Vegetation coverage space-time fusion algorithm based on proximity effect

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a vegetation coverage space-time fusion algorithm based on a proximity effect, which comprises the following steps of: acquiring coarse spatial resolution image data and fine spatial resolution image data of multiple time points, and acquiring a coarse spatial resolution image of a target time point; selecting samples based on a Monte Carlo random sampling method in a set moving window, and constructing a proximity factor calculation formula to screen an effective pixel set; and constructing a scale transformation model by using a proximity factor, converting the coarse spatial resolution image data into fine spatial resolution image data of a target time point, and generating a vegetation coverage prediction image with unified high time resolution and fine spatial resolution. According to the method, Monte Carlo random sampling and multi-dimensional proximity factor calculation are introduced, so that the calculation efficiency is remarkably improved, the applicability of the algorithm in a complex earth surface scene in an arid region is enhanced, and the method is suitable for the fields of ecological monitoring, global vegetation dynamic monitoring and the like.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a spatiotemporal fusion algorithm of vegetation coverage based on proximity effect. Background Art

[0002] Vegetation coverage is an important indicator for measuring the dynamic changes and distribution of surface vegetation, and is also an important basis for ecological restoration, natural disaster response and environmental protection research. However, since satellite image acquisition is often affected by ice and snow cover, cloud cover and equipment defects, high spatial resolution vegetation coverage data has problems such as missing pixels and spatial discontinuity, especially in dynamic monitoring of large-scale areas. This problem is particularly significant. In addition, fine spatial resolution data is difficult to meet the temporal resolution requirements for efficient monitoring of large areas due to its small coverage width and long acquisition cycle. In response to the above challenges, spatiotemporal fusion algorithms have gradually become an important means to solve the spatiotemporal consistency of vegetation coverage products.

[0003] Existing technologies mainly use spatiotemporal fusion algorithms based on demixing methods, weighted function methods and machine learning methods, such as multi-sensor multi-resolution technology (MMT), spatiotemporal adaptive reflectance fusion model (STARFM) and enhanced spatiotemporal adaptive reflectance fusion model (ESTARFM). However, these methods have the following shortcomings: the demixing method relies on spectral characteristics to decompose pixels, and the accuracy decreases in areas with complex surface features (such as heterogeneous landscapes); the moving window algorithm requires pixel-by-pixel iterative calculations, which is especially inefficient in large areas and difficult to meet the needs of efficient monitoring; for example, the STARFM algorithm is only applicable to uniform landscape areas, and the prediction accuracy is poor in heterogeneous oases or complex surface features. Summary of the invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a spatiotemporal fusion algorithm of vegetation coverage based on the proximity effect. Based on the spatial autocorrelation and temporal continuity characteristics of the ground objects, the present invention introduces the Monte Carlo random sampling method in the moving window, screens the effective pixel set adjacent to the target pixel, and finally generates a vegetation coverage prediction image with high temporal resolution and fine spatial resolution by constructing a proximity factor calculation formula and a scale transformation model. The present invention significantly improves the computational efficiency and applicability, and is suitable for dynamic monitoring of large-scale heterogeneous landscapes.

[0005] A spatiotemporal fusion algorithm of vegetation coverage based on proximity effect includes the following steps:

[0006] Collect coarse spatial resolution images and fine spatial resolution image data at multiple time points to obtain coarse spatial resolution image data at a target time point;

[0007] In a set moving window, samples are selected based on a random sampling method, and sample screening rules are constructed to screen a set of valid pixels, wherein the valid pixels are determined by the spectral similarity, spatial distance, and temporal characteristics of adjacent pixels;

[0008] Based on the effective pixel set, a proximity factor is calculated to comprehensively characterize the contribution of the effective pixel to the reflectivity of the target pixel;

[0009] The proximity factor is used to construct a scale transformation model, and the coarse spatial resolution image data is converted into fine spatial resolution image data of the target time point, so as to generate a vegetation coverage prediction image at the target time point.

[0010] Preferably, the size of the moving window is dynamically adjusted according to the complexity of the surface features.

[0011] Preferably, the random sampling method adopts Monte Carlo random sampling to improve computational efficiency by reducing the number of samples in the moving window.

[0012] Preferably, the effective pixels are determined by a sample screening rule, wherein the samples are screened jointly based on spectral similarity and spatial distance.

[0013] Preferably, the spatial distance D i It is calculated by the following Manhattan distance formula:

[0014] D i =1+(|x w / 2 -x i |+|y w / 2 -y i |) / (w / 2)

[0015] Among them, w represents the size of the moving window, x w / 2 ,y w / 2 Respectively represent the horizontal and vertical coordinates of adjacent pixels in the moving window, x i ,y i Respectively represent the horizontal and vertical coordinates of the center pixel.

[0016] Preferably, the proximity factor W i Calculated by the following formula:

[0017]

[0018] Among them, S i A comprehensive measure representing spectral similarity and spatial distance characteristics.

[0019] Preferably, the scale transformation model is constructed by a proportional relationship between a coarse spatial resolution image and a fine spatial resolution image, wherein the proportional relationship is determined by a spectral difference and a time weight.

[0020] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0021] The present invention introduces Monte Carlo random sampling to reduce the number of similar pixel screening times, thereby significantly improving the computational efficiency of large-scale areas while maintaining the accuracy of the algorithm.

[0022] The present invention adopts multi-dimensional proximity factors such as spectral similarity, Manhattan distance and time weight to solve the fusion accuracy problem caused by the complexity of surface features in heterogeneous areas.

[0023] The present invention ensures the unification of high temporal resolution and fine spatial resolution data by constructing a scale transformation model and a proximity effect index. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the flow chart of the algorithm of the present invention;

[0025] Figure 2 Schematic diagrams showing the comparison of using the Euclidean distance describing a single path to describe proximity and the multi-path Manhattan distance used in the present invention. DETAILED DESCRIPTION

[0026] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0027] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "include", "comprising", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0028] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0029] Collect coarse spatial resolution (CSR) images and fine spatial resolution (FSR) image data at multiple time points to obtain coarse spatial resolution image data at the target time point. The spatiotemporal fusion method is a method that uses the spatial autocorrelation and temporal continuity characteristics of the ground object to convert the coarse resolution reflectivity change information into fine spatial resolution reflectivity. As shown in formula (1):

[0030] L(x w / 2 ,y w / 2 ,t p ,F)=T m ×L1(x w / 2 ,y w / 2 ,t p ,F)+T n ×L2(x w / 2 ,y w / 2 ,t p ,F)

[0031] Among them, L(x w / 2 ,y w / 2 ,t p ,F) represents the image of the predicted time of FSR (t p ), F represents the FVC image at the predicted time, L1(x w / 2 ,y w / 2 ,t p ,F) and L2(x w / 2 ,y w / 2 ,t p , F) represent t k (k=1or2) FSR image at time (t p ) image prediction results, T m and T n Represents the CSR image distance t at time t1 and t2 respectively p The time weight of the CSR image at the moment. The calculation of L1 and L2 requires the input of the prediction time (t p ) and at least two sets of reference time (t k ,) to calculate the CSR and FSR images of t p FSR image of time.

[0032] Calculating L1 and L2, the method includes three steps: target time sampling pixel selection, screening sampling pixels to obtain effective pixels, and calculating the proximity of effective pixels.

[0033] Among them, the prediction time t p The image is based on the reference time t k The image data at a certain moment is calculated. Formula (2) is as follows:

[0034]

[0035] In the formula, L(x w / 2 ,y w / 2 ,t p , F) and L(x w / 2 ,y w / 2 ,t k , F) represent the prediction time (t p ) and reference time (t k ), F represents the FVC image at that moment, M(x i ,y i ,t p , F) and M(x i ,y i ,t k , F) represent the prediction time (t p ) and reference time (t k ) is the image of the moving window size (x, y) in the CSR image; w represents the size of the moving window, V i represents the scale conversion factor between CSR and FSR. At the pixel scale, the land type in arid areas does not change significantly at coarse and fine spatial resolutions. Therefore, a simple linear equation can be used as the scale transformation equation between CSR and FSR images to further describe the relationship between CSR and FSR images. In order to solve W in the above expression i , W i represents the neighboring factor in the moving window, which represents the relationship between each selected pixel and the center pixel (x w / 2 ,y w / 2 ), the subscript w represents the size of the moving window. In order to calculate W i , the moving window in previous studies used a 25*25 moving window. The moving window is used to calculate the entire fine spatial resolution image. The 25*25 moving window is slid with a step size of one pixel, and all pixels within the window range are involved in the calculation. Therefore, the calculation efficiency is very low. In order to solve this problem, the present invention experimentally designs the window size and the number of random samples. In order to randomly determine the selected pixels in the moving window, the present invention uses the Monte Carlo method. Then, the judgment function selects the valid pixels according to the different values ​​between the selected pixels and the center pixel. Finally, in order to determine the proximity effect between the valid pixels and the center pixel in the moving window, the proximity index is used.

[0036] To calculate W in the moving window i, the present invention designs a Monte Carlo random sampling method to determine the selected pixels. Unlike pixel-by-pixel iterative calculation, the present invention uses a Monte Carlo random sampling method to randomly select pixels in a moving window, which improves the calculation efficiency. Monte Carlo simulation is a method for simulating statistical systems. Its purpose is to generate a representative set of configurations to access thermodynamic quantities without analytically solving the system or performing precise enumeration.

[0037] According to formula (2), it is necessary to select images at two time points, namely T1 and T2, for the CSR image and the FSR image. p is the general symbol of the CSR image or FSR image at the prediction time, T1 is the general symbol of the CSR image or FSR image before the prediction time, and T2 is the general symbol of the CSR image or FSR image after the prediction time. p CSR image at time point to predict T p When obtaining an FSR image at a time point, a moving window needs to be used. The present invention takes into account the characteristics of the desert landscape and designs a moving window. The size of the moving window has been determined, and the Monte Carlo method needs to be used to select the number of selected samples. The number of selected samples is based on the setting of the number of random samples, which is regarded as prior knowledge. In order to determine the similarity between the sampled pixels and the center pixel, a judgment function needs to be introduced. Considering that vegetation growth may change from T1 to T2, in order to obtain the sampled pixels in the common area of ​​T1 and T2, the present invention uses the high-resolution images acquired at T1 and T2 to select the sampled pixels, respectively, and then extracts the intersection of the two results to obtain a more accurate set of valid pixels. Therefore, function construction is the key to effective pixel screening for high-resolution images acquired at T1 and T2.

[0038] In the moving window, the present invention selects a small number of samples by Monte Carlo sampling. In order to determine whether these samples are adjacent pixels of the central pixel, the present invention introduces a proxy function, as shown in formula (3):

[0039] |L(x i ,y i ,t k ,F)-L(x ω / 2 ,y ω / 2 ,t k ,F)|≤C v (B)

[0040] Among them, L(x i ,y i ,t k ,F) and L(x w / 2 ,y w / 2 ,t k, F) represent the adjacent pixels and the central pixel in the moving window respectively; C v (B) is the reflectance variation coefficient of the B band in the FSR, which is defined as the ratio of the standard deviation of the reflectance to the mean value. Formula (3) is similar to the cost function in the inversion. Most previous studies considered a fixed value, that is, the impact of the number of land types in the study area on the results, which requires a lot of prior knowledge and is difficult to determine. Therefore, the present invention does not consider the impact of the number of land types and minimizes the error in selecting effective pixels, which makes the FSRFVC (vegetation coverage) in formula (2) using effective pixels more accurate.

[0041] After selecting neighboring pixels, the next step is to calculate the proximity of each neighboring pixel. The proximity between pixels is mainly determined by spectral similarity, distance and time. The temporal metric is used as the weight of the two reference temporal images when calculating the predicted image (t k ), so the temporal metric is not considered when calculating the predicted image of a single reference temporal image. i Represents the neighborhood effect of each adjacent similar pixel on the reflectance estimate of the central pixel, considering that constraints need to be imposed in multiple aspects to ensure the robustness of the algorithm:

[0042]

[0043] Among them, S i The normalized inverse of is used as the weight W i , S i Similar pixels with larger values ​​contribute less to the change in the reflectance of the central pixel. Formula (5) is:

[0044] S i =(1-R i )×D i

[0045] Among them, D i Represents the distance between the effective pixel and the center pixel in the moving window. The distance between them is calculated using Manhattan distance. Formula (6) is:

[0046] D i =1+(|x w / 2 -x i |+|y w / 2 -y i |) / (w / 2)

[0047] Among them, D irepresents the Manhattan distance between the adjacent pixels and the central pixel in the moving window, and w represents the size of the moving window. Previous studies used Euclidean distance to calculate the distance between adjacent pixels and the central pixel. Unlike non-arid areas, arid areas are highly homogeneous, and Monte Carlo algorithms have been used to reduce the sample size of adjacent pixels. Therefore, it is difficult to describe proximity using only Euclidean distance that describes a single path (e.g. Figure 2 (a)). Therefore, the present invention adopts multi-path Manhattan distance (such as Figure 2 (as shown in (b)).

[0048] Secondly, the similarity of vegetation coverage between the high-resolution and low-resolution pixels of the effective pixels is also used as one of the weight factors. The closer the vegetation coverage of the effective pixel is to that of the central pixel, the smaller the distance is, and the greater the weight of the effective pixel (i.e., the greater the contribution). The spectral similarity here is determined by the correlation coefficient between each effective pixel and its corresponding low-resolution pixel, as shown in formula (7):

[0049]

[0050] Among them, R i is the spectral correlation coefficient between high-resolution pixels and low-resolution pixels of similar pixels, L i and M i They are t m and t n The vegetation coverage vector of similar pixels at the moment and their corresponding low-resolution pixels, E(·) represents the expected value, D(L i ) and D(M i ) are L i and M i The variance of R i The value range is -1 to 1, R i The larger it is, the higher the proximity of vegetation cover.

[0051] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware.

[0052] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A spatiotemporal fusion algorithm for vegetation coverage based on proximity effect, characterized in that: The following steps are involved: Collect coarse spatial resolution images and fine spatial resolution image data at multiple time points to obtain coarse spatial resolution image data at a target time point; In a set moving window, samples are selected based on a random sampling method, and sample screening rules are constructed to screen a set of valid pixels, wherein the valid pixels are determined by the spectral similarity, spatial distance, and temporal characteristics of adjacent pixels; Based on the effective pixel set, a proximity factor is calculated to comprehensively characterize the contribution of the effective pixel to the reflectivity of the target pixel; The proximity factor is used to construct a scale transformation model, and the coarse spatial resolution image data is converted into fine spatial resolution image data of the target time point, so as to generate a vegetation coverage prediction image at the target time point.

2. The spatiotemporal fusion algorithm of vegetation coverage based on proximity effect according to claim 1 is characterized in that: The size of the moving window is dynamically adjusted according to the complexity of the surface features.

3. The spatiotemporal fusion algorithm of vegetation coverage based on proximity effect according to claim 2 is characterized in that: The random sampling method adopts Monte Carlo random sampling and improves computational efficiency by reducing the number of samples in the moving window.

4. The spatiotemporal fusion algorithm of vegetation coverage based on proximity effect according to claim 1 is characterized in that: The effective pixels are determined by a sample screening rule, wherein the samples are screened jointly based on spectral similarity and spatial distance.

5. The spatiotemporal fusion algorithm of vegetation coverage based on proximity effect according to claim 1 is characterized in that: The spatial distance D i It is calculated by the following Manhattan distance formula: D i =1+(|x w / 2 -x i |+|y w / 2 -y i |) / (w / 2) Among them, w represents the size of the moving window, x w / 2 ,y w / 2 Respectively represent the horizontal and vertical coordinates of adjacent pixels in the moving window, x i ,y i Respectively represent the horizontal and vertical coordinates of the center pixel.

6. The spatiotemporal fusion algorithm of vegetation coverage based on proximity effect according to claim 1 is characterized in that: The proximity factor W i Calculated by the following formula: Among them, S i A comprehensive measure representing spectral similarity and spatial distance characteristics.

7. The spatiotemporal fusion algorithm of vegetation coverage based on proximity effect according to claim 1 is characterized in that: The scale transformation model is constructed by the proportional relationship between the coarse spatial resolution image and the fine spatial resolution image, wherein the proportional relationship is determined by the spectral difference and the time weight.

Images