Geographically weighted non-parametric regression-based remote sensing data normalization method and system

Abstract

The application discloses a remote sensing data normalization method and system based on geographical weighted nonparametric regression, and the method comprises the following steps: building a geographical weighted nonparametric regression estimation model; obtaining a target image and a reference remote sensing image set, and forming a remote sensing image pair based on the target image and the corresponding reference remote sensing image; preprocessing the target image in the remote sensing image pair to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing image; judging the coefficient of variation in each local window of the resampled target image and the reference remote sensing image, and screening out qualified homogeneous pixels based on the coefficient of variation and a preset threshold to form a homogeneous pixel set; and performing pixel-by-pixel calculation on the image to be normalized based on the geographical weighted nonparametric regression estimation model and the homogeneous pixel set, so as to obtain a normalization result. The remote sensing data with high width, high time resolution, consistent spectrum and consistent radiation can be used as a radiation reference benchmark to realize the normalization of any target image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing data processing, and particularly relates to a remote sensing data normalization method and system based on geographic weighted non-parametric regression. BACKGROUND

[0002] The rapid increase in the number of earth observation satellites is a necessary factor for the construction of large-scale remote sensing data sets at global and regional scales, providing reliable support for high-frequency and high-precision ground surface monitoring. Although the surge in data volume and the diversity of sensors have improved the capabilities of remote sensing monitoring, traditional radiation correction techniques still face significant challenges in the radiation consistency or normalization processing of multi-temporal and multi-source remote sensing images, limiting the collaborative use of massive multi-source remote sensing data. The key to radiation normalization is to construct the radiation conversion relationship between multi-source remote sensing data, which is complex, nonlinear, spatially and temporally heterogeneous, and highly variable due to differences in sensor characteristics, surface conditions, imaging processes, and atmospheric conditions. The commonly used methods are traditional regression estimation models such as least squares linear model and Theil regression, and machine learning methods such as support vector machine regression and random forest. However, these methods are not sufficient to depict the complex nonlinear spatio-temporal heterogeneity of the radiation conversion relationship caused by multiple factors, making it difficult to address the radiation inconsistency problem under the influence of complex factors such as surface differences, atmospheric interference, and sensor differences, and limiting the radiation consistency effect of multi-source optical remote sensing data.

[0003] To effectively address the problem of radiation inconsistency and the complexity caused by changes in land cover types and features over time, accurately depicting the radiation difference correlation under the influence of multiple factors is a key technical difficulty in achieving high-precision radiation normalization of multi-source remote sensing images. SUMMARY

[0004] The present application provides a remote sensing data normalization method, system and device based on geographic weighted non-parametric regression, which overcomes the shortcomings of the prior art.

[0005] To solve the above technical problems, the present application solves the problem by the following technical scheme:

[0006] A remote sensing data normalization method based on geographic weighted non-parametric regression, comprising the following steps:

[0007] A geographic weighted non-parametric regression estimation model is built, and the adaptive bandwidth parameter selection weight in the geographic weighted non-parametric regression estimation model is determined by combining the geographic location of the current pixel point and the geographic spatial bandwidth and value domain spatial bandwidth, wherein the adaptive bandwidth parameter selection weight is determined by the value domain weight and the geographic weight.

[0008] obtain a target image and a set of reference remote sensing images, and form a remote sensing image pair based on the target image and corresponding reference remote sensing images;

[0009] perform preprocessing on the target image in the remote sensing image pair to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing images, wherein the preprocessing at least includes resampling and geometric registration;

[0010] determine the coefficient of variation in each local window in the resampled target image and the reference remote sensing images, and filter out homogeneous pixels that meet the conditions based on the coefficient of variation and a preset threshold to form a set of homogeneous pixels;

[0011] perform pixel-by-pixel calculation on the target image based on the geographically weighted non-parametric regression estimation model and the set of homogeneous pixels to obtain a normalization result, wherein the corresponding value range bandwidth and geographical space bandwidth are determined by the difference in gray value between the pixel to be normalized in the target image and the homogeneous pixels and the geographical spatial position relationship, and then the adaptive bandwidth parameter selection weight is determined.

[0012] As an implementable manner, the geographically weighted non-parametric regression estimation model is constructed by the following steps:

[0013] construct a linear regression estimation model in combination with the geographical position of the current pixel point, convert the linear regression estimation model by non-parametric regression in combination with the regression coefficient, geographical space bandwidth and value range bandwidth through Taylor expression, and obtain a non-parametric linear regression estimation model;

[0014] perform weighted least squares processing on the non-parametric linear regression estimation model to obtain a geographically weighted non-parametric regression estimation model, and obtain a predicted value at a certain geographical position through the geographically weighted non-parametric regression estimation model.

[0015] As an implementable manner, the linear regression estimation model is represented as follows:

[0016]

[0017] The non-parametric linear regression estimation model is represented as follows:

[0018]

[0019] The geographically weighted non-parametric regression estimation model is represented as follows:

[0020]

[0021] wherein W j p (x i ,θ i) represents an adaptive bandwidth parameter selection weight, b represents a value domain space bandwidth, h represents a geographic space bandwidth, j represents an actual value, represents an estimated value or a predicted value of a regression coefficient β at a geographic position θ i , j (x i ) represents a weight of the jth homogeneous pixel in the value domain space relative to x i , j (θ i ) represents a weight of the jth homogeneous pixel in the geographic space relative to the geographic position θ i , i represents the i-th geographic position, i.e., a spatial coordinate, Y j represents a reference remote sensing image, represents an adaptive bandwidth parameter selection weight.

[0022] As an implementable manner, the adaptive bandwidth parameter selection weight is determined by a value domain space weight and a geographic space weight, including the following steps:

[0023] A value domain space bandwidth is determined by a pixel to be normalized in the target image and a hyperparameter in the target image, and a value domain space weight is obtained by the value domain space bandwidth and a locally characteristic adaptive double square function;

[0024] A geographic space bandwidth is determined based on a spatial autocorrelation coefficient threshold, and a geographic space weight is obtained by an adaptive double square function;

[0025] The value domain weight is represented as follows:

[0026]

[0027] wherein, b i represents a value domain space bandwidth, which is set according to a pixel to be normalized x i in the target image and a hyperparameter d similar , and changes with the size of the current pixel gray value;

[0028] The geographic space weight is represented as follows:

[0029]

[0030] wherein, d j (θ i ) represents a distance between the pixel to be normalized in the target image and a homogeneous pixel with a geographic position θ i , h represents a geographic space bandwidth, which is determined by a threshold R' of a spatial autocorrelation coefficient R;

[0031]

[0032] where N is the number of pixels, ω ij (d) is the weight matrix of distance d, x i and x j are the reference values of pixel i and pixel j, respectively; is the average value of all pixels, and the distance d when R takes the threshold value R' is defined as the geographical spatial bandwidth h, which determines the neighborhood range on the geographical space, (u i ,v i ) represents the coordinates of the i-th homogeneous pixel in the target image in the homogeneous pixel set.

[0033] As an implementable manner, the target image in the remote sensing image pair is preprocessed to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing image, including the following steps:

[0034] The resampled target image has the same spatial scale as the reference remote sensing image, and the target image has the same geometric reference information as all the reference remote sensing images through geometric registration.

[0035] As an implementable manner, the coefficient of variation in each local window of the resampled target image and the reference remote sensing image is judged, and the homogeneous pixels meeting the conditions are screened out based on the coefficient of variation and a preset threshold to form a homogeneous pixel set, including the following steps:

[0036] The coefficient of variation in each local window of the resampled target image and the reference remote sensing image is calculated through a coefficient of variation calculation model, respectively;

[0037] An adaptive threshold of a preset adjustable hyperparameter is set, and the adaptive threshold is represented as: σ homo and σ homo = λ · std + mean, the pixels with a coefficient of variation less than the adaptive threshold are screened out to form a homogeneous pixel set, and the coefficient of variation is represented as CV, the pixels with CV < σ homo are screened out to form the homogeneous pixel set Ω homo ;

[0038] The homogeneous pixel set is represented as follows:

[0039] Ω homo = {(SI resampling1 ,RI1),(SI resampling2 ,RI2),...,(SI resamplingn ,RI n}

[0040] wherein, Ω homo denotes the homogeneous pixel set, SI resampling denotes the pixels in the down-sampled target image, SI denotes the target image, RI denotes the reference remote sensing image, n denotes the number of homogeneous pixels, σ homo denotes the adaptive threshold, mean, std respectively denote the mean and standard deviation of the gray values of all pixels in the current local window.

[0041] As an implementable manner, the coefficient of variation calculation model is represented as follows:

[0042]

[0043] wherein, CV denotes the coefficient of variation, mean, std respectively denote the mean and standard deviation of the gray values of all pixels in the current local window.

[0044] As an implementable manner, the homogeneous pixel is calculated based on the geographic weighted non-parametric regression estimation model to obtain the normalization result, specifically:

[0045] The homogeneous pixel is calculated pixel by pixel by the geographic weighted non-parametric regression estimation model to obtain the normalization result, and the normalization result of the homogeneous pixel set is represented as follows:

[0046]

[0047] wherein, SI i (x i , θ i ) denotes the normalization result of the i-th pixel in the homogeneous pixel set at the geographic position θ i , denotes the adaptive bandwidth parameter selection weight, θ i denotes the geographic position, i.e. spatial coordinates, at the i-th pixel, x i denotes the i-th pixel in the homogeneous pixel set, SI i denotes the homogeneous pixel set, RI j denotes the reference remote sensing image, Ω homo denotes the homogeneous pixel set.

[0048] As an implementable manner, the resampling is realized by area-weighted data resampling, specifically: the actual ground area in each local window in the converted remote sensing observation image is calculated;

[0049] Based on the actual ground area in each local window divided by the sum of all pixel areas, the weight in each local window is obtained, and then the weighted average value is obtained by the weight in each local window to realize image resampling.

[0050] A remote sensing data normalization system based on a geographically weighted non-parametric regression comprises a model building module, a data processing module and a result calculation module.

[0051] The model building module builds a geographically weighted non-parametric regression estimation model, determines adaptive bandwidth parameter selection weights in the geographically weighted non-parametric regression estimation model by combining the geographical position of a current pixel point and through a geographical space bandwidth and a value range space bandwidth, wherein the adaptive bandwidth parameter selection weights are determined by a value range space weight and a geographical space weight.

[0052] The data processing module is configured to obtain a target image and a reference remote sensing image set, and form a remote sensing image pair based on the target image and corresponding reference remote sensing images; pre-process the target image in the remote sensing image pair to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing images, wherein the pre-processing at least includes resampling and geometric registration; and determine a coefficient of variation in each local window in the resampled target image and the reference remote sensing images, and filter out homogeneous pixels meeting a condition based on the coefficient of variation and a preset threshold to form a homogeneous pixel set.

[0053] The result calculation module performs pixel-by-pixel calculation on the target image based on the geographically weighted non-parametric regression estimation model and the homogeneous pixel set to obtain a normalization result, wherein the corresponding value range space bandwidth and geographical space bandwidth are determined by the difference in gray scale values between a pixel to be normalized in the target image and the homogeneous pixels and the geographical position relationship, and then the adaptive bandwidth parameter selection weights are determined.

[0054] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the following method.

[0055] A geographically weighted non-parametric regression estimation model is built, adaptive bandwidth parameter selection weights in the geographically weighted non-parametric regression estimation model are determined by combining the geographical position of a current pixel point and through a geographical space bandwidth and a value range space bandwidth, wherein the adaptive bandwidth parameter selection weights are determined by a value range space weight and a geographical space weight.

[0056] A target image and a reference remote sensing image set are obtained, and a remote sensing image pair is formed based on the target image and corresponding reference remote sensing images.

[0057] The target image in the remote sensing image pair is pre-processed to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing images, wherein the pre-processing at least includes resampling and geometric registration.

[0058] Determine the coefficient of variation in each local window of the target image and the reference remote sensing image after resampling, and filter out the homogeneous pixels meeting the conditions based on the coefficient of variation and a preset threshold to form a homogeneous pixel set;

[0059] Based on the geographically weighted non-parametric regression estimation model and the homogeneous pixel set, the target image is calculated pixel by pixel to obtain a normalization result, wherein the corresponding value range space bandwidth and geographical space bandwidth are determined through the difference in gray value between the target image and the homogeneous pixel in the target image to be normalized and the geographical spatial position relationship, and then the adaptive bandwidth parameter selection weight is determined.

[0060] A remote sensing data normalization device based on a geographically weighted non-parametric regression includes a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the method as follows when executing the computer program:

[0061] A geographically weighted non-parametric regression estimation model is built, the geographical position of the current pixel point is combined, the adaptive bandwidth parameter selection weight in the geographically weighted non-parametric regression estimation model is determined through the geographical space bandwidth and the value range space bandwidth, and the adaptive bandwidth parameter selection weight is determined through the value range space weight and the geographical space weight;

[0062] A target image and a reference remote sensing image set are obtained, and a remote sensing image pair is formed based on the target image and the corresponding reference remote sensing image;

[0063] The target image in the remote sensing image pair is preprocessed to obtain a resampled target image with the same spatial resolution as the reference remote sensing image, and the preprocessing at least includes resampling and geometric registration;

[0064] Determine the coefficient of variation in each local window of the target image and the reference remote sensing image after resampling, and filter out the homogeneous pixels meeting the conditions based on the coefficient of variation and a preset threshold to form a homogeneous pixel set;

[0065] Based on the geographically weighted non-parametric regression estimation model and the homogeneous pixel set, the target image is calculated pixel by pixel to obtain a normalization result, wherein the corresponding value range space bandwidth and geographical space bandwidth are determined through the difference in gray value between the target image and the homogeneous pixel in the target image to be normalized and the geographical spatial position relationship, and then the adaptive bandwidth parameter selection weight is determined.

[0066] The present application has the following technical effects:

[0067] The application constructs a geographical weighted non-parametric regression estimation model, and data is used to drive the establishment of the non-parametric model, that is, the model has no fixed form and is good at processing complex nonlinear relationships, and a geographical space weighting term is added to the geographical weighted non-parametric regression estimation model to realize modeling of the geographical weighted non-parametric regression estimation model, and in the modeling process, the spatial and spectral dimension radiation correlation is mined and utilized through construction of a spatial domain and spectral domain bandwidth weight function, and the geographical weighted non-parametric regression estimation model can solve the modeling problem of nonlinearity and spatial non-stationarity existing in the field of earth science.

[0068] The application uses remote sensing data with large width, high time resolution, consistent spectrum and consistent radiation as a radiation reference benchmark, and can realize normalization of any target image.

[0069] The application determines the bandwidth parameter in the geographical weighted non-parametric regression through an adaptive bandwidth parameter, solves the problem of insufficient response of a fixed parameter model to a local model, can determine different parameters under different conditions according to data characteristics under different conditions, and thus provides a spatially different radiation conversion model, so as to better adapt to changes in local characteristics, improve generalization performance and radiation normalization precision. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0071] Figure 1 is a schematic diagram of the overall process of the method of the application;

[0072] Figure 2 is a schematic diagram of the overall structure of the system of the application;

[0073] Figure 3 is a schematic diagram of the target image, reference image and normalization result comparison of the application;

[0074] Figure 4 is a schematic diagram of the visual performance obtained by various methods of the application. DETAILED DESCRIPTION

[0075] The application will be further described in detail below in combination with embodiments, and the following embodiments are used to explain the application, but the application is not limited to the following embodiments.

[0076] Embodiment 1:

[0077] A remote sensing data normalization method based on geographical weighted non-parametric regression, as shown inFigure 1 as shown, comprising the following steps:

[0078] S100, a geographic weighted non-parametric regression estimation model is built, the geographic position of the current pixel point is combined, the adaptive bandwidth parameter selection weight in the geographic weighted non-parametric regression estimation model is determined through the geographic space bandwidth and the value range space bandwidth, wherein the adaptive bandwidth parameter selection weight is determined through the value range space weight and the geographic space weight;

[0079] S200, a target image and a reference remote sensing image set are obtained, and a remote sensing image pair is formed based on the target image and the corresponding reference remote sensing image;

[0080] S300, the target image in the remote sensing image pair is preprocessed to obtain a resampled target image with the same spatial resolution as the reference remote sensing image, wherein the preprocessing at least includes resampling and geometric registration;

[0081] S400, the coefficient of variation in each local window of the resampled target image and the reference remote sensing image is determined, and the homogenous pixels meeting the conditions are screened out based on the coefficient of variation and the preset threshold to form a homogenous pixel set;

[0082] S500, based on the geographic weighted non-parametric regression estimation model and the homogenous pixel set, the target image is calculated pixel by pixel to obtain the normalization result, wherein the corresponding value range space bandwidth and geographic space bandwidth are determined through the difference between the gray value of the target image in the target image and the homogenous pixel and the geographic space position relationship, and then the adaptive bandwidth parameter selection weight is determined.

[0083] Non-parametric regression model is a statistical method that does not assume a fixed model form when modeling and analyzing data. Unlike traditional parametric regression models, non-parametric regression does not require pre-defined model structure, making non-parametric regression model very flexible in handling complex data relationships, especially when the exact relationship between data is not clear or when the relationship is difficult to express in simple mathematical form. Non-parametric regression method can capture the nonlinear patterns and structures in the data without specifying the model form in advance, so it can adapt to the structure of the data regardless of its complexity. It is very suitable for various types of data and relationships, including continuous and discrete variables; in addition, the results of non-parametric regression model are usually easier to interpret because they can be directly derived from the data without complex mathematical assumptions.

[0084] Based on this, the non-parametric model is established by data driving, that is, the model has no fixed form, is good at processing complex nonlinear relationship, and considers the first law of geography, that is, similar objects in space are similar, constructs a geographical space weighting term, and realizes the modeling of the spatial non-stationary model. The model will also help to solve the modeling problem of nonlinearity and spatial non-stationarity which is widespread in the field of earth science.

[0085] Without assuming any specific form of the regression function, the target is to find the minimum "residual sum of squares" of the estimate of the regression function The initial non-parametric regression estimation model is as follows:

[0086]

[0087] Local parameterization is performed through a Taylor expression of p order, for example, when x approaches x i , a locally parameterized linear regression estimation model is obtained, and is expressed as follows:

[0088]

[0089] The regression coefficient β c is constructed as f (j) (x) / c!, the non-parametric regression problem is converted into a linear regression problem, a linear regression estimation model is generated, and is expressed as follows:

[0090]

[0091] The contribution of each sample to the predicted value is weighted, and the weighting is performed according to the proximity of and , that is, , wherein is an observation value bandwidth, is a user-defined kernel function, and the above formula is modified as:

[0092]

[0093] It can be understood that the geographical weighted non-parametric regression model (GWNPR) of the present application is realized by combining the scheme of geographical weighted regression (GWR), that is, the regression coefficient can change in different positions. The specific implementation includes the following steps:

[0094] This embodiment considers the spatial geographical position, defines the regression function as a spatial variable dependent on the geographical position, can weight all observation values according to the spatial proximity to the regression observation value according to the distance decay function. Assuming that a sample set is , for any point , the coordinates are its geographical position is expressed as , the geospatial weight , is the spatial bandwidth. Then, considering the non-parametric regression estimation model of the initial construction of the spatial geographical position, it is expressed as:

[0095] ;

[0096] Through the Taylor expression, combined with the regression coefficient, the geospatial bandwidth and the value range spatial bandwidth, the non-parametric regression conversion of the linear regression estimation model is carried out, and the non-parametric linear regression estimation model is obtained, which is expressed as: ;

[0097] The weighted least squares processing is carried out on the non-parametric linear regression estimation model, and the geographically weighted non-parametric regression estimation model is obtained. The prediction value of a certain geographical position is obtained through the geographically weighted non-parametric regression estimation model, which is expressed as: ;

[0098] wherein, is the adaptive bandwidth parameter selection weight, , is the value range spatial bandwidth, is the geospatial bandwidth, is the actual value, is the estimated value or the prediction value of the regression coefficient at the geographical position , is the weight of the th homogeneous pixel in the value range space relative to , is the weight of the th homogeneous pixel in the geospatial space relative to the geographical position , is the geographical position at the th, i.e. the spatial coordinate, is the reference remote sensing image, is the adaptive bandwidth parameter selection weight.

[0099] The geographically weighted non-parametric regression estimation model constructed by this method can simplify the complex relationship through modeling to find a reasonable correlation, and can reasonably use the spatial geographical position in the modeling process, and then determine the adaptive bandwidth parameter selection weight by constructing the geospatial bandwidth and the value range spatial bandwidth weight, and then realize the mining of the correlation of the spatial and spectral dimensions and the subsequent application. Further, the data with large bandwidth, high temporal resolution, consistent spectrum and consistent radiation are used as the radiation reference benchmark to normalize any target image.

[0100] In addition, in other embodiments, an adaptive bandwidth parameter determination method is also proposed, which is used to determine the adaptive bandwidth parameter selection weight in the geographically weighted non-parametric regression estimation model, and can solve the problem of insufficient response of the fixed parameter model to the local model. The specific scheme is as follows: the adaptive bandwidth parameter selection weight in the geographically weighted non-parametric regression estimation model is determined through the geographic space bandwidth and the value range space bandwidth, and the adaptive bandwidth parameter selection weight can be expressed as:

[0101] The value range space bandwidth corresponding to the to-be-normalized pixel in the target image is determined through the target image in the target image and the hyperparameter; the value range space weight is obtained through the value range space bandwidth and the locally adaptive double square function, and the value range space weight is expressed as:

[0102] The geographic space bandwidth is determined based on the spatial autocorrelation coefficient threshold value, and then the geographic space weight is obtained through the adaptive double square function, and the geographic space weight can be expressed as:

[0103]

[0104] That is, in the entire calculation process, the value range space weight in the value range space is calculated through the locally adaptive double square function, the value range space bandwidth b is determined by the data characteristics of the to-be-normalized pixel in the current target image and the hyperparameter d similar is determined according to the local data characteristics, b i represents the value range space bandwidth, and is determined according to the to-be-normalized pixel x i in the target image and the hyperparameter d similar is set and changes with the size of the current pixel gray value. The adaptive bandwidth setting can better fit the data, and the weight can be adjusted according to the local characteristics of the data, so as to more accurately describe the non-stationarity of the radiation transformation relationship, If the bandwidth setting is not appropriate, such as too high, it will be similar to the global OLS model, and it is difficult to describe the non-stationarity of the radiation transformation relationship; if it is too low, it will cause high variance because the model is only applicable to observations closer to the regression point; and the geographic space bandwidth in the geographic space is calculated through the spatial distance adaptive double square function, d j (θ i ) represents the distance between the to-be-normalized pixel in the target image and the geographic location θ i in the homogeneous pixel set, and in combination with the sample, it can be known that The geographic space bandwidth h is determined through the threshold value R' of the spatial autocorrelation coefficient R;

[0105]

[0106] where N is the number of observation points, ω ij (d) is the weight matrix of distance d, x i and x j are the observation values of observation points i and j, respectively; is the average value of all observation points, and the distance d when R takes the threshold value R' is defined as the geographic spatial bandwidth h, which determines the range of the "neighborhood" on the geographic space and affects which observation points are considered when calculating the weight.

[0107] The fixed bandwidth involved in these two weights is set to weight the data at different locations on the geographic space, which makes the observation data closer to the regression point have greater weight than the data farther away, and the observation values outside the specified neighborhood are assigned to zero and do not affect the local regression of the target point, better considering the influence of geographic location on the regression relationship. Through this method, the data characteristics under different temporal and spatial conditions can be determined, and the parameters under different conditions can be determined, better adapting to the changes in local characteristics and improving the generalization performance and radiometric normalization accuracy of the model.

[0108] After the construction of the geographic weighted non-parametric regression estimation model, the process of realizing radiometric normalization through the geographic weighted non-parametric regression estimation model is as follows:

[0109] The target image and the reference remote sensing image set are obtained, and a remote sensing image pair is formed based on the target image and the corresponding reference remote sensing image. The target image in the remote sensing image pair is preprocessed to obtain a resampled target image with the same spatial resolution as the reference remote sensing image. The coefficient of variation in each local window of the resampled target image and the reference remote sensing image is determined, and the homogeneous pixels that meet the conditions are selected based on the coefficient of variation and the preset threshold to form a homogeneous pixel set. Then, the target image is calculated pixel by pixel through the geographic weighted non-parametric regression estimation model combined with the homogeneous pixel set to obtain the normalization result.

[0110] The specific process is as follows:

[0111] After obtaining the target image and the reference remote sensing image set, a remote sensing image pair is formed as an example from the target image and the corresponding reference remote sensing image. The target image is resampled, and the resampled target image and the reference remote sensing image are geometrically registered, that is, the resampled target image and the reference remote sensing image are in a corresponding relationship. Specifically, the target image is denoted as SI, and the reference remote sensing image is denoted as RI. The reference remote sensing image is a coarse resolution image. The points of each image in the image pair are represented by the corresponding values and positions (i.e., spatial coordinates), for example, SI i = (x i , θ i) represents the i-th point from the target image, the geographical position is denoted as θ i . The pixel mean resampling is performed on the to-be-normalized image (i.e., the target image), and the resampled target image has the same spatial scale and geometric characteristics as the reference remote sensing image. The pixels in the down-sampled target image can be denoted as SI resampling .

[0112] The pixels in the reference remote sensing image and the pixels in the target image are also filtered to obtain the homogeneous pixels that meet the conditions, that is, by calculating the coefficient of variation in each local window in the resampled target image and the reference remote sensing image, and extracting the homogeneous pixels, the homogeneous pixels that meet the conditions are filtered according to the coefficient of variation and the preset threshold to form a homogeneous pixel set. The homogeneous pixel set will be used as a reference or medium for the to-be-normalized pixels in the target image in the subsequent process. The specific method is as follows:

[0113] According to the pixels of the resampled target image, the coefficient of variation calculation model CV is used to measure the coefficient of variation in each local window in the resampled target image. By using the adaptive threshold σ homo = λ · std + mean, the pixels with CV < σ homo are filtered to form the homogeneous pixels as the homogeneous pixel set Ω homo , where mean and std are the mean and standard deviation of the gray values of all pixels in a local window, and the homogeneous pixel set is represented as:

[0114] Ω homo = {(SI resampling1 , RI1), (SI resampling2 , RI2),..., (SI resamplingn , RI n}.

[0115] In processing the scale effect in radiometric normalization, the homogeneity of pixels needs to be measured to select appropriate pixels for fitting the radiometric transformation model. The coefficient of variation calculation model (CV) is one of the commonly used homogeneity measurement methods, which belongs to the standardized probability distribution or frequency distribution dispersion measurement method and can reflect the diversity of data. It can eliminate the influence of different mean values on the comparison of dispersion, and then reflect the statistical variation of pixels. Here, the coefficient of variation calculation model applied is represented as: {(std, mean) ∈ R 2 : mean ≠ 0}, mean and std are the mean and standard deviation of the gray values of all pixels in the current local window. Since the coefficient of variation calculation model considers the deviation of all pixels from the mean, it provides more comprehensive information on the dispersion of pixels for the homogeneous pixel filtering process.

[0116] After obtaining the homogeneous pixel set, the target image is calculated pixel-by-pixel using a geographically weighted nonparametric regression estimation model and the homogeneous pixel set to obtain the normalized result. The calculation process also includes determining the weights for the adaptive bandwidth parameter. This process can be understood as the process of determining the weights for the adaptive bandwidth parameter in the geographically weighted nonparametric regression estimation model. The corresponding value space bandwidth is determined by the difference in gray values ​​between the pixel to be normalized and the homogeneous pixels in the target image and the hyperparameter. The value space weights are obtained through an adaptive double-square function of the value space bandwidth and local features. The value space weights are expressed as:

[0117] The bandwidth b of the value range is determined by the data characteristics of the pixels to be normalized in the current target image and the hyperparameter d. similar The decision is to adapt and change based on the characteristics of local data. b i Represents the bandwidth of the value range, based on the x-th pixel to be normalized in the target image. i With hyperparameter d similar It is set and changes with the current pixel grayscale value.

[0118] The geospatial bandwidth is determined based on the geospatial positional relationship between the pixels to be normalized and homogeneous pixels in the target image. Then, the geospatial weights are obtained through an adaptive double-square function, and the geospatial weights are expressed as follows:

[0119]

[0120] That is, the spatial weights of the value range need to be determined by the geographic spatial relationship between the pixels to be normalized and homogeneous pixels in the target image. Throughout the process, the observation points in the model are the set of homogeneous pixels. The sample points in the middle, here The geospatial bandwidth h is determined by the threshold R' of the spatial autocorrelation coefficient R;

[0121]

[0122] The final weights for selecting the adaptive bandwidth parameter were determined as follows:

[0123] Finally, by combining a geographically weighted nonparametric regression estimation model with a homogeneous pixel set, the normalized result of pixel-by-pixel calculation of the target image is:

[0124] Furthermore, the technical effects of the present invention can be illustrated through specific images, see reference. Figure 3 As shown,Figure 3 (a) is a target image to be normalized, Figure 3 (b) is a normalized result, Figure 3 (c) is a reference remote sensing image, and the normalized result obtained through normalization processing Figure 3 (b) is compared with the target image to be normalized Figure 3 (a) has obvious gray scale changes, and the normalized result Figure 3 (b) is compared with the reference remote sensing image Figure 3 (c) can also intuitively see the similar distribution pattern and gray scale value dynamic range.

[0125] NDVI data of synchronous cloud-free Landsat ETM+ and ASTER images of a certain border area in a certain region are used for experiments by using various methods, and the visual performance results of the normalized results are as shown in Figure 4 (a) is a scatter plot of original data; (b) is a relative radiation normalization result based on least square regression LSQ; (c) is a relative radiation normalization result based on a land cover model CLM; (d) is a relative radiation normalization result based on support vector machine regression SVM; (e) is a relative radiation normalization result based on a local calibration land cover model LCLM; (f) is a relative radiation normalization result based on geographical weighted non-parametric regression GWNPR of the present application; through comparison, it can be seen that the scatter plot between the normalized result obtained by the present application and the evaluation reference data is closely aligned with the 1:1 line, which indicates that the two data are highly coincident. The scatter plot before normalization processing shows an obvious nonlinear mode, which conforms to the assumption of nonlinear relationship between multi-source data. Compared with other radiation conversion relationship modeling methods, such as least square regression LSQ, global land cover classification model CLM, support vector machine regression SVM, and local calibration land cover model LCLM, the scatter plot of the relative radiation normalization result of the modeling method GWNPR of the present application is closer to the true value, and the fitting curve is also more consistent with the distribution of data. It can be directly proved that the present application exhibits superiority when modeling between nonlinear and non-stationary data.

[0126] In one embodiment, the data acquisition method needs to be explained: atmospheric radiation transmission 6S model is used to simulate the atmospheric radiation transmission process, and remote sensing observation data under different atmospheric conditions are simulated. The 6S model is mainly used to simulate the transmission process of solar radiation in the earth-atmosphere system and calculate the radiation brightness at the satellite entrance pupil. It can consider the absorption of water vapor, CO2, O3 and O2, molecular and aerosol scattering, and the influence of non-uniform ground and bidirectional reflectance distribution under the assumption of cloud-free atmosphere. The input parameters of the 6S model include: satellite geometric conditions, including solar zenith angle, solar azimuth angle, observation zenith angle, observation azimuth angle, etc.; atmospheric model, which defines the basic composition of the atmosphere and the temperature and humidity profile, which can select the preset model or input the measured data; aerosol model, which defines the main aerosol parameters in the world, such as aerosol phase function, asymmetric factor and single scattering albedo, etc.; spectral characteristics of the sensor, which defines the spectral response function of the sensor channel; surface reflectance, which defines the reflectance model of the ground, including uniform ground and non-uniform ground.

[0127] Based on the spectral response function, the spectral matching conversion factor between the two sensors is calculated to realize the data conversion between different sensors; data resampling is the process of converting a data set from one resolution or format to another resolution or format. Basically, it is to calculate the actual ground area of each pixel in the converted remote sensing observation image, based on the actual ground area of each pixel divided by the sum of the areas of all pixels to obtain the weight of each pixel, and then the weighted average value is obtained through each pixel weight to realize image resampling.

[0128] Finally, the normalized result, i.e. the normalized image, is obtained, which will be used subsequently. The image processed by normalization will make the results of affine reconstruction model or other models more accurate in subsequent applications.

[0129] Embodiment 2:

[0130] A remote sensing data normalization system based on geographic weighted non-parametric regression, as shown in Figure 2 Fig. 1, comprises a model building module 100, a data processing module 200 and a result calculation module 300.

[0131] The model building module 100 builds a geographic weighted non-parametric regression estimation model, combines the geographic location of the current pixel point, and determines the adaptive bandwidth parameter selection weight in the geographic weighted non-parametric regression estimation model through the geographic space bandwidth and the value range space bandwidth, wherein the adaptive bandwidth parameter selection weight is determined by the value range weight and the geographic weight.

[0132] The data processing module 200 is configured to acquire a target image and a reference remote sensing image set, and form a remote sensing image pair based on the target image and a corresponding reference remote sensing image; perform preprocessing on the target image in the remote sensing image pair to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing image, wherein the preprocessing at least includes resampling and geometric registration; and determine a coefficient of variation in each local window in the resampled target image and the reference remote sensing image, and filter out homogeneous pixels meeting a condition based on the coefficient of variation and a preset threshold to form a homogeneous pixel set.

[0133] The result calculation module 300 is configured to perform pixel-by-pixel calculation on the homogeneous pixel set based on the geographically weighted non-parametric regression estimation model to obtain a normalized result.

[0134] Each of the embodiments in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be referred to each other.

[0135] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, device, or computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media containing computer usable program code (including but not limited to disk memory, CD-ROM, optical memory, etc.).

[0136] The present application is described with reference to flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The means for implementing the functions specified in the flowcharts and / or block diagrams.

[0137] These computer program instructions can also be stored in a computer readable memory capable of guiding the computer or other programmable data processing terminal device to work in a specific way, so that the instructions stored in the computer readable memory produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The means for implementing the functions specified in the flowcharts and / or block diagrams.

[0138] These computer program instructions can also be loaded into computer or other programmable data processing devices to cause a series of operational steps to be performed on the computer or other programmable devices to produce computer-implemented processes such that the instructions which execute on the computer or other programmable devices provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1 The flowchart blocks or blocks in the flowcharts represent steps in a process or processes and / or functions specified in one or more blocks in a block or blocks.

[0139] It should be noted that:

[0140] The phrase "one embodiment" or "an embodiment" as used throughout this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. Therefore, the appearances of the phrase "one embodiment" or "an embodiment" throughout the specification are not necessarily all referring to the same embodiment.

[0141] Furthermore, it is to be understood that the specific embodiments described herein can be different, have different compositions, features, structures or otherwise be practiced under inherently different conditions and / or can have different results without departing from the spirit and scope of the claims. Therefore, other modifications orequivalents are intended to be covered.

Claims

1. A method for remote sensing data normalization based on geographically weighted non-parametric regression, characterized in that, The method comprises the following steps: The method comprises the following steps: A geographically weighted non-parametric regression estimation model is established, and the adaptive bandwidth parameter selection weight in the geographically weighted non-parametric regression estimation model is determined by combining the geographical position of the current pixel point and the geographical space bandwidth and the value range space bandwidth. A target image and a reference remote sensing image set are obtained, and a remote sensing image pair is formed based on the target image and the corresponding reference remote sensing image. The target image in the remote sensing image pair is preprocessed to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing image, wherein the preprocessing at least includes resampling and geometric registration. The coefficient of variation in each local window of the resampled target image and the reference remote sensing image is determined, and the homogenous pixels meeting the conditions are screened out based on the coefficient of variation and a preset threshold to form a homogenous pixel set. The target image is calculated pixel by pixel based on the geographically weighted non-parametric regression estimation model and the homogenous pixel set to obtain a normalization result, wherein the corresponding value range space bandwidth and geographical space bandwidth are determined by the difference between the gray values of the target image in the target image and the homogenous pixels and the geographical position relationship, and then the adaptive bandwidth parameter selection weight is determined. The geographically weighted non-parametric regression estimation model is constructed by the following steps: A linear regression estimation model is constructed by combining the geographical position of the current pixel point, and the linear regression estimation model is converted by non-parametric regression by combining the regression coefficient, the geographical space bandwidth and the value range space bandwidth through the Taylor expression to obtain a non-parametric linear regression estimation model. The non-parametric linear regression estimation model is processed by weighted least squares to obtain a geographically weighted non-parametric regression estimation model, and a predicted value at a certain geographical position is obtained through the geographically weighted non-parametric regression estimation model. The linear regression estimation model is represented as follows: The non-parametric linear regression estimation model is represented as follows: wherein denotes an adaptive bandwidth parameter selection weight, , denotes a value domain spatial bandwidth, denotes a geospatial bandwidth, denotes an actual value, denotes an estimated or predicted value of the regression coefficient at the geographical location , denotes a weight of the th homogenous pixel in the value domain space relative to , denotes a weight of the th homogenous pixel in the geospatial space relative to the geographical location , denotes the geographical location, i.e. spatial coordinates, at the th location, denotes a reference remote sensing image, denotes an adaptive bandwidth parameter selection weight.

2. The method of claim 1, wherein, The geographically weighted non-parametric regression estimation model is represented as follows: The adaptive bandwidth parameter selection weight is determined by the value range space weight and the geographical space weight, comprising the following steps: The value range space bandwidth is determined by the target image in the target image and the hyperparameter, and the value range space weight is obtained by the value range space bandwidth and the locally characteristic adaptive double square function. The geographical space bandwidth is determined based on the spatial autocorrelation coefficient threshold, and then the geographical space weight is obtained by the adaptive double square function. wherein, , denotes the value range space bandwidth, according to the target image in the target image in the pixel to be normalized and hyperparameters are set, which vary with the size of the current pixel grayscale value; The value range weight is represented as follows: wherein, represents the distance between the pixel to be normalized in the target image and the geographical position of the homogenous pixel in the homogenous pixel set, , represents the geographical spatial bandwidth determined by the threshold value of the spatial autocorrelation coefficient .​​ wherein, is the number of pixel points, is the distance is the weight matrix of the pixel points, and are the reference values of the pixel and the pixel respectively; is the average value of all pixel points, when the distance is taken as a threshold is defined as the geographical spatial bandwidth , the geographical spatial bandwidth determines the neighborhood range on the geographical space, represents the coordinate of the th homogenous pixel in the target image.

3. The method of claim 1, wherein, The geographical space weight is represented as follows: The target image in the remote sensing image pair is preprocessed to obtain a resampled target image with a spatial resolution consistent with that of the reference remote sensing image, comprising the following steps:

4. The method of claim 1, wherein, The target image is resampled by pixel mean, and the resampled target image has the same spatial scale as the reference remote sensing image, and the target image has the same geometric reference information as all the reference remote sensing images through geometric registration. The coefficient of variation in each local window of the resampled target image and the reference remote sensing image is determined, and the homogenous pixels meeting the conditions are screened out based on the coefficient of variation and a preset threshold to form a homogenous pixel set. The coefficient of variation is calculated by a coefficient of variation calculation model for each local window in the target image and the reference remote sensing image after resampling; An adaptive threshold of preset adjustable hyperparameters is represented as: and The pixel with a coefficient of variation less than the adaptive threshold is screened to form a homogeneous pixel set, and the coefficient of variation is represented as The pixel group of forms the homogeneous pixel set ; The homogeneous pixel set is represented as follows: wherein, denotes the homogeneous pixel set, denotes the pixels in the down-sampled target image, denotes the target image, denotes the reference remote sensing image, denotes the number of homogeneous pixels, denotes the adaptive threshold, denotes the mean and standard deviation of the gray values of all pixels in the current local window, respectively.

5. The method for remote sensing data normalization based on geographically weighted non-parametric regression according to claim 1 or 4, characterized in that, The coefficient of variation calculation model is represented as follows: wherein, denotes the coefficient of variation, respectively denote the mean and standard deviation of the gray scale of all pixels within the current local window.

6. The method of claim 1, wherein, The homogeneous pixel is calculated based on the geographic weighted non-parametric regression estimation model to obtain a normalization result, specifically: The homogeneous pixel is calculated by a geographic weighted non-parametric regression estimation model to obtain a normalization result, and the normalization result of the homogeneous pixel set is represented as follows: wherein, denotes a normalized result of a geographical position of the i-th pixel of the homogeneous pixel set, denotes a normalized result of a geographical position of the i-th pixel of the homogeneous pixel set, denotes a normalized result of a geographical position of the i-th pixel of the homogeneous pixel set, denotes an adaptive bandwidth parameter selection weight, denotes a geographical position, i.e. a spatial coordinate, at the i-th pixel, denotes a geographical position, i.e. a spatial coordinate, at the i-th pixel, denotes the i-th pixel of the homogeneous pixel set, denotes the i-th pixel of the homogeneous pixel set, denotes a reference remote sensing image, denotes a homogeneous pixel set.

7. The method of claim 1, wherein, The resampling is implemented by area weighting data resampling, specifically: the actual ground area in each local window in the converted remote sensing observation image is calculated; Based on the actual ground area in each local window divided by the sum of all pixel areas, the weight in each local window is obtained, and then the weighted average value is obtained by the weight in each local window to realize image resampling.

8. A system for remote sensing data normalization based on geographically weighted non-parametric regression, characterized in that, The model building module, the data processing module and the result calculation module are included. The model building module builds a geographic weighted non-parametric regression estimation model, determines the adaptive bandwidth parameter selection weight in the geographic weighted non-parametric regression estimation model by the geographic position of the current pixel point, the geographic space bandwidth and the value range space bandwidth, wherein the adaptive bandwidth parameter selection weight is determined by the value range space weight and the geographic space weight; The data processing module is used for obtaining a target image and a reference remote sensing image set, and forming a remote sensing image pair based on the target image and the corresponding reference remote sensing image; pre-processing the target image in the remote sensing image pair to obtain a resampled target image with the same spatial resolution as the reference remote sensing image, wherein the pre-processing at least includes resampling and geometric registration; judging the coefficient of variation in each local window in the resampled target image and the reference remote sensing image, and screening out the homogeneous pixels meeting the conditions based on the coefficient of variation and the preset threshold to form a homogeneous pixel set; The result calculation module calculates the target image pixel by pixel based on the geographic weighted non-parametric regression estimation model and the homogeneous pixel set to obtain a normalization result, wherein the corresponding value range space bandwidth and geographic space bandwidth are determined by the difference between the gray value of the target pixel in the target image and the homogeneous pixel and the geographic spatial position relationship, and then the adaptive bandwidth parameter selection weight is determined; The geographic weighted non-parametric regression estimation model is built by the following steps: A linear regression estimation model is built by combining the geographic position of the current pixel point, and a non-parametric linear regression estimation model is obtained by converting the linear regression estimation model through Taylor expression, combining the regression coefficient, the geographic space bandwidth and the value range space bandwidth; The non-parametric linear regression estimation model is subjected to weighted least square processing to obtain a geographic weighted non-parametric regression estimation model, and a predicted value at a certain geographic position is obtained by the geographic weighted non-parametric regression estimation model; The linear regression estimation model is represented as follows: The non-parametric linear regression estimation model is represented as follows: The geographic weighted non-parametric regression estimation model is represented as follows: wherein denotes an adaptive bandwidth parameter selection weight, , denotes a value domain spatial bandwidth, denotes a geospatial bandwidth, denotes an actual value, denotes an estimated or predicted value of the regression coefficient at the geographical location , denotes a weight of the th homogenous pixel in the value domain space relative to , denotes a weight of the th homogenous pixel in the geospatial space relative to the geographical location , denotes the geographical location, i.e. spatial coordinates, at the th location, denotes a reference remote sensing image, denotes an adaptive bandwidth parameter selection weight.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program, which when executed by a processor, implements the method of any one of claims 1 to 7.

10. A device for remote sensing data normalization based on geographically weighted non-parametric regression, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the method of any one of claims 1 to 7.

Images