A method and apparatus for surface reflectance fusion based on Gaussian model

Abstract

This invention provides a method and apparatus for surface reflectance fusion based on a Gaussian model, belonging to the field of remote sensing technology. The method includes: identifying various types of land cover based on first image data; acquiring the first reflectance of the first image data, the first observed value of a target pixel in the second image data, and the second observed value of a target pixel in the third image data; determining the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to each type of land cover in the second image data, the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to each type of land cover in the third image data, and the first area proportion of each type of land cover in the target pixel; and predicting the second reflectance of the fourth image data based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameters, and the first area proportion. This invention improves the computational efficiency of spatiotemporal fusion and has low computational resource requirements.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and in particular to a method and apparatus for fusion of surface reflectance based on a Gaussian model. Background Technology

[0002] High spatiotemporal resolution, dense time-series satellite remote sensing data is crucial for accurate monitoring of Earth's surface dynamics. To acquire high temporal and spatial resolution imagery simultaneously, various spatio-temporal fusion (STF) methods have been developed.

[0003] The effectiveness of these STF methods has been tested in many specific land monitoring tasks; however, the heavy computational demands and inefficiency of these complex algorithms make them difficult to apply on a large scale. On the one hand, learning-based methods, especially deep learning methods, have achieved unprecedented success in prediction accuracy, but most require at least two high-resolution datasets as input, demanding significant computational resources. On the other hand, many weighted, unmixed, Bayesian, and mixture-based methods can work with only one high-resolution dataset, suitable for situations where fine data is scarce, but they typically rely on empirical and pixel-by-pixel filtering steps, which are very time-consuming. In many cases, a slight improvement in accuracy comes at the cost of adding a large number of additional empirical steps.

[0004] In other words, existing technologies suffer from high demands for computing resources and low computing efficiency. Summary of the Invention

[0005] This invention provides a surface reflectance fusion method and apparatus based on a Gaussian model to solve the technical problems of high computational resource requirements and low computational efficiency in the spatiotemporal fusion of satellite image data in the prior art.

[0006] This invention provides a surface reflectance fusion method based on a Gaussian model, comprising:

[0007] Various types of land cover are identified based on the first image data; the first image data is the first satellite image data at the first moment.

[0008] The first reflectance of the first image data, the first observed value of the target pixel in the second image data, and the second observed value of the target pixel in the third image data are obtained; the second image data is the second satellite image data at the first time, and the third image data is the second satellite image data at the second time.

[0009] The Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area proportion of various types of land cover in the target pixel are determined respectively.

[0010] Based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameter, and the first area ratio, the second reflectance of the fourth image data is predicted; the fourth image data is the first satellite image data at the second time point.

[0011] The spatial resolution of the second satellite image data is lower than that of the first satellite image data.

[0012] According to the present invention, a surface reflectance fusion method based on a Gaussian model is provided, wherein identifying various types of surface cover based on first image data includes:

[0013] The first image data is classified or clustered to obtain a classification / clustering map;

[0014] Identify different types of land cover based on the classification / clustering diagram.

[0015] According to the present invention, a surface reflectance fusion method based on a Gaussian model is provided to determine each pure pixel corresponding to a target surface cover type, including:

[0016] Obtain the second area percentage of the target land cover type in any pixel of the second image data or the third image data;

[0017] If the second area ratio reaches a preset ratio, then any pixel is determined to be a pure pixel corresponding to the target land cover type.

[0018] According to a Gaussian model-based surface reflectance fusion method provided by the present invention, after determining that any pixel is a pure pixel corresponding to the target surface cover type if the second area proportion reaches a preset ratio, the method further includes:

[0019] If the number of pure pixels corresponding to the target land cover type exceeds the first threshold, the preset ratio is increased, and each pure pixel corresponding to the target land cover type is re-determined according to the new preset ratio, until the number of pure pixels corresponding to the target land cover type is less than the first threshold.

[0020] If the number of pure pixels corresponding to the target land cover type is less than the second threshold, the preset ratio is gradually reduced, and each pure pixel corresponding to the target land cover type is re-determined according to the new preset ratio, until the number of pure pixels corresponding to the target land cover type reaches the second threshold.

[0021] Wherein, the first threshold is greater than the second threshold.

[0022] According to the Gaussian model-based land reflectance fusion method provided by the present invention, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover are determined, including:

[0023] The mean and variance of the reflectance of each pure pixel corresponding to a single land cover type are determined to obtain the Gaussian distribution parameters corresponding to the single land cover type.

[0024] According to a Gaussian model-based surface reflectance fusion method provided by the present invention, after predicting the second reflectance of the fourth image data based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameter, and the first area proportion, the method further includes:

[0025] The confidence interval range of the second reflectance is determined based on the Gaussian distribution parameters corresponding to the third image data and the first area ratio;

[0026] If the second reflectance is outside the confidence interval, then the second reflectance is re-predicted based on the first reflectance, the first observation, and the second observation.

[0027] The present invention also provides a surface reflectance fusion device based on a Gaussian model, comprising:

[0028] The identification module is used to identify various types of land cover based on the first image data; the first image data is the first satellite image data at the first moment.

[0029] The acquisition module is used to acquire the first reflectance of the first image data, the first observed value of the target pixel in the second image data, and the second observed value of the target pixel in the third image data; the second image data is the second satellite image data at the first time, and the third image data is the second satellite image data at the second time.

[0030] The determination module is used to determine the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area proportion of various types of land cover in the target pixel.

[0031] The prediction module is used to predict the second reflectance of the fourth image data based on the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameter, and the first area ratio; the fourth image data is the first satellite image data at the second time point;

[0032] The spatial resolution of the second satellite image data is lower than that of the first satellite image data.

[0033] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the surface reflectance fusion method based on the Gaussian model as described above.

[0034] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the surface reflectance fusion method based on the Gaussian model as described above.

[0035] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the surface reflectance fusion method based on the Gaussian model as described above.

[0036] The surface reflectance fusion method and apparatus based on Gaussian model provided by this invention assumes that the reflectance of the surface cover follows an independent Gaussian distribution. Then, it predicts the reflectance of high-resolution image data through Gaussian distribution parameters, one high-resolution image data and two low-resolution image data. The spatiotemporal fusion of satellite image data does not rely on experience and pixel-by-pixel filtering, thus improving computational efficiency. The requirement for computational resources is low with only one high-resolution image data, thus reducing the demand for computational resources. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram illustrating the principle of the surface reflectance fusion method based on the Gaussian model provided by the present invention.

[0039] Figure 2 This is a flowchart illustrating the surface reflectance fusion method based on the Gaussian model provided by the present invention.

[0040] Figure 3This is one of the schematic diagrams showing a quantitative comparison of the accuracy of the three spatiotemporal fusion methods provided by this invention.

[0041] Figure 4 This is a schematic diagram illustrating the temporal statistical differences in reflectivity and the differences between sensors provided by the present invention.

[0042] Figure 5 This is one of the schematic diagrams illustrating the prediction results of the three spatiotemporal fusion methods provided by this invention.

[0043] Figure 6 This is one of the near-infrared scatter plots of the three spatiotemporal fusion methods provided by this invention.

[0044] Figure 7 This is the second of three schematic diagrams illustrating the prediction results of the spatiotemporal fusion methods provided by this invention.

[0045] Figure 8 This is a visual comparison diagram of the GAUTF predicted time series results and the actual Landsat images provided by the present invention.

[0046] Figure 9 This is a schematic diagram comparing the quantitative accuracy of the long-term prediction capabilities of the three spatiotemporal fusion methods provided by this invention.

[0047] Figure 10 This is the third of the three spatiotemporal fusion methods provided by this invention, illustrating the prediction results.

[0048] Figure 11 This is the second schematic diagram showing a quantitative comparison of the accuracy of the three spatiotemporal fusion methods provided by this invention.

[0049] Figure 12 This is the fourth of three schematic diagrams illustrating the prediction results of the spatiotemporal fusion methods provided by this invention.

[0050] Figure 13 This is the third schematic diagram showing a quantitative comparison of the accuracy of the three spatiotemporal fusion methods provided by this invention.

[0051] Figure 14 This is a schematic diagram of the surface reflectance fusion device based on the Gaussian model provided by the present invention.

[0052] Figure 15 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0054] Multi-temporal image sequence analysis is of significant value in numerous remote sensing applications, such as vegetation / crop tracking, evapotranspiration estimation, atmospheric monitoring, land use / cover change detection, and detailed investigation of human-nature interactions. Changes on the Earth's surface often occur rapidly, requiring high temporal and spatial resolution for timely and accurate characterization. However, due to limitations in remote sensing sensors and unstable observation conditions, high-resolution Earth observation data with both spatial and temporal continuity are extremely limited. For example, commonly used medium-to-high spatial resolution open-source sensors, such as Landsat 8 OLI, Landsat 9 OLI, and Sentinel-2 MSI, provide revisit cycles of 16 days at 30-meter spatial resolution and 5 days at 10-60-meter spatial resolution, respectively. In contrast, satellite instruments with medium-to-low spatial resolution, including MODIS (250m, 500m, and 1000m), AVHRR (1100m), Sentinel-3 (300m), and OrbView-2 (1000m), can achieve higher temporal resolutions of approximately 1-2 days. Furthermore, cloud and cloud shadow contamination significantly limit the spatiotemporal availability of observations, thus hindering the practical application of remote sensing data. Fusion of high spatial resolution data with high temporal resolution data can combine their complementary characteristics to generate synthetic observational data with enhanced spatial detail and temporal continuity.

[0055] Over the past decade, spatio-temporal fusion (STF) has received widespread attention and spawned a variety of new methods, which can be mainly divided into two categories: learning-based methods and non-learning-based methods.

[0056] Process-based methods follow a strict sequence of operations, such as the traditionally defined weighted methods, originally proposed in the STARFM method. STARFM simplifies the fusion process by assigning residuals to weighted coarse-resolution data temporal variations, with weights focusing on spectral, temporal, and spatial differences within a moving window. Improved versions are designed to handle more complex land cover, such as the Spatiotemporally Adaptive Reflectance Variation Mapping Algorithm (STAARCH) and Enhanced STARFM (ESTARFM). These methods are efficient and simple, suitable as baseline methods, but they may not be able to account for temporal variations between different land cover types.

[0057] Similar to weight-based methods, unmixing-based methods typically follow a strict sequence of operations and are categorized as process-based methods. Unmixing-based methods originate from multi-sensor multi-resolution (MMT) techniques and employ unmixing methods such as least squares to address the mixed pixel problem. Directly using unmixing methods for STF (Spectral Transmission Forecasting) results in significant spectral prediction errors. To minimize these errors, subsequent unmixing-based methods have been developed. Unmixing-based data fusion (UBDF) introduces simple positive values ​​and appropriate range constraints during linear unmixing. STDFA (Standard Unmixing Data Fusion) first introduced the important concept of unmixing time variations rather than single-image variations, and added the estimated variations back to the reference high-resolution image to obtain the prediction result, which later inspired Fit-FC and FSDAF methods.

[0058] Statistical methods typically rely on Bayesian theory and other statistically based machine learning approaches. Bayesian methods address the fusion problem from a purely statistical perspective. Within the Bayesian framework, spatiotemporal fusion can be viewed as a maximum a posteriori (MAP) problem. They address the MAP problem by constructing spatial relationships (i.e., scale models) between coarse and fine data from the same date based on the point spread function (PSF). Another approach, establishing relationships between coarse-resolution images observed on different dates (called temporal models), is based on a joint Gaussian distribution. While providing theoretical rigor, this approach often lacks geographical context. Learning-based methods in spatiotemporal image fusion focus on learning nonlinear relationships or mappings between low-resolution and high-resolution images. These can be categorized into dictionary-based learning methods, traditional machine learning methods, and deep learning methods. Sparse representation / dictionary pair learning / compressed sensing methods establish correspondences between high- and low-resolution images based on structural similarity and utilize the learned structural similarity to predict fine-data pixel values ​​from coarse data. Similarly, traditional machine learning methods establish nonlinear mapping functions between coarse and fine data and use these correspondences to predict high-resolution data from another date. However, learning a mapping function based solely on spatial similarity is an ill-posed problem, making traditional learning-based methods potentially ineffective. Unlike traditional machine learning methods that learn mapping functions from coarse to fine data, learning-based methods are a fusion of implicit models and data-driven approaches. This is because they do not rely on explicit mathematical formulas to model the relationships between data points; instead, they learn these relationships directly from large datasets by training and optimizing parameters. The large number of parameters in deep learning networks allows them to be viewed as specialized methods for establishing relationships between two distributions.

[0059] Finally, hybrid methods combine the advantages of different methods. For example, the FSDAF algorithm integrates weight functions and linear unmixing methods. Although it achieves high accuracy, the process is usually too complicated to be used in practice.

[0060] The following is combined with Figures 1-15This invention describes the surface reflectance fusion method and apparatus based on the Gaussian model provided by the present invention.

[0061] Since the reflectance of a single land cover type typically exhibits a symmetrical clustered distribution, conforming to Gaussian distribution characteristics, this invention models the reflectance of a single land cover type as a Gaussian distribution. Furthermore, the central limit theorem provides theoretical support for the Gaussian model of reflectance. The Gaussian model can effectively analyze and verify reflectance variability; for example, the maximum likelihood classification method treats the reflectance of a single land cover type as a Gaussian distribution.

[0062] Based on the linear decomposition model:

[0063] ;

[0064] in, The pixel values ​​are from simulated low-resolution image data. Let be the reflectance of the i-th type of land cover. Let ε represent the area of ​​the i-th type of land cover in the target pixels of the low-resolution image data, and let ε be the residual error.

[0065] Will Treat them as independent random variables and assume that they follow an independent Gaussian distribution. The distribution is denoted as , The mean, Let Variance be the variance. Since a linear combination of independent Gaussian distributions still follows a Gaussian distribution, this linear combination... The distribution can be represented as:

[0066] ;

[0067] By determining the distribution of various land cover types, the distribution of mixed pixels can be calculated using this formula.

[0068] Due to factors such as observation angle, aerosol conditions, and topographic effects, the pixel observation values ​​of low-resolution image data with the same distribution characteristics vary. There may be discrepancies. However, by determining its distribution characteristics and combining them with measured data, it is possible to effectively infer... Its characteristics. This is because... The distribution characteristics are known. If its variance is large, it indicates that the reflectivity of the i-th type of land cover is more susceptible to changes in environmental or sensor observation conditions. Therefore, the model simulation values... Compared with pixel observations The significant deviations can be attributed to these factors. From a statistical perspective, this inference can be verified through conditional expectation, whose mathematical expression is as follows:

[0069] ;

[0070] at this time, Substituting into the above formula, we get:

[0071] ;

[0072] in, and Already in the formula We obtain the result from the middle equation, and substituting it into the above equation, we get:

[0073] ;

[0074] In the formula, Let be the reflectance of the i-th type of land cover.

[0075] Therefore, low-resolution image pixel observations can be used. , obtain the reflectance of the i-th type of land cover at any time. and It can be obtained by sampling low-resolution image data. Figure 1 A calculation example is shown.

[0076] Given that the reflectivity of a particular landscape is continuous over time, This can be adjusted using a Savitzky-Golay filter; as the prediction time increases, the variance of reflectance typically tends to increase during pure pixel sampling. Estimation can also be based on prior knowledge. In addition, some endmember extraction techniques can be used to reduce variance.

[0077] In existing technologies, Bayesian-based methods and other methods applying Bayesian rules assume that all coarse-resolution pixels follow a uniform distribution (e.g., a joint multivariate Gaussian distribution or a Gaussian distribution). Unlike existing technologies, this invention assumes that each coarse-resolution pixel follows an independent Gaussian distribution. Although these pixels are actually instances of a joint Gaussian distribution, treating them as independent distributions effectively simplifies the problem.

[0078] While the aforementioned theory can be directly applied to mixed pixel decomposition, its results may not be ideal because reflectance does not strictly follow a Gaussian distribution, and there are spectral differences between coarse and fine resolution data. If the application only focuses on coarse temporal variations without requiring extensive detail, the formula... It can still be used directly for decomposition. However, by drawing on the mechanism revealed by the STARFM model—that coarse and fine resolution data exhibit similar spectral differences over time—the alignment accuracy between the decomposition results and the fine resolution reflectance can be significantly improved. Methods such as STARFM and FSDAF are developed based on this principle. This invention also applies this principle and combines known fine pixel values ​​from a specific time phase to optimize the results; the derivation process is as follows.

[0079] Assuming the first moment With the second moment If they have the same error term, then it can be deduced that:

[0080] ;

[0081] because This linear combination also follows a Gaussian distribution, and based on the properties of the Gaussian distribution, its distribution expression can be derived as follows:

[0082] ;

[0083] According to the formula The same conditional expectation theory can be used to construct variables. The conditional expectation expression is as follows:

[0084] ;

[0085] At the same time, according to the covariance relationship We can obtain:

[0086] ;

[0087] Similarly, according to the formula The derivation principle is to use known observations , and the mean of the difference Substituting, we get The estimated value is:

[0088] ;

[0089] in, and They represent and Coarse pixel observations at time 1. Because and All follow a Gaussian distribution, therefore we have Substituting the values, we get:

[0090] ;

[0091] If we assume that the land cover type has not changed, that is Then the following equation can be derived:

[0092] This invention assumes that the land cover type has not changed.

[0093] For simplicity, the right side of the above equation is denoted as... . Through Known values ​​from high-resolution image data acquired at any given time can be obtained. Preliminary estimates: This algorithm takes into account both time-varying and local variation factors.

[0094] Based on the above analysis, the surface reflectance fusion method based on the Gaussian model of this invention is obtained, such as... Figure 2 As shown, steps S1, S2, S3 and S4 are included but are not limited to.

[0095] Step S1: Identify various types of land cover based on the first image data; the first image data is the first satellite image data at the first moment.

[0096] Step S2: Obtain the first reflectance of the first image data, the first observed value of the target pixel in the second image data, and the second observed value of the target pixel in the third image data; the second image data is the second satellite image data at the first time, and the third image data is the second satellite image data at the second time.

[0097] The spatial resolution of the second satellite imagery data is lower than that of the first satellite imagery data.

[0098] Because the spatial resolution of the second satellite imagery data is lower than that of the first satellite imagery data, the second satellite imagery data is considered low-spatial-resolution imagery data, while the first satellite imagery data is considered high-spatial-resolution imagery data. In other words, the first imagery data is high-spatial-resolution imagery data, while the second and third imagery data are low-spatial-resolution imagery data. Various types of land cover can be identified based on the high-resolution imagery data.

[0099] First reflectance of the first image data To pass the first moment The known value is obtained from the first image data. The target pixel can be any pixel in the low-resolution image data, i.e., any mixed pixel. The first observation value is... The second observation value is .

[0100] Step S3: Determine the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area ratio of various types of land cover in the target pixel.

[0101] The Gaussian distribution parameters include mean and variance. The mean reflectance of each pure pixel corresponding to the i-th type of land cover in the second image data is... variance is The mean reflectance of each pure pixel corresponding to the i-th type of land cover in the third image data is variance is The proportion of the first area of ​​the i-th type of land cover in the target pixel is .

[0102] Step S4: Based on the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameter, and the first area ratio, predict the second reflectance of the fourth image data; the fourth image data is the first satellite image data at the second time point.

[0103] The fourth image data is the high-resolution satellite image data from the second time point. Step S4 may further include:

[0104] Predict the change in reflectance based on the first observation, the second observation, the Gaussian distribution parameters, and the first area proportion;

[0105] Adding the change in reflectance to the first reflectance yields the second reflectance.

[0106] The formula for calculating the second reflectivity is as follows:

[0107] ; The second reflectivity is given by the change in reflectivity. .

[0108] As can be seen from the above, the surface reflectance fusion method based on the Gaussian model of the present invention assumes that the reflectance of the surface cover follows an independent Gaussian distribution. Then, it predicts the reflectance of the high-resolution image data through the Gaussian distribution parameters, one high-resolution image data and two low-resolution image data. The spatiotemporal fusion of satellite image data does not rely on empirical and pixel-by-pixel filtering, thus improving computational efficiency. The requirement for computational resources is low with only one high-resolution image data, thus reducing the demand for computational resources.

[0109] In one embodiment, step S1 of the present invention may specifically include:

[0110] The first image data is classified or clustered to obtain a classification / clustering map;

[0111] Identify different types of land cover based on classification / clustering diagrams.

[0112] In sampling the Gaussian distribution parameters of land cover reflectance, the number of clusters in the first image data is crucial, and the number of clusters must be balanced to ensure that the signal approximates a Gaussian distribution. If the number of clusters is too small, the mixed categories will be difficult to conform to a Gaussian distribution; if the number of clusters is too large, it may result in an insufficient number of pure pixels, making it impossible to accurately model the Gaussian distribution of a specific land cover category. This invention can use software in an integrated development environment (IDE) (such as ENVI) for clustering, as it allows for more convenient control over whether pure pixels in the second time step (coarse resolution data) are determined by searching for pixels containing only a single land cover category within their pixel boundaries; such pixels are considered pure pixels.

[0113] In one embodiment, the present invention may determine each pure pixel corresponding to a target land cover type, including:

[0114] Obtain the second area percentage of the target land cover type in any pixel of the second or third image data;

[0115] If the second area ratio reaches the preset ratio, then any pixel is determined as the pure pixel corresponding to the target land cover type.

[0116] The Gaussian distribution parameters corresponding to each land cover type need to be obtained by sampling the reflectance of each pure pixel corresponding to that land cover type. Therefore, this invention needs to first determine each pure pixel corresponding to each land cover type in the low-resolution image data.

[0117] The preset ratio can be 95%. If the area of ​​a certain land cover type in a certain pixel reaches 95%, the pixel can be considered as a pure pixel corresponding to that land cover type.

[0118] This allows us to determine the pure pixels corresponding to each land cover type in low-resolution image data by measuring the area proportion of each land cover type, in order to determine the Gaussian distribution parameters of reflectance.

[0119] In one embodiment, if the second area proportion reaches a preset ratio, after determining any pixel as a pure pixel corresponding to the target land cover type, the method may further include:

[0120] If the number of pure pixels corresponding to the target land cover type exceeds the first threshold, the preset ratio is increased, and each pure pixel corresponding to the target land cover type is re-determined according to the new preset ratio until the number of pure pixels corresponding to the target land cover type is less than the first threshold.

[0121] If the number of pure pixels corresponding to the target land cover type is less than the second threshold, the preset ratio is gradually reduced, and each pure pixel corresponding to the target land cover type is re-determined according to the new preset ratio until the number of pure pixels corresponding to the target land cover type reaches the second threshold.

[0122] The first threshold is greater than the second threshold.

[0123] If the number of pure pixels corresponding to a certain land cover type exceeds the first threshold, it means that there are too many pure pixels corresponding to that land cover type, the preset ratio setting is unreasonable, and impure pixels will be introduced.

[0124] When the number of pure pixels corresponding to a certain land cover type is too large, the present invention can increase the preset ratio, such as to 100%, so that the number of pixels with the second area ratio reaching the preset ratio will decrease, that is, the number of pure pixels will decrease, which can avoid the introduction of impure pixels.

[0125] According to the Central Limit Theorem, it is generally recommended that the sample size be no less than 30, so the second threshold can be 30. If the number of pure pixels corresponding to a certain land cover type is less than 30, it means that the number of pure pixels corresponding to that land cover type is insufficient, the sample size cannot meet the requirements for constructing a Gaussian distribution, and the preset ratio setting is unreasonable.

[0126] When the number of pure pixels corresponding to a certain land cover type is too small, the present invention can gradually reduce the preset ratio, such as by gradually reducing the preset ratio in steps of 5%, until the number of pure pixels corresponding to the land cover type reaches the second threshold, which can ensure that the number of pure pixel samples meets the requirements for constructing a Gaussian distribution.

[0127] In one embodiment, the present invention determines the Gaussian distribution parameters of reflectance for each pure pixel corresponding to various types of land cover, including:

[0128] The mean and variance of reflectance of each pure pixel corresponding to a single land cover type are determined to obtain the Gaussian distribution parameters corresponding to a single land cover type.

[0129] In other words, this invention determines the Gaussian distribution parameters corresponding to each type of land cover.

[0130] Although a Gaussian distribution can be used to approximate the reflectance distribution of a specific land cover type, the actual distribution may not strictly conform to the Gaussian property, in which case the second reflectance predicted according to step S4 will be inaccurate. Therefore, in one embodiment, after step S4, the land reflectance fusion method based on the Gaussian model of the present invention may further include:

[0131] The confidence interval range of the second reflectivity is determined based on the Gaussian distribution parameters corresponding to the third image data and the first area proportion.

[0132] If the second reflectance is outside the confidence interval, the second reflectance is re-predicted based on the first reflectance, the first observation, and the second observation.

[0133] This invention can use a 95% confidence level statistical test to determine whether the surface cover type of coarse pixels can be constituted by a linear combination of Gaussian distributions. Based on the characteristics of the standard normal distribution, the corresponding Z-value at the 95% confidence level is 1.96.

[0134] like Assuming that the land cover type remains unchanged, based on the 95% confidence level requirement of the Gaussian distribution, we can obtain:

[0135] ;

[0136] Based on this, the range of the confidence interval can be derived:

[0137] ;

[0138] Changes in land cover types were also included in this estimate, as they should not be included in the 95% confidence level at the second time point.

[0139] If the second reflectance is outside the confidence interval, the predicted target pixel is an uncertain pixel, and the second reflectance can be estimated using the following formula:

[0140] ;

[0141] Furthermore, the introduction of high-resolution data can improve the performance of uncertain pixels. The estimate can be achieved using the following formula: This operation needs to be performed independently for each band: uncertain pixels in the near-infrared band may not be uncertain pixels in the red band. In this case, the red band will be predicted using an RC plot, while the near-infrared band will be predicted using a formula. Fill in the gaps.

[0142] This invention compares the existing technology with the present invention through experiments.

[0143] I. Study Area and Dataset

[0144] This invention uses two datasets. The first dataset is the Landsat-MODIS paired dataset (Emelyanova et al. 2013), which is widely used in academia as a benchmark. The second dataset contains a benchmark dataset of Sentinel-2 (S2) and Sentinel-3 (S3) images (Boumahdi et al. 2023), which is characterized by greater resolution differences and finer data that provides higher spatial resolution.

[0145] 1. First stop

[0146] The dataset at this site is frequently used to test the effectiveness of the Spatiotemporal Fusion (STF) method in predicting fine-resolution data in heterogeneous landscapes with significant phenological variations. The dataset contains 16 pairs of Landsat 7 ETM+ and MODIS09GA imagery, acquired from October 2001 to May 2002, covering an area of ​​approximately 20 km × 20 km (800 × 800 pixels, spatial resolution 25 meters). The first site exhibits significant landscape heterogeneity, containing numerous small paddy fields and woodland patches. Significant temporal variations have occurred in this area, including phenological changes in vegetation and land use cover changes caused by agricultural activities.

[0147] 2. Sentinel-2 / 3 (S2 & S3) dataset

[0148] This invention utilizes the Waterband dataset from the Sentinel-2 / 3 (S2 & S3) dataset. The dataset uses S2 level 1C and S3 OLCI level 1B data. The S2 and S3 data underwent atmospheric correction and reprojection to ensure projection matching between S2 and S3. The second site is a rural site. This region exhibits rapid temporal changes due to forest fires. It is known for its high frequency of early-season fires, resulting in varying phenological changes in vegetation throughout the year. The data coverage is 18km × 18km (1800 × 1800 Sentinel-2 pixels). This invention uses bands 2 (blue light), 3 (green light), 4 (red light), and 8 (near-infrared) of S2, respectively, to fuse with bands Oa4, Oa6, Oa8, and Oa17 of S3. These correspond to bands 1, 2, 3, and 7 of the S2 data and bands 4, 6, 8, and 13 of the S3 data in the second site dataset.

[0149] II. Comparison and Evaluation Indicators

[0150] Since this invention uses only a single-temporal high-resolution dataset and does not compare it with learning-based fusion methods, such methods typically require high-resolution time-series data. This invention chooses to compare with STARFM and FSDAF methods, which also only require single-temporal high-resolution data. Qualitative evaluation is performed through visual comparison between false-color synthetic images and real-predicted images, and scatter plots are used for qualitative analysis of the fusion results. For quantitative evaluation, the following metrics are used: Root Mean Square Error (RMSE) measures the difference between predicted and true reflectance; Correlation Coefficient (CC) characterizes the linear correlation between predicted and true reflectance; Absolute Difference (AD) is the average of the absolute errors between predicted and true pixel values.

[0151] ;

[0152] In addition to the quantitative assessment of spectral similarity mentioned above, this invention also employs Local Binary Pattern (LBP) and Robert's Edge detection for spatial similarity testing. The calculation formula for Robert's Edge detection is as follows:

[0153] ;

[0154] Among them, D i,j This represents the grayscale value of the pixel in the i-th row and j-th column. The edge detection value (Edge) ranges from [-1, 1]: a value of 0 indicates that the fused image has ideal edge preservation; the closer the value is to negative infinity, the more smooth the edge features of the fused image are; the closer the value is to positive infinity, the more sharp the edge features of the fused image are. The calculation formula for Local Binary Pattern (LBP) is as follows:

[0155] ;

[0156] Where Di represents the grayscale values ​​of all pixels surrounding the center pixel in the 3×3 moving window; Dc represents the grayscale value of the center pixel. decimal represents the operation of converting a binary number to decimal. The LBP value ranges from [-1, 1]: a value of 0 indicates that the fused image has an ideal texture preservation effect; the closer the value is to negative infinity, the more the texture features of the fused image are over-smoothed; the closer the value is to positive infinity, the more the texture features of the fused image are over-sharpened.

[0157] III. Experimental Results

[0158] 1. MODIS-Landsat images for experiments at the first site

[0159] The hardware configuration used in this invention is as follows: Intel(R) Core(TM) i9-10850K CPU @ 3.60GHz processor, 64.0 GB DDR4 memory, 1TB NVMe solid-state drive, and Windows 10 Professional system environment. All speed comparison tests were run without GPU acceleration. The software development environment used was PyCharm Community Edition 2021.3.2 integrated development environment, with Python 3.10 as the interpreter.

[0160] Clustering maps were obtained using the K-means clustering algorithm in the ENVI Classic unsupervised classification tool, dividing the entire image into 3 categories in the first site dataset. The FSDAF source code uses the IsoData method for prediction (minimum number of categories set to 5, maximum number of categories set to 6). The FSDAF parameters were set as follows: pure pixel count 360, block size 30, window size and similar pixel count both 20.

[0161] The average running time of each algorithm on the first site dataset is as follows: GAUSTF approximately 63 seconds, FSDAF approximately 1660 seconds, STARFM approximately 550 seconds, and Fit-FC approximately 2560 seconds (calculated for the entire image). Due to the large missing area in the lower right corner of the original data, this paper only shows the results for a sub-region.

[0162] 1 / 1, Comprehensive Performance Comparison of Single-Item Pairing

[0163] This invention selects data from November 25, 2001, to predict two time phases: January 12, 2002, and March 10, 2002. Based on the imagery from November 25, 2001, the data is divided into three spectral categories. Only real MODIS imagery is used for pairing in the test. Some original MODIS data exhibits incomplete registration; therefore, the bilinearly resampled data is directly input into the algorithm of this invention (this algorithm requires coarse-resolution data to maintain its original resolution; the first site dataset is 500 meters). The resampling operation is performed using the bilinear resampling method in ArcGIS. To meet the input requirements of the STARFM algorithm (requiring absolute registration between MODIS data pairs), the mode (dominant) method in Python is used to resample the 500-meter MODIS data to a 25-meter resolution. FSDAF directly uses the original 25-meter resampled MODIS data provided in the first site dataset.

[0164] A quantitative comparison of the accuracy of three spatiotemporal fusion (STF) methods, for example. Figure 3 The figure shows the accuracy evaluation results of three data fusion methods when predicting two target dates using data from November 25, 2001, for the first site study area. The numerical units in the figure are reflectance (RMSE = root mean square error, CC = correlation coefficient, AD = mean absolute difference from true reflectance, Edge = Robert's edge detection index, LBP = local binary mode). Overall, FSDAF achieved the best RMSE in all bands for both prediction dates, and also obtained the best CC value in the vast majority of bands. The study found that in predictions for more recent dates (January 12, 2002), the CC value and RMSE obtained by this invention are close to the performance of FSDAF. This invention achieves the best performance in the absolute difference (AD) index, indicating that its spectral prediction ability is close to FSDAF and superior to STARFM. Figure 6 A scatter plot showing the near-infrared band prediction results is presented. (a)-(c) represent the prediction results on January 12, 2002; (d)-(f) represent the prediction results on March 10, 2002. The scatter plot shows that FSDAF's prediction results are more concentrated in the near-infrared band. The near-infrared band RMSE of this invention (GAUSTF) is 0.681, which is better than STARFM, but its prediction ability for heterogeneous regions with rapidly changing reflectance is weaker than FSDAF (FSDAF's near-infrared band RMSE is 0.0624).

[0165] The final prediction results of all methods are mainly affected by two key factors: the time-varying reflectivity difference and the differences between different sensors, such as... Figure 4 As shown.

[0166] The differences between the two MODIS data from different time phases reflect the spectral variation of reflectance over time. It can be seen that the near-infrared band exhibits an extremely low correlation coefficient (CC) and an extremely high root mean square error (RMSE) compared to other bands, indicating a significant phenological change between the input image date and the predicted date. Furthermore, the differences in the images... It can be observed that the consistency difference between the two sensors in the near-infrared band is not significant. However, all three methods (GAUSTF, STARFM, and FSDAF) exhibit low prediction performance in the near-infrared band, indicating that the temporal change in reflectance (Δy) has a far greater impact on the final prediction result than the differences in sensor systems. This phenomenon is reasonable because all comparative methods rely on Δy (i.e., the change in reflectance over time) in the time-dimensional reflectance prediction rather than directly using the original observation value at time t2. .

[0167] The prediction results for January 12, 2002 using the STARFM, FSDAF, and GAUSTF methods are as follows: Figure 5 As shown, the original Landsat images of the heterogeneous region acquired at times t1 and t2 are (a) and (b), respectively. Sub-image 1 is located within the upper yellow rectangular area of ​​the entire image; sub-image 2 is located within the lower yellow rectangular area of ​​the entire image.

[0168] from Figure 5It can be seen that STARFM exhibits a significant blockiness effect. FSDAF, on the other hand, produces over-smoothing results, which is reflected in its relatively poor performance in the LBP and Edge indices. GAUSTF performs excellently in the Edge index, indicating its good prediction effect on edge features, and its LBP index also performs reasonably well. This is due to GAUSTF's smoothing prediction of the RC mapping; when this RC value is superimposed on the Landsat data at time t1, it helps to better preserve the original edge features of land cover.

[0169] The prediction results for March 10, 2002 using the STARFM, FSDAF, and GAUSTF methods are as follows: Figure 7 As shown, the original Landsat images of the heterogeneous region acquired at times t1 and t2 are (a) and (b), respectively. Sub-image 1 is located in the upper yellow rectangular area of ​​the entire image; sub-image 2 is located in the lower yellow rectangular area of ​​the entire image.

[0170] Figure 7 This paper presents a visual comparison of three spatiotemporal fusion (STF) methods using data from March 10, 2002. The sub-region comparison shows that GAUSTF demonstrates superior predictive ability for changed ground features. In sub-image 1, features that appeared as bright white in the original input image had transformed into bare land by the prediction date. Only STARFM correctly predicted the transition from white to bare land, primarily due to its high prediction accuracy in the green and red bands. Within the yellow rectangular area, the original input image showed bare land surrounded by flooded areas, which had transformed into growing crops by the prediction date. STARFM, FSDAF, and GAUSTF all successfully predicted this change. FSDAF generated a brighter prediction, closer to the actual image, mainly due to its accurate prediction in the near-infrared band.

[0171] Within the yellow rectangular area of ​​sub-image 2, GAUSTF detected the change from bright white to bare ground, while STARFM and FSDAF failed to predict this change. Visual comparison clearly shows that the three methods generally produced similar results. STARFM exhibits a significant blockiness effect, but performs exceptionally well in the green and red bands; FSDAF performs best in detecting phenological changes, particularly benefiting from its excellent near-infrared predictive capabilities; GAUSTF achieved an acceptable moderate result overall.

[0172] In the quantitative assessment, this invention noted that STARFM performed well and was close to FSDAF in most bands, but its results in the near-infrared band were not ideal. STARFM performed well in long-term predictions because it did not use clustering diagrams during the process, unlike FSDAF and GAUSTF. Long-term predictions are more likely to involve plot changes and therefore may be more sensitive to clustering diagrams. This invention found that GAUSTF produced better overall results in the prediction for March 10, 2002, when using a 4-class clustering diagram (currently using a 3-class clustering diagram). However, similar to STARFM, its results in the near-infrared band remained unsatisfactory.

[0173] The near-infrared scatter plot for the second predicted date (March 10, 2002) is shown below. Figure 6 As shown in (d)-(f) in the figure. The result of FSDAF is closest to the 1:1 line and has the best convergence, while the performance of GAUTF and STARFM seems similar, but GAUTF is better than STARFM in both RMSE and CC in the near-infrared band. However, FSDAF still achieves the best results overall in terms of RMSE and CC. In the first three bands, GAUTF performs worse than STARFM, mainly because the RMSE of Δy is relatively smaller in these bands than in other bands, giving STARFM an advantage (see Figure 1). Figure 3 However, GAUSTF outperforms STARFM in the latter four bands. Overall, FSDAF is the best choice of the three methods, demonstrating the strongest spectral prediction capability.

[0174] However, FSDAF does not perform best across all metrics, such as AD, Edge, and LBP. Its average RMSE is also quite close to that of GAUSTF. Users can choose based on their needs: if a balanced performance and faster speed are preferred, GAUSTF is a better choice; if superior spectral prediction capabilities are desired and slower speed is not a concern, then FSDAF is the better option.

[0175] 1 / 2, Evaluation of Time Series Prediction Results

[0176] To test GAUTF's ability to recover reflectance time series, this invention used November 25, 2001 as a baseline and made predictions for all subsequent dates, focusing on evaluating its spectral prediction capabilities. Visual comparison with real imagery is also included. Figure 8As shown, (a)-(e) are Landsat images of the time series for December 1, 2001, January 5, 2002, February 13, 2002, February 22, 2002, and May 17, 2002, respectively; (f)-(j) are the time series results predicted by GAUTF on the same dates; (k)-(o) are Landsat images of the time series for April 2, 2002, April 11, 2002, April 18, 2002, April 27, 2002, and May 4, 2002, respectively; and (p)-(t) are the time series results predicted by GAUTF on the same dates.

[0177] Visual comparisons of the time series predicted by GAUSTF demonstrate that this method successfully predicted the overall trend of crop phenological changes. Detailed visual comparisons of STARFM and FSDAF results are provided in the appendix. Quantitative comparisons of the three methods are also included. Figure 9 As shown, the root mean square error (RMSE) between the prediction results of each method and the actual Landsat image on the corresponding date is as follows: (a) Blue band (b) Green band (c) Red band (d) Near-infrared band (e) Shortwave infrared 1 band (f) Shortwave infrared 2 band.

[0178] from Figure 9 As can be seen, in most cases, GAUSTF achieves moderate performance between STARFM and FSDAF. In the near-infrared band, GAUSTF performs comparably to FSDAF and is significantly better than the STARFM method. Overall, while GAUSTF's long-term reflectance prediction capability is not as good as FSDAF, it is better than STARFM.

[0179] 2. Sentinel-2 & Sentinel-3 dataset experiments

[0180] To test the performance of GAUTF in high-resolution fusion, this invention uses Sentinel-2 and Sentinel-3 data to conduct a fusion experiment on the Waterbank site. The input date was May 14, 2019, and the prediction target dates were June 18, 2019, and September 11, 2019. GAUTF uses a 3-class clustering map generated based on K-means clustering in ENVI Classic; FSDAF uses the IsoData method for 5-class prediction, with the following parameters set: 855 pure pixels, block size 20, window size, and number of similar pixels both 20.

[0181] The original Landsat data in the dataset was upsampled to a 10-meter resolution. To maintain the accuracy of the pixel values, this invention uses the mode method to resample it to a 30-meter resolution before inputting it into GAUTF. STARFM and FSDAF, on the other hand, directly use the original Landsat data in the dataset. The average running time of each method at the Waterbank site is: approximately 20 seconds for GAUTF, approximately 1280 seconds for FSDAF, and approximately 315 seconds for STARFM.

[0182] The first visual pair of predicted dates Figure 10 As shown, the original Landsat images acquired at the Waterbank site at times T1 and T2 are (a) and (b), respectively, and (f)-(g) are magnified views of the green rectangular areas in the entire image. All three methods can predict the overall changes over the entire area well, but there are differences in detail: FSDAF and GAUSTF have similar visual prediction results, but FSDAF predicts detailed phenological changes better (especially in the central circular area); STARFM exhibits a block effect, and the predicted phenological changes within the circular area spill over to the outside (as shown in pixel S3). Quantitative comparison: Figure 11 As shown, the numerical unit is reflectance (RMSE = root mean square error, CC = correlation coefficient, AD = mean absolute difference from true reflectance, Edge = Robert's edge detection index, LBP = local binary mode).

[0183] The result of the second prediction date is as follows Figure 12 As shown, the original Landsat images acquired at the Waterbank site at times T1 and T2 are (a) and (b), respectively, and (f)-(g) are magnified views of the green rectangular areas within the entire image. These dates are four months apart from the input dates, but no significant land cover or phenological changes occurred in this area. Visual comparison reveals that the prediction results of the three methods are generally similar, but STARFM's prediction is sharper than FSDAF and GAUSTF. In the magnified areas, the results of GAUSTF and FSDAF show high similarity.

[0184] according to Figure 11 The results show that GAUSTF achieved the highest correlation coefficient (CC) values ​​across all bands, with its root mean square error (RMSE) being roughly on par with FSDAF. In terms of mean absolute difference (AD), the three methods performed similarly. STARFM, due to the block effect, performed best in edge detection; while GAUSTF performed best in Local Binary Pattern (LBP), preserving richer texture details. Quantitative comparison results for magnified areas are shown below. Figure 11In the second half: FSDAF achieved the best RMSE values ​​in the first three bands, while GAUSTF achieved the best RMSE performance in the near-infrared band. STARFM had better AD values ​​in the green and red bands, but overall the AD values ​​of the three methods were quite similar. Furthermore, GAUSTF and FSDAF had similar CC values ​​and both outperformed STARFM. Within the sub-regions, GAUSTF also maintained the best LBP performance.

[0185] Accuracy evaluation of three data fusion methods applied to the Waterbank research site (based on predictions for September 11, 2019, using data from May 14, 2019). Figure 13 As shown. The numerical unit is reflectance (RMSE = root mean square error, CC = correlation coefficient, AD = mean absolute difference from true reflectance, Edge = Robert's edge detection index, LBP = local binary mode).

[0186] according to Figure 13 The results show that GAUSTF achieved the highest correlation coefficient (CC) values ​​in the blue, green, and red bands. While FSDAF maintained the lowest root mean square error (RMSE), its value was essentially on par with GAUSTF. In terms of mean absolute difference (AD), the three methods performed similarly. GAUSTF demonstrated performance comparable to FSDAF in spectral prediction, and both outperformed STARFM. STARFM showed the best performance in edge detection due to blockiness, while GAUSTF achieved the best local binary mode (LBP) performance across the entire image and magnified areas, indicating its superior texture preservation capabilities.

[0187] Through three experiments, this invention demonstrates that GAUTF's overall performance in spectral prediction is close to FSDAF and superior to STARFM, while generating fusion results with richer texture details. In the first experiment, facing regions with high land cover heterogeneity and significant phenological changes over time, GAUTF achieved a moderate accuracy between FSDAF and STARFM in time-series spectral prediction. In the second experiment, even with large-scale land cover changes and without a change detection mechanism, the current version of GAUTF still exhibits effective recovery capabilities for changed land cover. In the third experiment, when using Sentinel-2 and Sentinel-3 data as input for prediction, GAUTF achieved the best spectral prediction results on the first prediction date, indicating its better applicability in higher-resolution data fusion. GAUTF's most significant advantage lies in its computational efficiency. It runs approximately 26-64 times faster than FSDAF and approximately 8-15 times faster than STARFM on the three datasets. For fusion tasks involving fine-resolution pixels of approximately 3000×3000 and coarse-resolution pixels of 120×120, the processing time can be reduced to around 120 seconds, demonstrating strong engineering practicality. If users need to balance accuracy and efficiency in spatiotemporal fusion applications, GAUTF is undoubtedly one of the best choices currently available.

[0188] like Figure 14 As shown, the surface reflectance fusion device based on the Gaussian model provided by this invention includes, but is not limited to:

[0189] The identification module is used to identify various types of land cover based on the first image data; the first image data is the first satellite image data at the first moment.

[0190] The acquisition module is used to acquire the first reflectance of the first image data, the first observed value of the target pixel in the second image data, and the second observed value of the target pixel in the third image data; the second image data is the second satellite image data at the first moment, and the third image data is the second satellite image data at the second moment;

[0191] The determination module is used to determine the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area proportion of various types of land cover in the target pixel.

[0192] The prediction module is used to predict the second reflectance of the fourth image data based on the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameter, and the first area proportion; the fourth image data is the first satellite image data at the second time point.

[0193] The spatial resolution of the second satellite imagery data is lower than that of the first satellite imagery data.

[0194] It should be noted that the surface reflectance fusion device based on the Gaussian model provided by the present invention can execute the surface reflectance fusion method based on the Gaussian model described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0195] Figure 15 This is a schematic diagram of the electronic device provided by the present invention. The electronic device may include: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. The processor can call logical instructions in the memory to execute a land surface reflectance fusion method based on a Gaussian model. The method includes: identifying various types of land cover based on first image data; obtaining the first reflectance of the first image data, the first observation value of the target pixel in the second image data, and the second observation value of the target pixel in the third image data; determining the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area proportion of various types of land cover in the target pixel; and predicting the second reflectance of the fourth image data based on the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameters, and the first area proportion.

[0196] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0197] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, when the program instructions are executed by a computer, the computer is able to execute the land surface reflectance fusion method based on the Gaussian model provided in the above embodiments, the method including: identifying various types of land cover according to first image data; obtaining a first reflectance of the first image data, a first observation value of a target pixel in the second image data, and a second observation value of a target pixel in the third image data; determining the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to various types of land cover in the third image data, and a first area proportion of various types of land cover in the target pixel; predicting the second reflectance of the fourth image data according to the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameters, and the first area proportion.

[0198] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the Gaussian model-based land reflectance fusion method provided in the above embodiments. The method includes: identifying various types of land cover based on first image data; acquiring a first reflectance of the first image data, a first observation value of a target pixel in second image data, and a second observation value of a target pixel in third image data; determining the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of the reflectance of each pure pixel corresponding to various types of land cover in the third image data, and a first area proportion of each type of land cover in the target pixel; and predicting the second reflectance of a fourth image data based on the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameters, and the first area proportion.

[0199] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0200] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A Gaussian model-based surface reflectance fusion method, characterized in that, include: Identify various types of land cover based on the first image data; The first image data is the first satellite image data at the first moment; The first reflectance of the first image data, the first observed value of the target pixel in the second image data, and the second observed value of the target pixel in the third image data are obtained; the second image data is the second satellite image data at the first time, and the third image data is the second satellite image data at the second time. The Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area proportion of various types of land cover in the target pixel are determined respectively. Based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameter, and the first area ratio, the second reflectance of the fourth image data is predicted; the fourth image data is the first satellite image data at the second time point. The spatial resolution of the second satellite image data is lower than that of the first satellite image data. The Gaussian distribution parameters include mean and variance. Based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameters, and the first area proportion, the second reflectance of the fourth image data is predicted, including: ； wherein, is the second reflectivity, is the first time, is the second time, is the first reflectivity, , are respectively the mean and the variance of the reflectivity of each pure pixel corresponding to the i-th type of surface cover in the second image data, , are respectively the mean and the variance of the reflectivity of each pure pixel corresponding to the i-th type of surface cover in the third image data, is the first area ratio of the i-th type of surface cover in the target pixel, is the first observation value, is the second observation value.

2. The surface reflectance fusion method based on the Gaussian model according to claim 1, characterized in that, The identification of various land cover types based on the first image data includes: The first image data is classified or clustered to obtain a classification / clustering map; Identify different types of land cover based on the classification / clustering diagram.

3. The surface reflectance fusion method based on the Gaussian model according to claim 1, characterized in that, Identify the individual pure pixels corresponding to the target land cover type, including: Obtain the second area percentage of the target land cover type in any pixel of the second image data or the third image data; If the second area ratio reaches a preset ratio, then any pixel is determined to be a pure pixel corresponding to the target land cover type.

4. The surface reflectance fusion method based on the Gaussian model according to claim 3, characterized in that, If the second area ratio reaches a preset ratio, then after determining that any pixel is a pure pixel corresponding to the target land cover type, the method further includes: If the number of pure pixels corresponding to the target land cover type exceeds the first threshold, the preset ratio is increased, and each pure pixel corresponding to the target land cover type is re-determined according to the new preset ratio, until the number of pure pixels corresponding to the target land cover type is less than the first threshold. If the number of pure pixels corresponding to the target land cover type is less than the second threshold, the preset ratio is gradually reduced, and each pure pixel corresponding to the target land cover type is re-determined according to the new preset ratio, until the number of pure pixels corresponding to the target land cover type reaches the second threshold. Wherein, the first threshold is greater than the second threshold.

5. The surface reflectance fusion method based on the Gaussian model according to claim 1, characterized in that, Determine the Gaussian distribution parameters of reflectance for each pure pixel corresponding to various types of land cover, including: The mean and variance of the reflectance of each pure pixel corresponding to a single land cover type are determined to obtain the Gaussian distribution parameters corresponding to the single land cover type.

6. The surface reflectance fusion method based on the Gaussian model according to claim 1, characterized in that, After predicting the second reflectance of the fourth image data based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameter, and the first area proportion, the method further includes: The confidence interval range of the second reflectance is determined based on the Gaussian distribution parameters corresponding to the third image data and the first area ratio; If the second reflectance is outside the confidence interval, then the second reflectance is re-predicted based on the first reflectance, the first observation, and the second observation.

7. A surface reflectance fusion device based on a Gaussian model, characterized in that, include: The identification module is used to identify various types of land cover based on the first image data; The first image data is the first satellite image data at the first moment; The acquisition module is used to acquire the first reflectance of the first image data, the first observed value of the target pixel in the second image data, and the second observed value of the target pixel in the third image data; the second image data is the second satellite image data at the first time, and the third image data is the second satellite image data at the second time. The determination module is used to determine the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the second image data, the Gaussian distribution parameters of reflectance of each pure pixel corresponding to various types of land cover in the third image data, and the first area proportion of various types of land cover in the target pixel. The prediction module is used to predict the second reflectance of the fourth image data based on the first reflectance, the first observation value, the second observation value, the Gaussian distribution parameter, and the first area ratio; the fourth image data is the first satellite image data at the second time point; The spatial resolution of the second satellite image data is lower than that of the first satellite image data. The Gaussian distribution parameters include mean and variance. Based on the first reflectance, the first observed value, the second observed value, the Gaussian distribution parameters, and the first area proportion, the second reflectance of the fourth image data is predicted, including: ； in, The second reflectivity, For the first moment, For the second moment, The first reflectivity, , These are the mean and variance of the reflectance of each pure pixel corresponding to the i-th type of land cover in the second image data, respectively. , Let be the mean and variance of the reflectance of each pure pixel corresponding to the i-th type of land cover in the third image data. The first area percentage of the i-th type of land cover in the target pixel. The first observation value, This is the second observation value.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the surface reflectance fusion method based on the Gaussian model as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the surface reflectance fusion method based on the Gaussian model as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the surface reflectance fusion method based on the Gaussian model as described in any one of claims 1 to 6.

Images