Sentinel-MODIS day-by-day reflectivity fusion method based on sparse Bayesian learning

Abstract

The invention discloses a Sentinel-MODIS day-by-day reflectivity fusion method based on sparse Bayesian learning, and the method comprises the steps: taking MODIS day-by-day observation and a Sentinel high-resolution image as the basis, firstly carrying out the image pairing and spectral response consistency correction, constructing a multiband reflectivity historical sample set, and carrying out the spectral response consistency correction; under a sparse Bayesian learning framework, a target is screened through an adaptive sliding window, and automatic evaluation and fitting of variable contribution degree are realized by utilizing marginal likelihood maximization, so that the robustness of a prediction result is ensured at a probability modeling level. According to the method, day-by-day time continuity, fine description of 10-meter spatial resolution and consistent reconstruction of multi-band reflectivity can be realized, and the prediction precision is improved through residual correction and spatial smoothing. According to the invention, the limitation of the resolution ratio and the synthesis period of a traditional product is broken through, and the monitoring capability of rapid hydrology and vegetation processes in high-dynamic ecological environments such as lake wetlands and the like is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing data fusion and high-resolution spatio-temporal reconstruction, and particularly relates to a Sentinel-MODIS daily 10-meter reflectance fusion method based on sparse Bayesian learning, which is used to realize the generation of daily continuous, high spatial resolution and multi-band surface reflectance data. BACKGROUND

[0002] There has been a long-standing contradiction between the temporal resolution and spatial resolution of remote sensing data. Sentinel-2 satellite has 10-meter spatial resolution and 13 spectral bands, which can capture fine features and multi-element information of the ground. However, its revisit period is usually 5-10 days, which is limited by orbit and cloud cover, and it is difficult to form a continuous daily time series. In contrast, the MODIS sensor has an observation frequency of 1-2 days and can provide more than 20 years of global coverage time series data, but the spatial resolution is only 250-500 meters, which cannot meet the needs of fine ecological process research.

[0003] To balance the time and space advantages, the prior art combines Landsat and MODIS data by time proximity and spectral similarity to generate time series images at 30-meter resolution, but this technology generally relies on MODIS NDVI or EVI index products, and these products are mostly 8-day or 16-day composites, not daily observations. Therefore, although the high temporal resolution advantage of MODIS is utilized, the generated high-resolution time series images still have intervals, making it difficult to form a truly daily continuous sequence.

[0004] In addition, most of the current researches in the prior art focus on the time series fusion of a single vegetation index, ignoring the value of multi-band data in multi-element observation, so there is an urgent need for a method to realize daily-scale multi-band consistency reconstruction.

[0005] On the other hand, existing sparse modeling methods such as LASSO have been applied in image fusion, but they rely on manually set regularization parameters and are easily affected by noise and data redundancy, making it difficult to ensure robustness.

[0006] Therefore, the existing methods cannot simultaneously meet the comprehensive needs of 10-meter spatial resolution, daily continuity and multi-band consistency. Developing a new fusion method based on sparse Bayesian learning can combine the high spatial resolution of Sentinel with the daily observation advantage of MODIS, and can realize high-precision dynamic monitoring of multi-elements in high-dynamic ecological systems such as lakes and wetlands, providing key data support for hydrological process simulation, ecological assessment and carbon cycle research.

[0007] Sparse Bayesian Learning (SBL) automatically adjusts the importance of variables by maximizing marginal likelihood, thus preserving effective information and reducing redundant information in a probabilistic sense. Its advantage lies in maintaining stability and consistency across multi-band, high-dimensional tasks. Summary of the Invention

[0008] Invention aims / technical problems This invention addresses the technical problems existing in the prior art by proposing a Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning. By introducing a sparse Bayesian modeling framework, this invention combines the 10-meter spatial resolution of Sentinel with the daily observation advantages of MODIS, achieving the reconstruction of daily 10-meter multi-band reflectance, which can meet the needs of highly dynamic ecosystems such as lakes and wetlands for refined monitoring.

[0009] Technical Solution: To achieve the above-mentioned technical objectives, this invention provides the following technical solution: a Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning, comprising the following steps: S1. Acquire daily MODIS reflectance images and Sentinel multi-band high-resolution images to form the MODIS domain and Sentinel domain, respectively. Using the effective observation date k in the Sentinel domain as the benchmark, select the daily MODIS reflectance image M with the minimum cloud cover within k±1 days. k And compared with Sentinel multi-band high-resolution image S on observation date k k Perform time alignment and construct historical sample pairs K represents the total number of effective observation days for the Sentinel multi-band high-resolution image; the remaining reflectance image data in the MODIS domain are denoted as... N is the total number of remaining reflectance images in the MODIS domain; S2, through daily MODIS reflectance images M k Perform spectral response consistency correction to ensure that the corrected MODIS data is consistent with Sentinel multi-band high-resolution imagery in the spectral domain. S3, Reflectance image for the MODIS domain Through adaptive sliding window Get the target date in the sliding window Inner MODIS vector Thus, reflectance images in the MODIS domain are constructed. To M k The training relationship; S4. Based on sparse Bayesian learning, the MODIS vector of the target day within the sliding window is obtained. Sparse time weights ; S5, the MODIS vector of the target day within the sliding window obtained in step S4. Sparse time weights Migrate to the Sentinel domain to obtain window-level Sentinel initial predictions. This leads to the acquisition of the entire initial image. ; S6, calculate each sliding window The residuals of the internal MODIS are then used to perform timing corrections in the Sentinel domain, resulting in a windowed result of the timing consistency callback. And stitch the entire image to form a reflectance image for the target date. .

[0010] Furthermore, the specific process of spectral consistency correction in step S2 is as follows: S2.1, at the pixel scale, through historical samples... Establish an empirical linear mapping relationship from the MODIS domain to the Sentinel domain. This leads to the pixel-level slope a and intercept b; S2.2, the pixel-level slope a and intercept b are extended to the daily MODIS domain, and pixel-by-pixel fitting is performed to obtain the corrected equivalent emissivity of the target day. , .

[0011] Furthermore, the adaptively sized sliding window described in step S3 The specific construction method is as follows: S3.1, for reflectance images in the MODIS domain For each image, construct a sliding window. and make sliding windows The percentage of effective pixels within the range meets the threshold requirement, among which, N is the total number of remaining reflectance images in the MODIS domain; P represents the reflectance image in the MODIS domain. The total number of sliding windows in the internal structure; S3.2, in the sliding window that has been filtered Constructing MODIS reflectance images MODIS reflectance image M k The training relationship is used to obtain the sliding window. Feature matrix and sliding windows MODIS vector of the target day As shown in the following formula, , In the formula, Indicates the corrected number The first MODIS The equivalent reflectance of a sliding window. The MODIS feature matrix representing the historical sample set. Indicates the historical value of the same pixel. Indicates the time difference term. This represents the neighborhood mean.

[0012] Furthermore, step S4, based on sparse Bayesian learning, obtains the value within the sliding window. MODIS vector of the target day Sparse time weights The specific process is as follows: S4.1, in each sliding window A sparse Bayesian linear model is established as shown in the following equation. , In the formula, Indicates sliding window MODIS vector of the target day within the interior. Represents the weight vector; The error term due to observation noise has a mean of 0 and a noise variance of . Independent and identically distributed Gaussian noise , Represents the identity matrix, with dimensions AND same; S4.2, iteratively update hyperparameters by maximizing the marginal likelihood function. With noise variance After suppressing non-contributing components, the weight vector is obtained after the model converges. posterior mean , denoted as sparse time weight ; S4.3, based on each sliding window Sparse time weights The MODIS sparse Bayes prediction results are calculated using the following formula. , , In the formula, This represents the MODIS sparse Bayes prediction result.

[0013] Furthermore, the specific process of step S5 is as follows: S5.1, same sliding window Internal history retrieval Sentinel stack The sparse time weights are calculated using the following formula. The time components are weighted and synthesized to obtain the window-level Sentinel initial prediction. , , In the formula, Indicates date The Sentinel reflectivity vector, Indicates in sliding window The learned sparse time weight vector The One component;

[0014] S5.2, seamlessly stitch the overlapping areas of adjacent sliding windows according to consistency weights to obtain the entire initial predicted image. .

[0015] Furthermore, the specific process of step S6 is as follows: S6.1, in each sliding window The residuals of MODIS are calculated internally and then returned to the Sentinel domain for timing correction using the following formula, yielding the sliding window result after the return. , , In the formula, This represents the result of the sliding window after the timing consistency callback. These are the MODIS observation results for that day. This represents the MODIS sparse Bayes prediction result; S6.2, Sliding window result after timing consistency callback The reflectance image of the target date is formed by stitching together the entire area. .

[0016] Furthermore, the daily MODIS reflectance image and Sentinel multi-band high-resolution image include six bands: blue light, green light, red light, near-infrared, shortwave infrared 1, and shortwave infrared 2.

[0017] Furthermore, in step S3.1, the sliding window The effective pixel ratio within the area must be greater than a threshold, and the threshold value ranges from 10% to 15%.

[0018] Beneficial effects: Compared with the prior art, the present invention has the following advantages: (1) This invention automatically evaluates the importance of variables through the marginal likelihood maximization mechanism, without the need for manual setting of regularization parameters, and achieves robust feature sparsity and adaptive model optimization, significantly improving the robustness and generalization ability of the fusion.

[0019] (2) This invention uses daily MODIS observation as time-driven to achieve a unified output of 10-meter spatial resolution and daily time resolution, breaking through the previous time and space limitations of 30 meters and 8–16 days interval, and can capture the rapid change process of lake wetlands and agricultural ecosystems.

[0020] (3) Unlike traditional fusion models that only target indices such as NDVI or EVI, this invention maintains spectral consistency under a multi-band modeling framework and can simultaneously support the inversion of multiple ecological and environmental parameters, achieving a leap from single indicator to multi-element comprehensive monitoring. Attached Figure Description

[0021] Figure 1 This invention relates to the Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning.

[0022] Figure 2 This is the distribution of MODIS and Sentinel data in this embodiment of the invention, where: color represents cloud cover in MODIS09GA; asterisk represents Sentinel.

[0023] Figure 3 The blue light band fused reflectance and Sentinel consistency evaluation results in this embodiment of the invention include three time phases: October 20, 2015 (descent period), August 24, 2020 (flood season), and January 3, 2025 (dry season).

[0024] Figure 4 The green band fused reflectance and Sentinel consistency evaluation results in this embodiment of the invention include three time phases: October 20, 2015 (descent period), August 24, 2020 (flood season), and January 3, 2025 (dry season).

[0025] Figure 5 The red band fused reflectance and Sentinel consistency evaluation results in this embodiment of the invention include three time phases: October 20, 2015 (descent period), August 24, 2020 (flood season), and January 3, 2025 (dry season).

[0026] Figure 6 The near-infrared band fused reflectance and Sentinel consistency evaluation results in this embodiment of the invention include three time phases: October 20, 2015 (descent period), August 24, 2020 (flood season), and January 3, 2025 (dry season).

[0027] Figure 7The consistency evaluation results of shortwave infrared 1 (1.24μm) fused reflectance and Sentinel in this embodiment of the invention include three time phases: October 20, 2015 (dry season), August 24, 2020 (high water season), and January 3, 2025 (dry season).

[0028] Figure 8 The consistency evaluation results of shortwave infrared 2 (1.64μm) fused reflectance and Sentinel in this embodiment of the invention include three time phases: October 20, 2015 (dry season), August 24, 2020 (high water season), and January 3, 2025 (dry season).

[0029] Figure 9 This is a comparison between the water area extracted by fusion reflectance and Sentinel in this embodiment of the invention.

[0030] Figure 10 This is a comparison between vegetation area extracted by fusion reflectance and Sentinel in this embodiment of the invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise specifically stated, the relative arrangement, expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0032] like Figure 1As shown, this invention presents a Sentinel-MODIS daily 10-meter reflectance fusion method based on sparse Bayesian learning, aiming to address the shortcomings of existing spatiotemporal fusion technologies in terms of spatial resolution, temporal resolution, and multi-band consistency. Traditional spatiotemporal fusion methods mostly focus on Landsat data with a 30-meter resolution and rely on MODIS's 8-day or 16-day composite index products. Therefore, the output results are not continuous day by day. Furthermore, most studies only focus on a single vegetation index, neglecting the comprehensive value of multiple bands such as blue, green, red, near-infrared, and short-wave infrared in ecological environment monitoring. This invention, by introducing a sparse Bayesian modeling framework, combines the 10-meter spatial resolution of Sentinel with the daily observation advantages of MODIS, achieving the reconstruction of daily 10-meter multi-band reflectance, which can meet the needs of highly dynamic ecosystems such as lakes and wetlands for refined monitoring. The specific process of this invention is as follows:

[0033] Step S1: Time registration and construction of historical sample set

[0034] Establish the temporal registration relationship between Sentinel and MODIS and form a historical sample set. The method involves using the effective observation date k of Sentinel as a baseline, and selecting the MODIS image Mk with the lowest cloud cover within ±1 day of Sentinel for temporal matching to form historical sample pairs. .in, M represents the reflectance image of Sentinel on date k. k This represents the MODIS reflectance image closest in time. All images are uniformly resampled to a Sentinel 10 m grid to ensure spatial consistency. Step 1 yields a time-synchronized and spatially aligned set of historical samples. and daily MODIS images for the target date. (10 m grid).

[0035] Step S2, Spectral response uniformity correction

[0036] By eliminating the spectral response differences between the two types of sensors through spectral response consistency correction, the corrected MODIS data can be kept consistent with Sentinel in the spectral domain.

[0037] Using historical samples at the pixel scale An empirical linear mapping relationship from MODIS to Sentinel is established as shown below, thereby obtaining the cell-level slope a and intercept b. , Extending the values ​​of a and b obtained above to the daily MODIS domain, the corrected MODIS equivalent reflectance is obtained by pixel-by-pixel fitting using the following formula. and , , This yields the corrected historical MODIS stack. With target date This is used for subsequent modeling and fusion.

[0038] Step S3: Adaptive region construction and clean pixel selection based on sliding window To form a spatially homogeneous window with sufficient samples And to prepare a pair of data structures for subsequent sparse Bayesian mapping learning: target vector With characteristic matrix The method involves creating windows by sliding them across the entire image at set step sizes. The effective pixel ratio within each window is calculated. If the threshold requirement is met, fusion is performed on the entire window; otherwise, the window is expanded exponentially until the condition is met or the upper limit is reached. Within the filtered windows... upper structure Training relationship: , in, For window The MODIS equivalent reflectance vector of the inner target; The MODIS equivalent reflectance feature matrix for the historical sample set contains historical values ​​of the same pixel. Time difference item and neighborhood mean Get each sliding window Feature matrix With the target vector .

[0039] Step S4: Sparse Bayesian learning and time weight calculation Automatically identify a small number of key dates within the MODIS domain that are most similar to the target date. This is done in each sliding window. Establish a sparse Bayesian linear model: , For weight vector Apply independent Gaussian priors: , The hyperparameters are updated iteratively by maximizing the marginal likelihood function. With noise variance This suppresses non-contributing components. After the model converges, the posterior mean of the weights is obtained. And predict the MODIS reflectance for that day: , In the formula, Represents each sliding window The MODIS predicted reflectance will be used for subsequent residual consistency correction.

[0040] Step S5: Weights are transferred to Sentinel and initial predictions are generated. The goal of this step is to transfer the time weights obtained in step 4 to the Sentinel domain to achieve a preliminary high-resolution prediction of the target date. Specifically, this is done using the same sliding window... Internal history retrieval Sentinel stack ,use The time components are weighted and synthesized: , By averaging the overlapping areas of adjacent windows according to consistency weights, seamless stitching is achieved to obtain the entire initial predicted image. .

[0041] Step S6, Timing Consistency Residual Callback

[0042] This step ensures that the Sentinel fusion results are consistent with the MODIS observations on the same day along the timeline. The method involves performing this process within each sliding window. The residuals of MODIS are calculated internally and then called back to the Sentinel domain for timing correction. , in, This represents the result of the sliding window after the timing consistency callback. These are the MODIS observation results for that day. This represents the MODIS sparse Bayes prediction result. The window result after time-series consistency callback. And stitched together over the entire area to form a 10 m reflectance image of the target date. .

[0043] Step 7: Output and Application of Fusion Results

[0044] The final output is a daily continuous multi-band reflectance data product with a spatial resolution of 10 m, covering blue, green, red, near-infrared, and short-wave infrared bands. This data can be widely used for dynamic monitoring of multiple elements in lakes and wetlands, including water area, transparency, total suspended solids concentration, chlorophyll content, and vegetation type identification, providing key support data with high spatiotemporal resolution for hydrological process simulation, ecosystem assessment, and carbon cycle research.

[0045] Example

[0046] This embodiment uses Poyang Lake National Nature Reserve as an example to verify the performance of daily high-resolution remote sensing fusion achieved by the method of this invention. Poyang Lake National Nature Reserve is located in a typical area of ​​lakes connected to the Yangtze River in the middle and lower reaches of China. The region exhibits significant hydrological rhythms, with distinct alternations of high, low, and dry seasons throughout the year. Submerged plants, emergent plants, and water area fluctuate frequently with changes in water level. Water transparency, total suspended solids concentration, and underwater light environment can change significantly within a few days. Therefore, this area is the most representative wetland type for verifying the performance of daily high-resolution remote sensing fusion. This invention selects Poyang Lake National Nature Reserve as an example to fully demonstrate the advantages and necessity of daily 10-meter multi-band reflectance fusion in highly dynamic ecosystems.

[0047] 1. Data Source and Preprocessing This embodiment uses the Poyang Lake National Nature Reserve as the study area. A total of 3616 daily-scale surface reflectance images from the MODISMOD09GA satellite (2015–2025) and 248 images with cloud cover below 10% from the Sentinel-2 A / B satellite Level-2A product were acquired. To ensure complete regional coverage, multi-track Sentinel images from the same day were stitched together along their orbits. All images were reprojected to WGS 84 / UTM Zone 50N and cropped to fit the study area.

[0048] Cloud and cloud shadow pixels were removed using the MODIS QC band and the Sentinel Scene Classification Layer (SCL). Subsequently, a response mapping relationship was established based on the spectral response functions of the two types of sensors, and spectral consistency correction was performed on the Sentinel band to ensure its correspondence with MODIS in the reflectance domain. Using the MODIS date as a reference, the nearest Sentinel observation was retrieved within ±1 day for matching, thus constructing a daily time-series fusion input dataset.

[0049] Using MODIS diurnal series as a time reference, the nearest Sentinel image within a ±1 day range is searched to achieve time alignment of the target date. Figure 2 The matched MODIS cloud cover distribution and Sentinel coverage features are displayed.

[0050] 2. Sparse Bayesian Learning Fusion Process This embodiment performs independent fusion processing on six bands: blue, green, red, near-infrared, shortwave infrared 1, and shortwave infrared 2. A 30×30 pixel initial sliding window is constructed centered on each target pixel, with a window step size of 20 pixels. The window overlap rate is automatically determined by the relationship between the two to avoid brightness jumps along the stitching seams. After removing cloud contamination and invalid pixels within the sliding window using a cloud mask, if the number of valid pixels is not less than 30, it is directly used for sparse Bayesian model training; otherwise, a window expansion strategy is triggered.

[0051] The window expansion employs an exponential growth method, causing the window area to expand outward at approximately a doubling rate, with a maximum expansion of 10 times. This satisfies the minimum sample size requirement while avoiding computational overhead and heterogeneity interference from excessive expansion. Once the sample size requirement is met, the corresponding pixel sequences of all historical Sentinel and MODIS data within the window are selected as input. Redundant information is automatically compressed using sparse Bayesian learning, retaining only the time-nearest samples that contribute most to the prediction of the current date, thereby improving the model's generalization ability and its ability to suppress outlier observations.

[0052] During the prediction phase, the model reconstructs the Sentinel-level spatial details of the target date based on the daily dynamic changes of MODIS. Because each band is modeled independently, this invention can completely preserve the specific spectral response characteristics of different bands in the distribution of water bodies, submerged vegetation, mudflats, and emergent plants, without causing physical meaning shifts due to mutual interference between bands.

[0053] The window-by-window fusion results employ a seamless stitching mechanism using weighted averaging of overlapping regions to reduce boundary streaks caused by local window differences. This ultimately yielded a daily 10-meter reflectance product series covering the Poyang Lake National Nature Reserve from 2015 to 2025. Figures 3-8 The results show the consistency between the fused reflectance and the Sentinel reference results on different representative dates during the receding, high-water, and low-water periods, demonstrating that the present invention can maintain robust and continuous spectral reconstruction capabilities in highly dynamic environments.

[0054] 3. Validation of water body and vegetation area extraction based on fused reflectance To quantitatively assess the applicability of the fusion product in ecological monitoring, this invention employs a mature spectral index method for comparative verification. Water extraction utilizes the modified normalized difference water index (mNDWI). , by The threshold is used to extract the range of water bodies. Figure 9 The comparison between the water area extracted from the fused data and Sentinel's data shows that the two are highly consistent in shape and area, indicating that the present invention can accurately characterize the dynamic boundary of the water surface on a daily basis.

[0055] Vegetation was extracted using the Enhanced Vegetation Index (EVI). , by Extract wetland vegetation areas. Figure 10 The results showed that the vegetation area extracted from the fused data was consistent with the Sentinel monitoring results, and could identify the expansion and decline processes of emergent vegetation and seasonally exposed mudflat vegetation.

[0056] 4. Model Validation and Accuracy Evaluation

[0057] To verify the reliability of the method of this invention, representative dates were selected from the receding water period (October 20, 2015), the high water period (August 24, 2020), and the low water period (January 3, 2025) to evaluate the pixel-level consistency between the fused reflectance of the six bands and the Sentinel observations of the same period. Figures 3-8 As shown, all bands achieved good fitting results, with the shortwave infrared bands (SWIR1 and SWIR2) performing best, with R² values ​​all above 0.90 and RMSE as low as 0.0126. This is because the SWIR bands have a strong response to changes in moisture content and mudflat exposure, exhibiting large information variation amplitudes, which is beneficial for capturing key driving factors in time-series reconstruction using sparse Bayesian methods. Therefore, this invention is particularly suitable for monitoring ecological processes driven by wetland hydrological conditions, and can support high-precision inversion of important ecological parameters such as suspended solids concentration and soil moisture.

[0058] The fusion results of near-infrared (NIR) and red band measurements are also stable (R² mostly between 0.80 and 0.91), effectively distinguishing emergent vegetation, submerged vegetation, and water body boundaries. This forms an important basis for vegetation indices (such as EVI and SAVI) and biomass estimation. Due to increased heterogeneity of vegetation cover during the receding water period, the RMSE of the red band is relatively high in 2015 and 2025, which may lead to some bias in vegetation cover estimation. However, overall, it can still accurately characterize the seasonal vegetation change pattern.

[0059] The fusion accuracy of green and blue bands is slightly lower than that of red and near-infrared bands, especially during the dry season when the RMSE of the Blue band reaches 0.0649. This is because the blue band is more susceptible to water turbidity and atmospheric Rayleigh scattering, resulting in a higher noise ratio, which may introduce some local biases into water body inversion based on the green / blue bands (such as transparency and plume diffusion discrimination). However, during the wet season when the water body is continuous, both Blue and Green bands show high consistency (R²>0.78), indicating that this method can maintain robust expressive ability during periods of enhanced water connectivity.

[0060] Water area is extracted based on fusion results. Figure 9 ) and vegetation area ( Figure 10Compared with Sentinel, the water extraction results achieved R²=0.93 and RMSE=36.27 km², demonstrating the advantages of this invention in monitoring the rapid expansion and contraction of water bodies. The vegetation area accuracy was R²=0.84 and RMSE=50.55 km². Although affected by vegetation heterogeneity during the receding water period and uncertainties in the blue light band, it can still accurately respond to the stage succession of wetland vegetation and the exposure of mudflats.

[0061] In summary, this invention enables reliable reconstruction of daily, 10-meter, multi-band reflectance sequences, with the shortwave infrared and near-infrared bands showing the best performance, which is beneficial for the comprehensive monitoring of wetland hydrological and vegetation processes. Furthermore, the error structures of each band exhibit physical consistency, ensuring that this invention can be directly used for high-precision dynamic inversion of water bodies, vegetation, and multiple ecological elements, providing reliable data support for the assessment of carbon cycling, hydrological connectivity, and ecological health in lakes and wetlands.

[0062] This invention first performs geometric correction and spectral consistency processing to ensure the comparability of data from different sensors. Then, it constructs a multi-band fusion dictionary and, within a sparse Bayesian learning framework, adaptively selects optimal atoms through marginal likelihood maximization to automatically assess and narrow the contribution of variables, thereby ensuring the robustness of the results at the probabilistic modeling level. This invention can simultaneously achieve daily-scale temporal continuity, fine characterization with 10-meter spatial resolution, and consistent reconstruction of multi-band reflectance, further improving prediction accuracy through residual correction and spatial smoothing. Compared with existing methods, this invention overcomes the resolution and synthesis cycle limitations of traditional products, significantly improving the monitoring capabilities of rapid hydrological and vegetation processes in highly dynamic ecological environments such as lakes and wetlands, providing reliable data support for water quality inversion, ecosystem assessment, and carbon cycle research.

[0063] The above description is only a partial embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning, characterized in that... Includes the following steps: S1. Acquire daily MODIS reflectance images and Sentinel multi-band high-resolution images to form the MODIS domain and Sentinel domain, respectively. Using the effective observation date k in the Sentinel domain as the benchmark, select the daily MODIS reflectance image M with the minimum cloud cover within k±1 days. k And compared with Sentinel multi-band high-resolution image S on observation date k k Perform time alignment and construct historical sample pairs K represents the total number of effective observation days for the Sentinel multi-band high-resolution image; the remaining reflectance image data in the MODIS domain are denoted as... N is the total number of remaining reflectance images in the MODIS domain; S2, through daily MODIS reflectance images M k Perform spectral response consistency correction to ensure that the corrected MODIS data is consistent with Sentinel multi-band high-resolution imagery in the spectral domain. S3, Reflectance image for the MODIS domain Through adaptive sliding window Get the target date in the sliding window Inner MODIS vector Thus, reflectance images in the MODIS domain are constructed. To M k The training relationship; S4. Based on sparse Bayesian learning, the MODIS vector of the target day within the sliding window is obtained. Sparse time weights ; S5, the MODIS vector of the target day within the sliding window obtained in step S4. Sparse time weights Migrate to the Sentinel domain to obtain window-level Sentinel initial predictions. This leads to the acquisition of the entire initial image. ; S6, calculate each sliding window The residuals of the internal MODIS are then used to perform timing corrections in the Sentinel domain, resulting in a windowed result of the timing consistency callback. And stitch the entire image to form a reflectance image for the target date. .

2. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to claim 1, characterized in that: The specific process of spectral consistency correction in step S2 is as follows: S2.1, at the pixel scale, through historical samples... Establish an empirical linear mapping relationship from the MODIS domain to the Sentinel domain. This leads to the pixel-level slope a and intercept b; S2.2, the pixel-level slope a and intercept b are extended to the daily MODIS domain, and pixel-by-pixel fitting is performed to obtain the corrected equivalent emissivity of the target day. , .

3. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to claim 1, characterized in that: The adaptive-size sliding window described in step S3 The specific construction method is as follows: S3.1, for reflectance images in the MODIS domain For each image, construct a sliding window. and make sliding windows The percentage of effective pixels within the range meets the threshold requirement, among which, N is the total number of remaining reflectance images in the MODIS domain; P represents the reflectance image in the MODIS domain. The total number of sliding windows in the internal structure; S3.2, in the sliding window that has been filtered Constructing MODIS reflectance images MODIS reflectance image M k The training relationship is used to obtain the sliding window. Feature matrix and sliding windows MODIS vector of the target day As shown in the following formula, ， In the formula, Indicates the corrected number The first MODIS The equivalent reflectance of a sliding window. The MODIS feature matrix representing the historical sample set. Indicates the historical value of the same pixel. Indicates the time difference term. This represents the neighborhood mean.

4. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to claim 1, characterized in that: Step S4, based on sparse Bayesian learning, obtains the value within the sliding window. MODIS vector of the target day Sparse time weights The specific process is as follows: S4.1, in each sliding window A sparse Bayesian linear model is established as shown in the following equation. ， In the formula, Indicates sliding window MODIS vector of the target day within the interior. Represents the weight vector; The error term due to observation noise has a mean of 0 and a noise variance of . Independent and identically distributed Gaussian noise , Represents the identity matrix, with dimensions AND same; S4.2, iteratively update hyperparameters by maximizing the marginal likelihood function. With noise variance After suppressing non-contributing components, the weight vector is obtained after the model converges. posterior mean , denoted as sparse time weight ; S4.3, based on each sliding window Sparse time weights The MODIS sparse Bayes prediction results are calculated using the following formula. , ， In the formula, This represents the MODIS sparse Bayes prediction result.

5. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to claim 1, characterized in that: The specific process of step S5 is as follows: S5.1, same sliding window Internal history retrieval Sentinel stack The sparse time weights are calculated using the following formula. The time components are weighted and synthesized to obtain the window-level Sentinel initial prediction. , ， In the formula, Indicates date The Sentinel reflectivity vector, Indicates in sliding window The learned sparse time weight vector The One component; S5.2, seamlessly stitch the overlapping areas of adjacent sliding windows according to consistency weights to obtain the entire initial predicted image. .

6. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to claim 1, characterized in that: The specific process of step S6 is as follows: S6.1, in each sliding window The residuals of MODIS are calculated internally and then returned to the Sentinel domain for timing correction using the following formula, yielding the sliding window result after the return. , ， In the formula, This represents the result of the sliding window after the timing consistency callback. These are the MODIS observation results for that day. This represents the MODIS sparse Bayes prediction result; S6.2, Sliding window result after timing consistency callback The reflectance image of the target date is formed by stitching together the entire area. .

7. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to any one of claims 1-6, characterized in that: The daily MODIS reflectance images and Sentinel multi-band high-resolution images include six bands: blue, green, red, near-infrared, shortwave infrared 1, and shortwave infrared 2.

8. The Sentinel-MODIS daily reflectance fusion method based on sparse Bayesian learning according to claim 3, characterized in that: In step S3.1, the sliding window The effective pixel ratio within the area must be greater than a threshold, and the threshold value ranges from 10% to 15%.

Images