] To solve the minimization energy constraint given in Step 3, using the idea that the zero point of the first derivative of the above formula is the best extreme point, the iterative solution formulas for R, S 1 and S 2 are obtained. The obtained iterative optimization solution model is as follows:

[0051]

[0052]

[0053]

[0054]

[0055] Among them, represents the reciprocal of the shading component of pixel i of the remote sensing image I at Phase 1 1 ; I 1ib represents the gray value of the i-th point and the b-th band of the remote sensing image I at Phase 1 1 ; R 1ib represents the reflectance of the i-th point and the b-th band of the remote sensing image at Phase 1; represents the reciprocal of the shading component of pixel i of the remote sensing image I at Phase 2 2 ; I 2ib represents the gray value of the i-th point and the b-th band of the remote sensing image I at Phase 2 2 ; R 2ib represents the reflectance of the i-th point and the b-th band of the remote sensing image at Phase 2; R ib represents the reflectance of the i-th point and the b-th band of the multi-temporal remote sensing image; R jb represents the reflectance of the j-th point and the b-th band of the multi-temporal remote sensing image; w ji represents the similarity between pixel j and its neighboring point i (3x3, 5x5, 7x7 or 9x9); N(j) represents the pixels within the local neighborhood around pixel point j (the local neighborhood refers to 3x3, 5x5, 7x7 or 9x9); w jk represents the similarity between pixel j and its neighboring point k; R kb represents the reflectance of the k-th point and the b-th band of the multi-temporal remote sensing image.

[0056] Embodiment 5: The difference between this embodiment and any one of Embodiments 1 to 4 is that in step 4, given multi-temporal remote sensing images (Image 1 and Image 2) and initialization parameters, based on the iterative optimization solution model in step 3, the final intrinsic reflectance (the reflectance R of the multi-temporal remote sensing image, the shading component S of the remote sensing image I at time phase 1 1 , the shading component S of the remote sensing image I at time phase 2 1 , the shading component S of the remote sensing image I at time phase 2 2 ) is obtained. 2 )

[0057] By setting corresponding threshold parameters and iteration end settings, through iterative solution, the final inversion results of the three parameters R, S 1 and S 2 are obtained;

[0058] The specific process is as follows:

[0059] The setting range of the local neighborhood scale (the selection range of local neighborhood similar points) is [3, 5, 7, 9]. The larger the value, the higher the smoothness. If there is no special setting, it is recommended to select 5 for the local neighborhood scale; the setting range of the neighborhood similarity control parameter σ is 0.1 to 10. The larger the value, the greater the contrast of the similarity distribution weight. When there are no special requirements, it is recommended to set it to 1. The adjacent iteration parameter error threshold is 10 -5 ;

[0060] The iteration end conditions are as follows:

[0061]

[0062] where k' is the number of iterations.

[0063] Other steps and parameters are the same as any one of Embodiments 1 to 4.

[0064] Embodiment 6: The difference between this embodiment and any one of Embodiments 1 to 5 is that the local neighborhood scale is set to 5.

[0065] Other steps and parameters are the same as any one of Embodiments 1 to 5.

[0066] Embodiment 7: The difference between this embodiment and any one of Embodiments 1 to 6 is that the similarity control parameter σ is set to 1.

[0067] Other steps and parameters are the same as any one of Embodiments 1 to 6.

[0068] The following embodiments are used to verify the beneficial effects of the present invention:

[0069] Embodiment 1:

[0070] In this embodiment, a method for extracting the spectral reflectance image of ground objects based on multi-temporal intrinsic image decomposition is specifically as follows:

[0071] The data used in the experiment are remote sensing images of two time phases in the Dalian area of GF-2, with the image size of 450×450. The number of image bands is 4, and the data has been pre-processed such as atmospheric and geometric correction. Figure 2a It is the RGB band image of the original data in time phase 1. Figure 2b It is the RGB band image of the original data in time phase 2. Figure 3a It is the RGB band image of the result of extracting the reflectance of a single image in time phase 1. Figure 3b It is the RGB band image of the result of extracting the reflectance of a single image in time phase 2. Figure 3c It is the RGB band image of the result of extracting the reflectance of the multi-temporal remote sensing image according to the present invention.

[0072] It can be seen from the image comparison that, compared with the result of extracting the reflectance of a single image, the reflectance image extracted by multi-temporal intrinsic decomposition is closer to the color of the original image, without color difference, and the spectral reflectance difference of the same type of ground objects is very small. The experimental results verify the effectiveness of the method for extracting the spectral reflectance image of ground objects based on multi-temporal intrinsic image decomposition proposed by the present invention.

[0073] The present invention can also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and deformations according to the present invention, but these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.

Claims

1. Method for extracting ground object spectral reflectance image based on multi-temporal intrinsic image decomposition, characterized in that: The specific process of the method is as follows: Step 1: Establish an expression model for the intrinsic reflectance information of multi-temporal remote sensing images, where the multi-temporal remote sensing images consist of the same reflectance component and independent shadow components; Step 2: Based on the expression model for the intrinsic reflectance information of multi-temporal remote sensing images, construct a decomposition constraint model under local spatio-temporal energy constraint. By minimizing the decomposition constraint model under local spatio-temporal energy constraint, obtain the optimal solution of the decomposition constraint model, and invert the common reflectance R and the shadow components of each image; Step 3: Convert the optimal solution of the decomposition constraint model into an iterative optimization solution model for the common reflectance and two shadow components; Step 4: Given multi-temporal remote sensing images and initial parameters, based on the iterative optimization solution model in Step 3, obtain the final intrinsic reflectance; The initial parameters are the local neighborhood scale range, neighborhood similarity regulation parameter, adjacent iteration parameter error threshold, and the initial values of the reflectance R and the shadow components of each image; The initial values of the shadow components of each image are set as all-zero matrices; In Step 1, the process of establishing the expression model for the intrinsic reflectance information of multi-temporal remote sensing images is as follows: Collect multi-temporal multi / hyperspectral remote sensing images at the same location at different times. The expression model for the intrinsic reflectance information of multi-temporal remote sensing images is: I 1 = S 1 R I 2 = S 2 R Among them, I 1 is the remote sensing image of phase 1, I 2 is the remote sensing image of phase 2, R is the common reflectance of the multi-temporal remote sensing images, I 1 、I 2 and R are all of size m rows by n columns in the B band, S 1 is the shadow component of the remote sensing image I 1 in phase 1, S 2 is the shadow component of the remote sensing image I 2 in phase 2, S 1 and S 2 are both of size m rows by n columns; B is the number of spectral bands of the multi- / hyperspectral remote sensing image; The decomposition constraint model under local spatio-temporal energy constraint in Step 2 is expressed as: Among them, E represents the local spatio-temporal energy; P are all the pixels in the image, the number of which is m×n, and N(i) represents the pixels in the local neighborhood around pixel point i; R i represents the reflectance of the i-th point of the remote sensing image; where w ij represents the neighborhood similarity between pixel point i and its neighboring point j, and the neighborhood similarity calculation formula is as follows: where, I ib represents the gray value of the i-th point in the b-th band of the image I, and I jb represents the gray value of the j-th point in the b-th band of the image I, and σ is the neighborhood similarity regulation parameter; By minimizing the above decomposition constraint model under local spatio-temporal energy constraint, obtain the optimal solution of the decomposition constraint model, and invert the common reflectance R and the shadow components of each image; That is: Among them, R b represents the gray value of the b-th band of the reflectance R; In Step 3, the process of converting the optimal solution of the decomposition constraint model into an iterative optimization solution model for the common reflectance and two shadow components is as follows: The iterative optimization solution model is as follows: Among them, represents the remote sensing image I at phase 1 1 the reciprocal of the shadow component of pixel point i; I 1ib represents the remote sensing image I at phase 1 1 the gray value of the i-th point in the b-th band; R 1ib represents the reflectance of the i-th point in the b-th band of the remote sensing image at phase 1; represents the remote sensing image I at phase 2 2 the reciprocal of the shadow component of pixel point i; I 2ib represents the remote sensing image I at phase 2 2 the gray value of the i-th point in the b-th band; R 2ib represents the reflectance of the i-th point in the b-th band of the remote sensing image at phase 2; R ib represents the reflectance of the i-th point in the b-th band of the multi-temporal remote sensing image; R jb represents the reflectance of the j-th point in the b-th band of the multi-temporal remote sensing image; w ji represents the similarity between pixel point j and its neighboring point i; N(j) represents the pixels within the local neighborhood around pixel point j; w jk represents the similarity between pixel point j and its neighboring point k; R kb represents the reflectance of the k-th point in the b-th band of the multi-temporal remote sensing image; I 1b represents the gray value of the 1st point in the b-th band of image I, I 2b represents the gray value of the 2nd point in the b-th band of image I.

2. The method for extracting ground object spectral reflectance image based on multi-temporal intrinsic image decomposition according to claim 1, characterized in that: In Step 4, given multi-temporal remote sensing images and initial parameters, based on the iterative optimization solution model in Step 3, obtain the final intrinsic reflectance. The specific process is as follows: The local neighborhood scale setting range is [3, 5, 7, 9], the neighborhood similarity regulation parameter σ setting range is 0.1 to 10, and the adjacent iteration parameter error threshold is 10 -5 ; The iteration end condition is as follows: where k′ is the number of iterations.

3. The method for extracting ground object spectral reflectance image based on multi-temporal intrinsic image decomposition according to claim 2, characterized in that: The local neighborhood scale is set to 5.

4. The method for extracting ground object spectral reflectance image based on multi-temporal intrinsic image decomposition according to claim 3, characterized in that: The neighborhood similarity regulation parameter σ is set to 1.

Images