A method, device and medium for improving remote sensing estimation of object abundance within a pixel

Abstract

The application discloses a method, device and medium for improving remote sensing estimation of ground object abundance in a pixel. The application belongs to the technical field of remote sensing imaging detection, and solves the technical problem that the full constraint least square method is affected by a sensor spatial point spread function, and errors exist in obtained ground object abundance. The method comprises the following steps: S1, collecting remote sensing image data; S2, estimating ground object abundance in a remote sensing image pixel by using an improved full constraint least square method; in the step S2, the improved full constraint least square method introduces the influence of the sensor point spread function when estimating the ground object abundance in the remote sensing image pixel, and improves the mixed pixel value; and S3, obtaining the estimation result of the ground object abundance in the remote sensing image pixel. The method is suitable for the technical field of remote sensing imaging.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing imaging detection technology, specifically relating to a method and system for improving the estimation of object abundance in remote sensing pixels. Background Technology

[0002] As the basic unit for recording surface information, the pixel in a remote sensing image often contains complex spectral mixing phenomena. Due to the spatial resolution limitations of current remote sensing technology, the surface area corresponding to a single pixel often does not contain only a single homogeneous feature. For example, a 30m × 30m Landsat-7 ETM+ pixel may simultaneously encompass multiple feature types such as vegetation canopy, bare soil, building roofs, and shadows. These different features have distinct spectral reflectance characteristics, and the signal ultimately received by the sensor is actually a weighted average response of the spectral features of all features within that pixel. This "mixed pixel" phenomenon is widespread in low- and medium-resolution images, and even in high-resolution images, due to the transitional and complex nature of feature boundaries, spectral mixing within pixels is difficult to completely avoid.

[0003] In existing technologies, fully constrained least squares (FCLS) is a commonly used method for estimating the abundance of land cover in pixels using remote sensing. However, due to the influence of the sensor’s point spread function (PSF), the obtained land cover abundance has errors. Summary of the Invention

[0004] To address the technical problem that the fully constrained least squares (FCLS) method is affected by the sensor's point spread function (PSF) and thus results in errors in the land cover abundance obtained, this invention provides a method to improve the estimation of land cover abundance within pixels using remote sensing.

[0005] The specific method is as follows:

[0006] S1. Acquire remote sensing image data;

[0007] S2. An improved fully constrained least squares method is used to estimate the abundance of objects in remote sensing image pixels. The improved fully constrained least squares method introduces the influence of the sensor point spread function when estimating the abundance of objects in remote sensing image pixels and improves the mixed pixel values.

[0008] S3. Obtain the crop abundance estimation results in the pixels of the remote sensing image.

[0009] Furthermore, when using the fully constrained least squares method to estimate the abundance of vegetation within pixels of remote sensing images, the following method is used: Get does not include True abundance of ground features within the spatial resolution range ,pass Obtain the true abundance of ground features within the spatial resolution range. Specifically: value inserted into get ;

[0010] in, Indicates the first Abundance of crops express With the The difference in spectral values ​​of plant species, This indicates that the original fully constrained least squares spectral matrix has the first element removed. The remaining spectral matrix after the spectral values ​​of the plant species. , This is the weight matrix. This represents the original fully constrained least squares mixed pixel value and the 1st... The difference in spectral values ​​of plant species This indicates the transpose operation.

[0011] further, The solution process is as follows:

[0012] In the original fully constrained least squares method, ,in, Indicates the blended pixel value, Represents spectral radiation in different wavelength bands. This represents the land cover abundance obtained by the original fully constrained least squares method. It is a column vector representing white noise in different bands; , Indicates the first Spectral values ​​of the plant species.

[0013] further, ;in, Indicates an element with a value of 1 Column vector of order, This represents the variance of the random errors contained in each band of the mixed pixel. for An identity matrix of order 1. This represents the Hadamard product operation. The value and The number of rows is the same.

[0014] further, Further expressed as ,in, , , .

[0015] The beneficial effects of the method described in this invention are as follows: the estimation accuracy of the traditional fully constrained least squares method is lower than expected. Using the method described in this invention, the average value of the overall error does not exceed 0.1%, thus avoiding systematic errors. Compared with the FCLS method, the improved method reduces the maximum value of the root mean square error from 12.58% to 11.29% and the minimum value from 8.29% to 8.16%. Furthermore, the root mean square error exhibits a U-shaped nonlinear variation law with the change in the degree of fragmentation. Attached Figure Description

[0016] Figure 1 This is a scene simulation diagram in an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of four bands of simulated remote sensing imagery in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram showing the decomposition results of different hybrid pixel decomposition methods in embodiments of the present invention;

[0019] Figure 4 This is a schematic diagram illustrating the statistical characteristics of the mixed pixel decomposition error when the fragmentation parameter is 0.5 in an embodiment of the present invention;

[0020] Figure 5 This is a schematic diagram of the probability distribution of the mixed pixel decomposition error when the fragmentation parameter is 0.05 in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] Example 1

[0023] To provide a more detailed and accurate description of the solutions in this invention, this embodiment first introduces FCLS.

[0024] Pixels in remote sensing images contain spectral information of various land features. As a result, the accuracy of typical remote sensing image classification is reduced. Therefore, mixed pixel decomposition methods have been introduced into remote sensing information extraction. Mixed pixel decomposition methods can be divided into linear mixed models and nonlinear mixed models. Among them, linear mixed models have the advantages of simplicity, efficiency, and clear physical meaning, and are widely used in the field of mixed pixel decomposition. The fully constrained least squares (FCLS) method is one of the most commonly used linear mixed model methods for estimating land feature abundance.

[0025] As the basis of the FCLS method, the linear mixture model regards the mixed pixel value as a linear product of land cover type spectrum, land cover type abundance and white noise, as shown in equation (1):

[0026] (1);

[0027] In the equation, It is A column vector, representing the set of pixel values ​​for a pixel in different bands, i.e., the mixed pixel values. It refers to the number of frequency bands; It is The matrix represents the spectral radiation of the endmembers in different wavelength bands. It is the number of endmembers; It is The column vector represents the abundance of each land feature in a pixel; It is a column vector representing white noise in different bands.

[0028] As can be seen from equation (1), the DN value of the mixed pixel is only related to the spectra and abundance of each object within the instantaneous field of view (IFOV), and The white noise in each band belongs to the same Gaussian distribution of random variables, and they have the same mean and variance. FCLS, as a special least squares estimation method, assumes that... In each band The solution that minimizes the sum of squares of the components is then... The sum of squares of the components in each band can be considered as The function is shown in equation (2):

[0029] (2);

[0030] in, For white noise in the first Components in each band It means The transpose of .

[0031] According to the least squares principle, the minimum value of the above function is... This involves calculating the solution when the first derivative of the above function is 0. The expression for is shown in equation (3):

[0032] (3).

[0033] Example 2

[0034] Based on the analysis of Example 1, in order to solve the problem that the fully constrained least squares method is affected by the sensor spatial point spread function (PSF) in remote sensing estimation of land cover abundance in pixels, resulting in errors in the obtained land cover abundance, this example proposes an improved fully constrained least squares method (PFCLS) that considers the influence of PSF.

[0035] The above calculation process assumes that the sensor's spatial response is an ideal spatial response, meaning the spatial response function is equal everywhere within the instantaneous field of view (IFOV) and zero outside the IFOV. However, the real spatial response function not only has different values ​​everywhere, but also has a non-zero value outside the instantaneous IFOV. This real spatial response function can be represented by the point spread function (PSF). Under the influence of the PSF, even if the spectral radiance and abundance of ground features within the IFOV are the same, different landscape patterns will result in different density (DN) values. In the process of extracting ground feature information at the sub-pixel scale, considering the influence of the sensor's point spread function can improve the accuracy of information extraction.

[0036] This method is implemented based on the established systematic error model and random error model for the decomposition error of mixed pixels. First, it determines the variance of mixed pixels in each band based on the random error model and the linear mixture model, and using the error propagation law. Then, it determines the weight of mixed pixels in each band based on the inverse relationship between variance and weight. Finally, it uses the weighted fully constrained least squares method to estimate the area proportion of ground features. Since the spatial response characteristics of the sensor and the characteristics of the ground feature landscape pattern only affect the systematic error model and the random error model, the method proposed in this embodiment is applicable to the fully constrained least squares method in all cases.

[0037] Under the influence of the sensor point spread function, the land cover abundance in equation (1) is the apparent land cover abundance of the mixed pixels, which can be expressed as the sum of the true land cover abundance within the IFOV and the land cover abundance error caused by the point spread function. Considering the influence of the point spread function, the linear mixing model shown in equation (1) can be rewritten as shown in equation (4):

[0038] (4);

[0039] in, It represents the true abundance of ground features within the spatial resolution range.

[0040] These represent random errors contained in the mixed pixel values ​​under the influence of the sensor point spread function. yes Column vectors; This represents the abundance error caused by the point spread function. yes Column vectors.

[0041] Existing research has shown that the mixed pixel decomposition error caused by PSF is a random variable exhibiting a Gaussian or approximately Gaussian distribution, and the statistical characteristics of this random variable are related to the abundance of ground features. Therefore, in equation (4) It is also a random variable. Since an approximate Gaussian distribution can be mathematically transformed to conform to the characteristics of a Gaussian distribution, it can be considered to constitute... Each component of follows a Gaussian distribution; therefore, according to the law of error propagation, variance and variance and There is a functional relationship between the variances of the mixed pixels. The variance of the random errors contained in each band is expressed by equation (5):

[0042] (5);

[0043] In the equation, Indicates the mixed pixel in the first... Error per band, , They represent the first Planting species in the first The spectral radiation and point spread function of the first band cause the... Abundance error of crop species No. The white noise contained in each band For the first The variance of each band, For the first The variance of the estimation error of the abundance of crop species. For the first The variance of white noise in each band.

[0044] Variance of random errors contained in each band of the mixed pixels It can be expressed by equation (6):

[0045] (6);

[0046] in, The Hadamard volume, representing the spectral radiation of ground features, The variance representing the error in the estimation of the abundance of ground features. This represents the variance of white noise.

[0047] Existing research indicates a significant functional relationship between the land cover abundance estimation error caused by PSF and the estimated value. Therefore, the variance of the abundance estimation error can be expressed as a function of the variance of the estimated value. Since the function is , equation (6) can be rewritten as follows:

[0048] (7);

[0049] in, This represents the functional relationship between the variance of the abundance estimation error and the abundance estimate, and is a random error model.

[0050] Comparing equations (5) and (7), it can be found that the variances of the errors contained in each band are not equal. Therefore, the estimation accuracy of the traditional fully constrained least squares method is lower than expected.

[0051] The error propagation law not only describes the relationship between the variance of the function value and the variance of each variable, but also requires that all variables involved in the variance calculation be independent variables. The FCLS method requires the sum of all abundances to be 1, therefore equation (7) needs to be transformed. The transformation method is to... Plant abundance It is expressed as the difference between 1 and other abundances, i.e.: Under these conditions, Substituting into equation (4), equation (4) can be rewritten as equation (8):

[0052] (8);

[0053] Indicates the first Spectral values ​​of crops This indicates that the original spectral matrix has the first spectral element removed. The remaining spectral matrix after the spectral analysis of the plant species. express With the The difference in spectral values ​​of plant species Indicates that it does not contain The true abundance of ground features within the spatial resolution range, Indicates an element with a value of 1 Column vectors of order;

[0054] express The middle does not include the first The matrix of crop abundance error, with a dimension ratio of Lesser 1.

[0055] For land cover abundance estimation under variance conditions, weighted least squares can be used. Compared with traditional least squares, weighted least squares reduces the impact of variance differences between bands on the results by assigning different weights to each band. Therefore, determining the weight of each band is an important step in weighted least squares. Error theory shows that there is an inverse relationship between weight and variance. Therefore, the weight of each band can be determined by the variance of each band calculated by equation (7). Since equation (7) is a column vector and the weight matrix is ​​a diagonal matrix, it is necessary to transform the column vector into a diagonal matrix through linear transformation. The weight matrix is ​​shown in equation (9).

[0056] (9);

[0057] in, Represents the weight matrix. Indicates an element with a value of 1 Column vector of order, The value and The number of rows is the same. This represents the variance of the random errors contained in each band of the mixed pixel. for An identity matrix of order 1. This represents the Hadamard product operation.

[0058] For estimating land cover abundance using the traditional least squares method shown in equation (3), since the variance of each band is equal, the weight of each band is 1. Following the weighting representation method shown in equation (9), the equation in equation (3) can be rewritten as: ,in It is the identity matrix. When the weights of each band are not equal, it can be used... Instead of the identity matrix, the abundance is obtained by weighted least squares, and the result is shown in equation (10):

[0059] (10);

[0060] Comparing the traditional least squares method shown in equation (3) and the weighted least squares method shown in equation (10), it can be seen that implementing the FCLS method using equation (10) still requires addressing two issues: the mutual influence between weights and land cover abundance, and the inconsistency in the equation form causing the estimation results to be non-negative. Regarding the first issue, due to land cover abundance... and weight matrix Since they influence each other, an iterative method can be used to gradually increase the accuracy of land cover abundance estimation. Regarding the second problem, this embodiment proposes to use a weight matrix... The equation is split into two symmetric matrices, making the form of equation (10) consistent with that of equation (3), and then the abundance is estimated using the NCLS method. The specific transformation is as follows:

[0061] (11);

[0062] in, yes A matrix formed by the square roots of all its elements. .

[0063] Finally obtained The expression is: (12).

[0064] After simplification, we get: (13), among which, , , .

[0065] pass Obtain the true abundance of ground features within the spatial resolution range. Specifically: value inserted into get .

[0066] Through the above steps, the improved method proposed in this embodiment transforms the problem of land cover abundance estimation with variance into a problem with the same form as the traditional least squares method. Then, the NCLS method is used to estimate land cover abundance, and the corrected abundance is obtained through a systematic error correction model. During the estimation process, the method proposed in this embodiment employs iterative computation to improve the estimation accuracy.

[0067] Example 3

[0068] This embodiment further defines Embodiment 2, and uses a specific simulation verification method to verify the effectiveness of the method proposed in Embodiment 2 for improving the abundance of objects in remote sensing estimated pixels.

[0069] Ground feature simulation scene generation:

[0070] To facilitate the simulation of imaging scene characteristics and subsequent calculations, this embodiment uses raster data to simulate the imaging scene. This raster image contains elements representing different land cover types, and the types of land covers are distinguished by differences in element values. One imaging scene can generate a simulated remote sensing image containing 10*10 mixed pixels.

[0071] A modified random clustering (MRC) method is used to generate specific imaging scene images. This method can generate imaging scene images with different features based on set parameters, including the number of rows and columns of the image, the initial scale and element values ​​of ground features, the initial probability value, the minimum patch size, and neighborhood rules. The number of rows and columns of the imaging scene are set to 194 and 186, respectively, with a minimum patch size of 1, meaning a single element can represent a ground feature patch. To simplify the problem, the imaging scene is set to include only two ground feature types: vegetation and soil, and the sum of the area ratios of the two types of ground features is 1. Since all elements in the raster image are of equal size, the area ratio of ground features can be converted into the proportion of element values ​​in the image. The initial area ratio of one ground feature is set to increase from 1% to 99% in 2% increments, while the area ratio of the other ground feature decreases accordingly. For each initial ground feature area ratio, 1000 different imaging scene simulation images are generated. Therefore, 100,000 mixed pixels can be generated under one initial ground feature area ratio. The parameters that most significantly affect the spatial pattern characteristics of the imaging scene are the initial probability and the neighborhood rule. We set the neighborhood rule to 4 neighborhoods and the initial probabilities to 0.05, 0.1, 0.2, 0.3, and 0.5 to analyze the impact of landscape pattern characteristics on the method. Figure 1 The image shown is a scene simulation, with white representing vegetation and black representing soil.

[0072] When simulating imaging scenarios, in addition to considering the landscape structure characteristics of ground features, the spectral characteristics of ground features should also be considered. This embodiment assumes that the mixed pixels contain spectral information from four bands, from blue light to near-infrared, and uses Landsat 8 remote sensing imagery to statistically analyze the spectral values ​​of vegetation and soil in the four bands. Finally, a linear transformation is used to convert the spectral values ​​from 16 bits to 8 bits, and the specific values ​​are shown in Table 1.

[0073] Table 1 Simulated values ​​of image scene elements:

[0074]

[0075] Simulation of remote sensing images containing mixed pixels:

[0076] Remote sensing imagery can be viewed as the result of convolving the sensor's point spread function (PSF) with the ground object's spectrum. Therefore, generating simulated remote sensing imagery requires first generating a sensor PSF matrix. The number of rows and columns in this matrix represents the sensor's spatial response range, and the element values ​​of the grid represent the sensor's spatial response values. In this embodiment, the matrix has 59 rows and 51 columns, with the 15x15 grid near the center representing the ideal spatial response range. The element values ​​in the matrix are determined by generating a two-dimensional Gaussian function with a set parameter and performing equally spaced sampling. In this embodiment, by fitting the PSF characteristics of the IKONOS sensor as shown in the literature, the two standard deviation parameters in the two-dimensional Gaussian function are determined to be 0.6497 and 0.5563, respectively. For the generated two-dimensional Gaussian function, equally spaced sampling is performed in both the row and column directions between [-3*0.6497, 3*0.6497], thus forming a 59x51 two-dimensional matrix. Based on this, the element values ​​in the matrix are normalized to ensure that the sum of the element values ​​is 1.

[0077] The simulated remote sensing images generated in this embodiment are divided into images without white noise and images with white noise. The former is used to construct a mathematical model of the error caused by PSF, while the latter is used to verify the effectiveness of the improved method proposed in this invention. For the simulated mixed pixel values ​​in the simulated image containing white noise, this embodiment uses a Gaussian random number function to generate four random numbers with a mean of 0 and a standard deviation of 1. After convolving the imaging scene data of each band with the point spread function, a random number is added to the result of each band and then rounded to obtain the simulated mixed pixel values ​​containing multiple bands.

[0078] Figure 2 The four bands of the simulated remote sensing image are represented by (a) to (d), which are the first to fourth band images respectively.

[0079] Simulation of hybrid pixel decomposition:

[0080] Constructing systematic and random error models of mixed pixel decomposition caused by point spread function using remote sensing images without white noise is a prerequisite for obtaining land cover abundance using improved methods. For a simulated mixed pixel, the true abundance of vegetation within the ideal range is statistically analyzed; at the same time, using the pixel value of the mixed pixel in a certain band, the estimated value of vegetation abundance in the mixed pixel is calculated using the formula shown in equation (11) below. This estimated value and the true abundance can form an error data pair.

[0081] ;

[0082] In the equation, An estimate representing vegetation abundance. Indicates the mixed pixel in the first... Pixel values ​​on the band, Indicates the soil end-member in the first Pixel values ​​in each band Indicates the vegetation endmember in the first Pixel values ​​in each band.

[0083] Error data pairs are grouped and statistically analyzed to form an error statistics matrix, where rows represent the true land cover abundance and columns represent the estimated land cover abundance. Based on the error statistics matrix, the probability of the true land cover abundance occurring under the estimated land cover abundance is calculated. The difference between the expected value of the true land cover abundance and the estimated land cover abundance is the systematic error of the land cover abundance estimate, while the median error of the true land cover abundance is the random error of the land cover abundance estimate. By fitting the variation patterns of systematic error and random error with the estimated land cover abundance, systematic error functions and random error functions are constructed.

[0084] For each mixed pixel in a remote sensing image containing white noise, the estimated vegetation abundance within that mixed pixel is obtained using both the FCLS method and the improved PFCLS method. Since the improved method employs an iterative computation strategy in estimating land cover abundance, this embodiment uses a threshold of 0.01% as the iteration termination point for the difference between two consecutive estimation results. Simultaneously, the true vegetation abundance within the ideal range of the scene corresponding to that mixed pixel is statistically analyzed; the difference between the estimated and true vegetation abundance values ​​represents the error of the corresponding method.

[0085] Results and Analysis:

[0086] To demonstrate the existence of inter-band variance in remote sensing images, this embodiment first uses an ideal spatial response function to form error-free multi-band mixed pixels for the imaging scene; then, according to equation (1), mixed pixels containing only white noise are generated. Simultaneously, a point spread function is used to form multi-band mixed pixels, and white noise is added to form true mixed pixels containing both the influence of the point spread function and white noise. Based on this, the errors between the mixed pixel values ​​containing only white noise and the true mixed pixel values ​​and the ideal pixel values ​​are calculated respectively. By statistically analyzing the variance of errors in each band, the differences between band variances are revealed. Table 2 shows the standard deviation of the random errors contained in the pixel values ​​of each band when the fragmentation parameter is 0.5. It can be seen from the table that when there is no influence from the sensor's point spread function, the error contained in the mixed pixels is white noise; therefore, the standard deviation of the error in each band pixel value is equal to the standard deviation of white noise. However, under the influence of the sensor's point spread function, the noise of the mixed pixels is affected by both white noise and the point spread function; at the same time, there are significant differences between the standard deviations of noise in different bands, therefore, the least squares method cannot be used to estimate the abundance of ground features.

[0087] Table 2 Standard deviation of errors for each band:

[0088]

[0089] Figure 3 The diagram shows the decomposition results of different mixed pixel decomposition methods. (a) is the actual land cover abundance, (b) is the result of the FCLS method, and (c) is the result of the PFCLS method.

[0090] from Figure 3 As can be seen, the simulation process can simultaneously acquire the true abundance of each mixed pixel and the estimated abundance of different decomposition methods. The error between the estimated abundance and the true abundance is the estimation error of different methods. In this embodiment, the errors of different methods are grouped and statistically analyzed at 0.1% intervals to obtain the error probability distribution map corresponding to different estimated abundances. Based on this, the average error and root mean square error corresponding to different estimates are further calculated, as well as the probability distribution of the errors and the cumulative probability of errors occurring within a certain range based on 0. Among them, the average error and root mean square error represent the average error and root mean square error within a 2% estimation interval.

[0091] Figure 4 The statistical characteristics of the errors of the two methods are shown when the breakage parameter is 0.5. Figure 4 (a) shows the variation of the average error across different abundance estimation intervals. The figure reveals that the error of the FCLS method varies between -8% and 8%, exhibiting a clear sinusoidal trend: the average error reaches its maximum when the estimated abundance is 25%, its minimum when the estimated abundance is 75%, and is 0 when the estimated abundance is 50%. Compared to the FCLS method, the improved method has a smaller average error, varying between -3% and 2%. This difference indicates that the FCLS method suffers from systematic bias in a single estimation, and this bias is correlated with the estimation results, while the improved method effectively reduces this systematic bias. Figure 4 (b) shows the variation of the root mean square error (RMSE) for the two methods. As can be seen from the figure, the RMSE of the FCLS method varies between 2% and 13%, exhibiting a symmetrical distribution. Compared to the FCLS method, the RMSE of the PVCLS method also shows a symmetrical distribution, but the range is slightly smaller, varying between 3% and 12%. Furthermore, the RMSE of the PVCLS method is smaller than that of the FCLS method within the range of 16% to 84% abundance estimation, while the opposite is true in other ranges. This characteristic indicates that the PVCLS method has higher accuracy when the abundance estimation ratio is between 16% and 84%.

[0092] Figure 4(c) shows the overall probability distribution of errors from different decomposition methods. As can be seen from the figure, the mixed pixel decomposition errors caused by the two methods share several common statistical characteristics: First, the errors caused by both methods exhibit a clear symmetrical distribution, with the center of symmetry near 0. This characteristic indicates that, at the global scale, neither method has a systematic error in abundance estimation. Second, the variation range of mixed pixel decomposition errors caused by both methods is mainly distributed between -30% and 30%, indicating that the limits of errors introduced by the two methods are consistent. Besides these two similarities, the mixed pixel decomposition errors caused by the two methods also show significant differences. Compared to the mixed pixel decomposition error caused by the traditional FCLS method, the maximum probability of the mixed pixel decomposition error caused by the improved method is approximately 0.4%, while the maximum probability of the error caused by the FCLS method is 0.28% overall. Furthermore, the improved method has a lower probability of occurring at large error points and a higher probability of occurring at small error points. To quantitatively analyze the above differences, this embodiment calculated the kurtosis of the probability distribution of the errors introduced by the two methods in mixed pixel decomposition. The results show that the kurtosis of the improved method is -0.05, while that of the FCLS method is -0.51. The difference in the frequency distribution and kurtosis of the errors also indicates that, for a single mixed pixel decomposition, the abundance results obtained by the improved method are more reliable. The accuracy of abundance estimation is usually quantitatively expressed by the root mean square error. In this case, the root mean square error of the improved method is 11.29%, while that of the FCLS method is 12.58%. This result again shows that the improved method can improve the accuracy of abundance estimation.

[0093] Figure 4 (d) shows the cumulative probability of the mixed pixel decomposition error caused by the two methods occurring within a certain range near 0. For example, 5% in the figure represents the cumulative probability of the mixed pixel decomposition error occurring between -5% and 5%. It can be seen from the figure that when the range is small, the cumulative probability of the mixed pixel decomposition error caused by the improved PFCLS method is greater than that caused by the FCLS method. For example, approximately 30% of the error caused by FCLS is distributed between -5% and 5%, while approximately 35% of the error caused by PFCLS is distributed in this region. The fact that more errors appear in a smaller range of variation further demonstrates that the improved method can improve the accuracy of mixed pixel decomposition.

[0094] Figure 5 The statistical characteristics of the mixed pixel decomposition error are shown when the fragmentation parameter is 0.3. (Comparison) Figure 4 and Figure 5 It can be observed that in scenarios with different pattern characteristics, the errors introduced by the two methods maintain a similar pattern of change, but the specific feature values ​​change. First, compare... Figure 4 (a) and Figure 5(a) It can be observed that the variation pattern of systematic errors included in different estimation errors has not changed. The average variation range of the error caused by the FCLS method has decreased to between -5% and 5%, while the average variation range of the error caused by the improved method has changed to between -6% and 6%. However, when the abundance estimation ratio is between 16% and 84%, the accuracy of the PFCLS method is still higher. Secondly, comparing the changes in RMSE for each estimated abundance, it can be found that the variation range of the root mean square error of both methods has changed to between 6% and 9%, and the relative magnitude of the RMSE of the two methods at the same estimated abundance still has the characteristic of piecewise variation. However, the endpoints of the segments have changed from 16% and 84% under parameter 0.5 to 20% and 80% under parameter 0.3. Third, regarding the frequency variation of the overall abundance estimation error, compared to the scenario with a parameter of 0.5, the error variation range caused by the two methods narrows in the scenario with a parameter of 0.3. However, the probability of zero error for the PVCLS method increases from 0.40% to 0.55%, while the probability of zero error for the FCLS method increases from 0.3% to 0.45%. This characteristic indicates that the accuracy of both methods is improving, but the accuracy of the improved method is still higher than that of the FCLS method. Finally, in the scenario with a parameter of 0.3, the cumulative probability of error between the two methods no longer shows a significant difference.

[0095] Since the landscape pattern of the scene affects the accuracy of both methods, Table 3 shows the trends of the mean, root mean square error (RMSE), and kurtosis of the errors caused by the two methods under different landscape fragmentation parameters. The table shows that the average error of the mixed pixel decomposition caused by both methods does not exceed 0.1%, indicating that neither method causes significant systematic errors at the global scale. Under different landscape fragmentation parameters, the RMSE of the FCLS method varies between 8.29% and 12.58%, while the RMSE of the PFCLS method varies between 8.16% and 11.29%. This difference indicates that the improved method can improve the overall accuracy of the mixed pixel decomposition. Comparing the RMSE of the mixed pixel decomposition results caused by the two methods reveals that both methods exhibit a U-shaped trend, and the RMSE is smallest when the landscape fragmentation parameter is 0.2. This phenomenon indicates that the accuracy of both methods is highest when the landscape fragmentation parameter is 0.2. Furthermore, the root mean square error (RMSE) induced by the improved method is consistently smaller than that induced by the traditional FCLS method, further demonstrating that the improved method can enhance the accuracy of mixed pixel decomposition. The kurtosis variations of the probability distributions of the errors induced by the two methods differ significantly: the kurtosis of the error induced by the improved method ranges from -0.05 to 0 with little variation, while the kurtosis of the error induced by the FCLS method ranges from -0.51 to -0.10, exhibiting a distinct inverted U-shaped variation. Moreover, the kurtosis of the improved method is consistently greater than that of the FCLS method. This result indicates that, under different landscape fragmentation parameters, the results of the improved method are more reliable than those of the FCLS method.

[0096] Table 3. Error statistics of improvement methods under different landscape patterns (%)

[0097]

[0098] Since the landscape pattern of the scene significantly affects the accuracy of the method, it is necessary to determine the landscape structure characteristics of the imaging scene of the pixel before applying the improved method, and then determine the required systematic error model and random error model. However, there may be obvious differences in landscape structure characteristics between different regions of a remote sensing image, which will cause the accuracy of the method to change in different regions. To solve this problem, the entire image can be divided into different sub-regions by image segmentation to ensure that the pixels in the sub-regions have similar landscape structures. Then, error models are constructed in each sub-region and the improved method is applied to perform mixed pixel decomposition. Combining equation (5) and Table 2, it can be found that in addition to the landscape pattern, the quantization level and signal-to-noise ratio of the remote sensing image will also affect the variance difference between bands. The increase of the quantization level or the decrease of the signal-to-noise ratio will increase the variance difference between bands, thereby reducing the accuracy of the FCLS method and making the improved method more effective in improving accuracy.

[0099] In the linear hybrid model, both land cover abundance and endmember spectral matrix have errors. This embodiment did not consider the error in the spectral matrix caused by the "different spectra for the same object" phenomenon during the analysis and improvement process. Combining the error propagation law and equation (5), it can be seen that, considering the errors in land cover abundance caused by the sensor point spread function, endmember spectral errors, and white noise, the error contained in the band can be expressed as the nonlinear sum of the above three errors. For the above problem, the error propagation law of the nonlinear equation can be used to calculate the total variance of the band based on the variance of the three components, and then the new method proposed in this invention can be used for hybrid pixel decomposition. In this case, the method can simultaneously reduce the influence of both the sensor point spread function and the "different spectra for the same object" factor on the linear hybrid pixel decomposition method.

Claims

1. A method for improving the estimation of object abundance in pixels using remote sensing, characterized in that, The method is specifically as follows: S1. Acquire remote sensing image data; S2. An improved fully constrained least squares method is used to estimate the abundance of objects in remote sensing image pixels. The improved fully constrained least squares method introduces the influence of the sensor point spread function when estimating the abundance of objects in remote sensing image pixels and improves the mixed pixel values. When the improved fully constrained least squares method estimates the abundance of vegetation within pixels of remote sensing images, it does so by: Get does not include True abundance of ground features within the spatial resolution range ,pass Obtain the true abundance of ground features within the spatial resolution range. Specifically: value inserted into get ; in, Indicates the first Abundance of crops express With the The difference in spectral values ​​of plant species, This indicates that the original fully constrained least squares spectral matrix has the first element removed. The remaining spectral matrix after the spectral values ​​of the plant species. , This is the weight matrix. This represents the original fully constrained least squares mixed pixel value and the 1st... The difference in spectral values ​​of plant species, Indicates the transpose operation; S3. Obtain the crop abundance estimation results in the pixels of the remote sensing image.

2. The method for improving the abundance of objects in remotely sensed pixels according to claim 1, characterized in that, The solution process is as follows: In the original fully constrained least squares method, ,in, Indicates the blended pixel value, Represents spectral radiation in different wavelength bands. This represents the land cover abundance obtained by the original fully constrained least squares method. It is a column vector representing white noise in different bands; , Indicates the first Spectral values ​​of the plant species.

3. The method for improving the abundance of vegetation in remotely sensed pixels according to claim 2, characterized in that, ;in, Indicates an element with a value of 1 Column vector of order, This represents the variance of the random errors contained in each band of the mixed pixel. for An identity matrix of order 1. This represents the Hadamard product operation. The value and The number of rows is the same.

4. The method for improving the abundance of objects in remotely sensed pixels according to claim 3, characterized in that, Further expressed as ,in, , , .

5. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-4.

6. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-4.

Images