Remote sensing spatio-temporal data fusion method based on phenological constraint and meteorological driving
By constructing meteorological-driven and phenological-constrained models and combining them with high spatial and temporal resolution remote sensing images, the problem of insufficient phenological patterns and meteorological factors in remote sensing spatiotemporal fusion was solved, resulting in more stable and reliable high spatiotemporal resolution remote sensing products suitable for multi-year vegetation monitoring and environmental assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INSPUR OPTOELECTRONICS SATELLITE TECHNOLOGY (SHANDONG) CO LTD
- Filing Date
- 2025-12-23
- Publication Date
- 2026-05-05
AI Technical Summary
Existing remote sensing spatiotemporal fusion methods ignore crop phenological patterns and do not consider meteorological factors, resulting in insufficient prediction accuracy in complex environments and a lack of reliability assessment of the fusion results.
The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving forces constructs meteorological driving models and phenological models, combines high spatial resolution and high temporal resolution remote sensing images, and uses meteorological driving factors and phenological constraint functions to correct the fusion results, thereby generating high spatiotemporal resolution remote sensing products.
This enhances the scientific validity and application value of remote sensing data fusion, generating more stable and reliable high spatiotemporal resolution remote sensing products suitable for large-scale, multi-year vegetation monitoring and environmental assessment.
Smart Images
Figure CN121982518A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving, belonging to the field of data processing technology. Background Technology
[0002] Remote sensing imagery, as an important means of Earth observation, is widely used in fields such as agricultural monitoring, ecological environment assessment, resource surveys, and global change research. Existing remote sensing data often exhibit complementarity in spatial and temporal resolution: high spatial resolution imagery (such as Landsat, Sentinel-2, and the GF series) can provide detailed spatial texture and ground feature information, but its revisit period is long, it is significantly affected by cloud cover, and its temporal continuity is insufficient; high temporal resolution imagery (such as MODIS and VIIRS) can provide continuous time-series data, but its spatial resolution is low, making it difficult to meet the needs of precision agriculture and regional ecological monitoring.
[0003] To address these contradictions, the academic community has proposed various spatio-temporal data fusion methods. These methods combine images from different sources to generate remote sensing products that simultaneously possess high spatial and temporal resolution. Typical methods include: STARFM (Spatial and Temporal Adaptive Reflectance Fusion Model), which predicts high-resolution images using weighted similar pixels; ESTARFM (Enhanced STARFM), which introduces dual-temporal images on top of STARFM to improve prediction accuracy under complex terrain; and FSDAF (Flexible Spatiotemporal Data Fusion), an image decomposition-based method that enhances adaptability to heterogeneous terrain.
[0004] However, most of the above methods are based on the assumption of linear change, which assumes that the trend of pixel spectral change can be approximated as linear between two known time phases. This assumption often does not hold true in actual agriculture and the ecological environment. For example, the growth process of crops has obvious phenological stages, and its remote sensing indices (NDVI, EVI, etc.) show typical nonlinear change patterns (rapid growth - plateau period - decline); vegetation dynamics are also strongly driven by meteorological conditions (temperature, precipitation, light, etc.). Under extreme climate events (drought, flood, high temperature), the changes in vegetation indices often show abrupt changes, which are difficult to accurately characterize by linear models; existing methods usually only rely on the remote sensing image itself and lack the utilization of external driving factors, resulting in insufficient prediction accuracy in complex spatiotemporal environments.
[0005] Therefore, existing spatiotemporal fusion methods still have the following shortcomings in agricultural and ecological monitoring applications: (1) Ignoring crop phenological patterns leads to deviations in the generated time series images during key growth stages; (2) The forecasts are inaccurate and lack robustness under extreme weather conditions because they do not take meteorological factors into account. (3) The lack of uncertainty quantification makes it impossible to effectively assess the reliability of the fusion results. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of existing remote sensing spatiotemporal fusion methods and provide a remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving. On the basis of multi-source data fusion, crop phenological curves and meteorological driving information are combined to overcome the limitations of linear assumptions and improve the scientificity and application value of spatiotemporal fusion products.
[0007] The technical solution adopted in this invention is as follows: A remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological drivers includes the following steps: S1. Acquire high spatial resolution remote sensing images, high temporal resolution remote sensing images, and concurrent meteorological product data; preprocess them; and calculate the vegetation index WDRVI for land cover classification. S2. Weather-driven modeling: Meteorological driving factors are constructed based on meteorological reanalysis data, and a mapping relationship between changes in meteorological variables and short-term responses of vegetation indices is established to form a meteorological driving model for characterizing the impact of short-term meteorological disturbances on vegetation indices. S3. Perform phenological modeling: Based on land cover classification, typical sample plots of the target crop are selected. Time series of the sample area are extracted from high temporal resolution data. The time series are fitted with nonlinear curves using a Gaussian model to establish a phenological growth model. The model is then optimized by solving the parameters to obtain a regional representative phenological model for the crop type. Key phenological periods are extracted from the fitted curve using the phenological model. The fitted phenological curve is used as a prior constraint function for the time series to establish a phenological constraint model. The phenological constraint model is then spatiotemporally generalized to generate a pixel-level phenological constraint function. S4. Using high spatial resolution remote sensing images and high temporal resolution remote sensing images, a three-step process of regression modeling, spatial filtering, and residual compensation is used to perform preliminary spatiotemporal fusion of fine-resolution images and coarse-resolution images; S5. Calculate the deviation index between the preliminary spatiotemporal fusion results and the predicted values of the phenological model curves and detect anomalies. Normal results are corrected using the phenological constraint function; results with large deviations or anomalies are first corrected using the phenological constraint function, then the meteorological weights of the meteorological driving factors are calculated, and the short-term changes of the vegetation index are calculated based on the weights. The preliminary spatiotemporal fusion results are then corrected by meteorological driving, and the results corrected by phenological constraints and meteorological driving are fused to generate the final high spatiotemporal resolution image sequence.
[0008] In the above method, step S1 uses ERA5 product meteorological data.
[0009] The process of constructing the meteorological driving model in step S2 is as follows: 1) Study the influence of meteorological driving factors on vegetation indices based on historical data, and obtain the sensitivity coefficients of meteorological factors. ; 2) Calculation of meteorological weights: , in As a sensitive factor for the growing season, it is determined based on the calculated vegetation index value. If the vegetation index exceeds a fixed threshold, it is assumed that the growing season has begun. During the growing season, If it is not the growing season, then , Determined based on the overall quality of remote sensing data; when the overall quality of remote sensing data is high... Taking a larger value indicates that the overall quality of the remote sensing data is higher. Take the smaller value; 3) Predicting short-term increases in vegetation index: This involves forecasting short-term changes in the vegetation index. It is represented as a linear combination of changes in multiple meteorological factors, with each meteorological factor corresponding to a regression coefficient, indicating the sensitivity of that variable to the vegetation index. , in, Describe the intensity of the impact of meteorological variables on vegetation status. This describes the deviation of a meteorological variable from its baseline state, and is the change of the Kth meteorological variable at time t relative to the multi-year climate average.
[0010] The meteorological driving factors are selected as accumulated temperature index AGDD and moisture stress factor VPD.
[0011] In step S3, the phenological growth model and phenological curves are discussed. The calculation formula is as follows: , Where a represents the peak amplitude, reflecting the intensity of crop growth; b represents the peak time, corresponding to the critical period of crop growth; c represents the curve width parameter, reflecting the duration of the growing season; and t represents the time variable, used to describe the change of vegetation index over time.
[0012] In step S3, the parameters are solved and optimized using the nonlinear least squares method to fit the parameters, so that the model curve matches the observation data best. A set of parameters is saved for each sample plot, and a weighted average is performed on the same type of land cover area to obtain the regional representative phenological model of the crop type.
[0013] In step S3, the fitted phenological curves will be... As a priori constraint function for time series, for any pixel If its temporal variation deviates significantly from the curve, the following constraint function is used for constraint correction: , in This is the phenological constraint weighting coefficient, determined based on the overall quality of the remote sensing image. A higher coefficient value corresponds to poorer image quality. (WDRVI) obs (x,y,t) represents the result after the initial fusion.
[0014] The spatiotemporal generalization process in step S3 is as follows: 1) Preparation of parameterization results: Obtain the parameter set of the Gaussian model obtained by fitting the sample area: i=1,2,...,N Where N is the number of sample regions; Construct a time-series climate vector for each cell: , in As an accumulated temperature index, It is a water stress factor; 2) Calculate the raster spatial distance: the cell index is Target pixel is The formula for calculating relative spatial distance is as follows: ; 3) Calculate climate similarity distance: First, the climate variables are standardized, as shown in the following formula: , in Indicates the first The mean of each climate variable, Indicates the first The standard deviation of each climate variable Then calculate the climatic Euclidean distance. The more similar the climatic conditions, the smaller the climatic Euclidean distance. The formula is as follows: , 4) Joint weight calculation: Introducing climate bandwidth parameters The scope of climate similarity is controlled and climate weights are calculated using the following formula: , Introducing spatial bandwidth parameters The spatial weights are calculated using the following formula: , The final joint weight is obtained by combining the climate weight and the spatial weight, as shown in the following formula: ; 5) Spatiotemporal generalization formulas for phenological parameters: For target pixel Its phenological parameters Promoted as: ; 6) Substitute the generalized phenological parameters into the phenological curve formula: , That is, a, b, and c use the parameters at (x0, y0).
[0015] The data fusion result after meteorological driving correction in step S5 is shown in the following formula: , This indicates the final fusion result; This indicates the index results after phenological constraints; △ This indicates the amount of vegetation index change caused by meteorological factors. This represents the meteorological correction weighting coefficient, which is determined based on the overall quality of the remote sensing image. The lower the image quality, the larger the coefficient value.
[0016] The beneficial effects of this invention are: This invention, by introducing vegetation phenological process models and meteorological driving variables into the fusion of high spatial resolution and high temporal resolution images, not only effectively improves the continuity and ecological rationality of time series reconstruction, but also significantly enhances the responsiveness of the fusion results to meteorological changes. This method can generate high spatiotemporal resolution remote sensing products with lower noise, more stable phenological characteristics, and greater reliability, and is applicable to various scenarios such as large-scale, multi-year vegetation monitoring, environmental assessment, and ecological process analysis. Attached Figure Description
[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0018] The following description, in conjunction with specific embodiments, provides further details.
[0019] Example 1: A remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving forces, comprising the following steps: S1. Acquire high spatial resolution remote sensing imagery, high temporal resolution remote sensing imagery, and contemporaneous meteorological product data; preprocess them; and calculate the vegetation index WDRVI for land cover classification. High spatial resolution remote sensing imagery (Sentinel-2) was acquired, primarily to provide information on surface spatial texture and detail; high temporal resolution remote sensing imagery (MODIS) provided continuous time-series information, mainly for capturing dynamic trends in vegetation change; and contemporaneous meteorological data (ERA5 product) provided hourly estimates of numerous atmospheric, terrestrial, and marine climate variables; atmospheric correction, geometric registration, and cloud detection were performed on the remote sensing imagery; the remote sensing data resolution was uniformly resampled to a spatial resolution consistent with Sentinel-2 imagery; and the vegetation index WDRVI was calculated. , , In the formula, α is a weighting coefficient (usually taken as 0.1~0.2), which is used to enhance the sensitivity of low vegetation cover areas. In this study, α=0.1 is selected according to the experimental requirements.
[0020] S2. Weather-driven modeling: Meteorological driving factors are constructed based on meteorological reanalysis data, and a mapping relationship between changes in meteorological variables and short-term responses of vegetation indices is established to form a meteorological driving model for characterizing the impact of short-term meteorological disturbances on vegetation indices. The meteorological-driven modeling of this invention aims to utilize ERA5 meteorological reanalysis data to correct the short-term trends of preliminary spatiotemporal fusion results by constructing a response relationship between meteorological variables and WDRVI, thereby improving the stability and accuracy under different climatic conditions. Based on actual needs, air temperature and soil moisture indicators from the ERA5 product were selected, as these two indicators have a significant impact on crop growth. The formula is as follows: Accumulated temperature index AGDD: , in For the first Daily temperature, This refers to accumulated temperature for phenological events.
[0021] Water stress factor VPD: Using the soil moisture layer directly provided by ERA5, an approximation of vapor pressure deficit (VPD) was constructed. , Saturated vapor pressure and actual water vapor pressure Calculated from temperature and relative humidity.
[0022] Considering the seasonal characteristics of vegetation growth, this invention employs a dynamic response model. Its core assumption is that the direction and magnitude of vegetation index changes can be predicted using the gradient of meteorological fluxes when meteorological conditions change. A meteorological-driven model is used to correct the trend of the initial spatiotemporal fusion results. If stress conditions such as drought or low temperature occur, the model will predict a decrease in NDVI and apply a negative correction to the fusion results. If accumulated temperature increases rapidly and sunlight is sufficient, the model will predict accelerated growth and apply a positive correction to the fusion results. This process ensures that the reconstructed time series remains scientifically sound even under abrupt weather changes.
[0023] S3. Perform phenological modeling: Based on land cover classification, typical sample plots of the target crop are selected. Time series of the sample area are extracted from high temporal resolution data. The time series are fitted with nonlinear curves using a Gaussian model to establish a phenological growth model. The model is then optimized by solving the parameters to obtain a regional representative phenological model for the crop type. Key phenological periods are extracted from the fitted curve using the phenological model. The fitted phenological curve is used as a prior constraint function for the time series to establish a phenological constraint model. The phenological constraint model is then spatiotemporally generalized to generate a pixel-level phenological constraint function. Based on historical remote sensing data and agronomic knowledge, curve fitting is performed on the growth process of major crops or vegetation to establish a phenological growth model, which serves as a time constraint in the spatiotemporal fusion process.
[0024] (1) Sample area selection: Based on the results of land parcel vector or land cover classification, select typical sample parcels of the target crop, and select areas with simple land cover type, little cloud cover and complete time series within the region.
[0025] (2) Time series extraction: Extract the time series of the sample area from the high temporal resolution data (MODIS), and perform Savitzky-Golay filtering to smooth the time series to remove noise, so as to obtain the crop phenological curve that can represent the crop growth trajectory.
[0026] (3) Curve Fitting: Nonlinear curve fitting was performed on the smoothed time series to establish a phenological growth model. The Gaussian model was selected for fitting. Gaussian curves have natural symmetry and smoothness, and will not show abrupt changes during the growth or decline stages. This model is suitable for unimodal vegetation changes, such as for crops with one harvest per year or obvious unimodal growth cycles (e.g., wheat, corn, rice, grassland, forest WDRVI annual changes). Gaussian curves can accurately describe the "rapid growth-slow decline" pattern, which is suitable for the selected study area. It can better maintain the smoothness of the curve and the consistency of the time series, and reduce the impact of fitting fluctuations on the subsequent fusion results. The formula for calculating the phenological curve is as follows: , Where a represents the peak amplitude, reflecting the crop growth intensity (such as the maximum WDRVI increase); b represents the peak time, corresponding to the key period of crop growth (such as the heading stage); and c represents the curve width parameter, reflecting the duration of the growing season.
[0027] (3) Parameter solution and optimization: The nonlinear least squares method is used to fit the parameters so that the model curve matches the observation data best. A set of parameters is saved for each sample plot, and the weighted average is performed on the same type of land cover area to obtain the regional representative phenological model of the crop type.
[0028] (4) Establish a phenological constraint model: The fitted phenological curves As a priori constraint function for time series, for any pixel If its time series changes significantly deviate from the curve, it can be constrained and corrected using the following formula: , in The phenological constraint weight coefficient (0-1) is determined based on the overall quality of the remote sensing image; the higher the value of the coefficient, the worse the image quality.
[0029] (5) Spatiotemporal extension of phenological models: The meteorological data used for the spatiotemporal extension of phenological parameters are only derived from ERA5 / ERA5-Land. Considering the relatively coarse spatial resolution of ERA5, this invention adopts a downscaling method combining weighted regression based on meteorological similarity and spatial interpolation. The ERA5 driving factors (such as effective accumulated temperature, accumulated precipitation, daily shortwave incidence, and root zone soil moisture) are downsampled or vertically interpolated to the pixel scale of the study area, and bias correction and standardization are performed on them. Then, these processed ERA5 derived features are used as explanatory variables of geographic weighted regression (GWR), and the Gaussian phenological parameters at each pixel are estimated by local weighted least squares. This process ultimately generates pixel-level phenological constraint functions. A corresponding phenological constraint function is established for each pixel, serving as input for the subsequent spatiotemporal fusion module.
[0030] a. Preparation of parameterization results: Obtain the parameter set of the Gaussian model obtained by fitting the sample area: i=1,2,...,N Where N is the number of sample regions. Construct a time-series climate vector for each cell: , in As an accumulated temperature index, It is a water stress factor.
[0031] b. Calculate raster spatial distance: cell index is Target pixel is The formula for calculating relative spatial distance is as follows: ; c. Calculate climate similarity distance: First, the climate variables are standardized, as shown in the following formula: , in Indicates the first The mean of each climate variable, Indicates the first The standard deviation of each climate variable.
[0032] Then calculate the climatic Euclidean distance. The more similar the climatic conditions, the smaller the climatic Euclidean distance. The formula is as follows: .
[0033] d. Joint weight calculation: Introducing climate bandwidth parameters The scope of climate similarity is controlled and climate weights are calculated using the following formula: , Introducing spatial bandwidth parameters The spatial weights are calculated using the following formula: , The final joint weight is obtained by combining the climate weight and the spatial weight, as shown in the following formula: ; e. Spatiotemporal generalization formulas for phenological parameters: For target pixel Its phenological parameters Promoted as: ; f. Substitute the generalized phenological parameters into the phenological curve formula: , That is, a, b, and c use the parameters at (x0, y0).
[0034] S4. Using high spatial resolution and high temporal resolution remote sensing images, preliminary spatiotemporal fusion of fine-resolution and coarse-resolution images is performed through three steps: regression modeling, spatial filtering, and residual compensation. This method employs a three-step process—regression modeling, spatial filtering, and residual compensation—to spatiotemporally fuse fine-resolution images (i.e., high spatial-to-low temporal resolution remote sensing images) with coarse-resolution images (i.e., high temporal-to-low spatial resolution remote sensing images), maximizing the preservation of spatial details and temporal continuity. The specific steps are as follows: (1) Regression model fitting: The relationship between the coarse resolution image and the fine resolution image at the target time is established through regression analysis.
[0035] The first step is local regression modeling. In the coarse-resolution image, linear regression is performed on each pixel and its neighborhood using time-series data, assuming that the pixel values of the coarse-resolution image change linearly over time. , In the formula, Let be the reflectance value of the i-th pixel at time t. and For regression coefficients, This is the error term.
[0036] Next is the estimation of regression model parameters. The regression coefficients are obtained by fitting the time series of each pixel using the least squares method. and : , Finally, the regression model was applied to the coarse-resolution image to obtain a preliminary high-resolution predicted image: , In the formula, It is a high-resolution image of band B with a known date. This is a fine-resolution image of the predicted band B. However, due to the limited spectral information in coarse-resolution images, regression models often produce obvious blocky artifacts, affecting the fusion of spatial details.
[0037] (2) Spatial filtering: Reduce spatial artifacts caused by the regression model and optimize the local spatial structure.
[0038] Based on spatial neighborhood similarity, similar pixels to the center pixel within the window are selected. The formula is as follows: , By calculating the spectral difference D between each pixel in the window and the center pixel, the smaller D is, the more similar the pixel is to the center pixel. The first n pixels (including the center pixel) that are most similar to the center pixel are identified as similar pixels.
[0039] Then, a normalized weighted method is used to process the spatial distance between similar pixels and the center pixel in the moving window. The weight is determined by the distance between pixels; the closer the pixels are, the higher the weight, and the greater their contribution to the center pixel. The weight is calculated by the following formula: , , In the formula, w represents the set window size, and d i W represents the spatial distance from the center pixel. i Indicates the weight.
[0040] Finally, the fine-resolution pixel values for the predicted date are weighted and assigned to obtain the spatially filtered and smoothed result. This preserves the local spatial features of the pixels while removing block artifacts in the regression prediction.
[0041] , In the formula This represents the prediction result of the x0 pixel in band B after spatial filtering. This preserves the local spatial characteristics of the pixel while removing block artifacts in the regression prediction.
[0042] (3) Residual compensation: Further correct the error of the regression model and improve the accuracy of the fused image.
[0043] This step fully utilizes the spectral information of the coarse-resolution image from the observation date to calculate the coarse-resolution image residual. Then, it resamples to the spatial resolution of the fine-resolution image using bicubic interpolation to obtain the initial residual. To avoid over-smoothing the interpolation, spatial filtering is applied to the residual. , In the formula, This indicates that high-resolution residuals are obtained through bicubic interpolation. This represents the residual corrected for pixel x0 in band B.
[0044] Finally, the optimized residuals are superimposed onto the spatially filtered prediction results to further improve the spectral consistency of the prediction results. , After three optimization steps—regression fitting, spatial filtering, and residual compensation—a preliminary fusion result with high resolution and spatiotemporal consistency was finally obtained.
[0045] S5. Calculate the deviation index between the preliminary spatiotemporal fusion results and the predicted values of the phenological model curves, and detect anomalies. Normal results are corrected using the phenological constraint function; results with large deviations or anomalies are first corrected using the phenological constraint function, then the meteorological weights of the meteorological driving factors are calculated, and the short-term changes in the vegetation index are calculated based on the weights to correct the preliminary spatiotemporal fusion results. Finally, the results corrected by phenological constraints and meteorological driving factors are fused to generate the final high spatiotemporal resolution image sequence. Phenological constraint optimization: The preliminary results are compared with the phenological curves to correct outliers that deviate from the phenological patterns. The steps are as follows: 1) Calculate the deviation index: This measures the degree of consistency between the preliminary fusion results and the phenological curve predictions. A smaller value indicates that the pixel is in a normal phenological state, while a larger value indicates abnormal fluctuations at that moment, such as residual cloud shadows, sensor noise, or phenological misalignment. ; 2) Detect outliers (mutations / noise points) The formula for calculating the first-order time difference is as follows: , , in This represents the average rate of change of that pixel over the entire time series. Vegetation growth is a gradual process, therefore... It is usually close to 0; The standard deviation represents the normal fluctuation range of pixel time variation. The larger the standard deviation, the more drastic the change, and the smaller the standard deviation, the smoother the change, such as the growing season of grassland or farmland.
[0046] like (r is set based on historical fluctuations), marked as abnormal fluctuations, and the fitted phenological curve is used first for repair: , in It is the result of phenological constraints.
[0047] 3) Local fitting: For non-outlier points, use the formula to and Weighted synthesis is used to reconstruct short-term missing data using interpolation under phenological constraints.
[0048] 4) Phenological constraints: The phenological constraint correction formula is shown below: , in The weighting of predicted and fitted values is controlled. This experiment is set up... The weight is 0.6, meaning the predicted value has a relatively large weight.
[0049] Meteorological-driven correction: For the anomalies obtained in the steps of calculating deviation indicators and detecting anomalies, the meteorological-driven model is used to adjust the short-term trend of change, improve the adaptability under extreme climate conditions, and superimpose the short-term positive or negative increments caused by meteorology onto the phenologically constrained sequence to ensure response to abnormal climate.
[0050] 1) Based on historical data, study the influence of meteorological driving factors on vegetation indices and obtain the sensitivity coefficients of meteorological factors. ; 2) Calculation of meteorological weights: , in As a growing season sensitive factor (0-1), if During the growing season, If it is not the growing season, then (Significantly reducing the weighting of weather corrections), this experiment Take 0.8; 3) Predicting short-term increases in vegetation index: Short-term changes in vegetation index It can be represented as a linear combination of changes in multiple meteorological factors, with each meteorological factor corresponding to a regression coefficient, indicating the sensitivity of that variable to the vegetation index. , in, Describe the intensity of the influence of meteorological variables on vegetation status. It describes the deviation of meteorological variables from the baseline state (such as wetter than usual, hotter than usual). Meteorological variables together determine the short-term changes in vegetation indices. This fusion result after incremental correction of phenological constraints enables it to respond to real meteorological events.
[0051] 4) Weather-driven corrections: The data fusion result after meteorological driving correction is shown in the following formula: , This indicates the final fusion result; This indicates the index results after phenological constraints; △ This indicates the amount of vegetation index change caused by meteorological factors. This represents the meteorological correction weighting coefficient, which is determined based on the overall quality of the remote sensing image. The lower the image quality, the larger the coefficient value.
[0052] The initial spatiotemporal fusion results were then subjected to dual-constraint fusion: combining phenological constraints and meteorological corrections, to generate the final high spatiotemporal resolution image sequence.
[0053] The above is a further description of the present invention in conjunction with embodiments, and the scope of protection of the present invention is not limited thereto.
Claims
1. A remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving forces, characterized in that, The steps include the following: S1. Acquire high spatial resolution remote sensing images, high temporal resolution remote sensing images, and concurrent meteorological product data; preprocess them; and calculate the vegetation index WDRVI for land cover classification. S2. Weather-driven modeling: Meteorological driving factors are constructed based on meteorological reanalysis data, and a mapping relationship between changes in meteorological variables and short-term responses of vegetation indices is established to form a meteorological driving model for characterizing the impact of short-term meteorological disturbances on vegetation indices. S3. Perform phenological modeling: Based on land cover classification, typical sample plots of the target crop are selected. Time series of the sample area are extracted from high temporal resolution data. The time series are fitted with nonlinear curves using a Gaussian model to establish a phenological growth model. The model is then optimized by solving the parameters to obtain a regional representative phenological model for the crop type. Key phenological periods are extracted from the fitted curve using the phenological model. The fitted phenological curve is used as a prior constraint function for the time series to establish a phenological constraint model. The phenological constraint model is then spatiotemporally generalized to generate a pixel-level phenological constraint function. S4. Using high spatial resolution remote sensing images and high temporal resolution remote sensing images, a three-step process of regression modeling, spatial filtering, and residual compensation is used to perform preliminary spatiotemporal fusion of fine-resolution images and coarse-resolution images; S5. Calculate the deviation index between the preliminary spatiotemporal fusion results and the predicted values of the phenological model curves and detect anomalies. Normal results are corrected using the phenological constraint function; results with large deviations or anomalies are first corrected using the phenological constraint function, then the meteorological weights of the meteorological driving factors are calculated, and the short-term changes of the vegetation index are calculated based on the weights. The preliminary spatiotemporal fusion results are then corrected by meteorological driving, and the results corrected by phenological constraints and meteorological driving are fused to generate the final high spatiotemporal resolution image sequence.
2. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, Step S1: Meteorological product data is obtained using ERA5 product meteorological data.
3. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, The process of constructing the meteorological driving model in step S2 is as follows: 1) Study the influence of meteorological driving factors on vegetation indices based on historical data, and obtain the sensitivity coefficients of meteorological factors. ; 2) Calculation of meteorological weights: ， in As a sensitive factor for the growing season, it is determined based on the calculated vegetation index value. If the vegetation index exceeds a fixed threshold, it is assumed that the growing season has begun. During the growing season, If it is not the growing season, then , Determined based on the overall quality of remote sensing data; when the overall quality of remote sensing data is high... Taking a larger value indicates that the overall quality of the remote sensing data is higher. Take the smaller value; 3) Predicting short-term increases in vegetation index: This involves forecasting short-term changes in the vegetation index. It is represented as a linear combination of changes in multiple meteorological factors, with each meteorological factor corresponding to a regression coefficient, indicating the sensitivity of that variable to the vegetation index.
4. Among them, Describe the intensity of the impact of meteorological variables on vegetation status. This describes the deviation of a meteorological variable from its baseline state, and is the change of the Kth meteorological variable at time t relative to the multi-year climate average.
5. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, In step S3, the phenological growth model and phenological curves are discussed. The calculation formula is as follows: ， Where a represents the peak amplitude, reflecting the intensity of crop growth; b represents the peak time, corresponding to the critical period of crop growth; c represents the curve width parameter, reflecting the duration of the growing season; and t represents the time variable, used to describe the change of vegetation index over time.
6. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, In step S3, the parameters are solved and optimized using the nonlinear least squares method to fit the parameters, so that the model curve matches the observation data best. A set of parameters is saved for each sample plot, and a weighted average is performed on the same type of land cover area to obtain the regional representative phenological model of the crop type.
7. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, In step S3, the fitted phenological curves will be... As a priori constraint function for time series, for any pixel If its temporal variation deviates significantly from the curve, the following constraint function is used for constraint correction: ， in This is the phenological constraint weighting coefficient, determined based on the overall quality of the remote sensing image. A higher coefficient value corresponds to poorer image quality. (WDRVI) obs (x,y,t) represents the result after the initial fusion.
8. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, The spatiotemporal generalization process in step S3 is as follows: 1) Preparation of parameterization results: Obtain the parameter set of the Gaussian model obtained by fitting the sample area: i=1,2,...,N, Where N is the number of sample regions; Construct a time-series climate vector for each cell: ， in As an accumulated temperature index, It is a water stress factor; 2) Calculate the raster spatial distance: the cell index is Target pixel is The formula for calculating relative spatial distance is as follows: ； 3) Calculate climate similarity distance: First, the climate variables are standardized, as shown in the following formula: ， in Indicates the first The mean of each climate variable, Indicates the first The standard deviation of each climate variable Then calculate the climatic Euclidean distance. The more similar the climatic conditions, the smaller the climatic Euclidean distance. The formula is as follows: ， 4) Joint weight calculation: Introducing climate bandwidth parameters The scope of climate similarity is controlled and climate weights are calculated using the following formula: ， Introducing spatial bandwidth parameters The spatial weights are calculated using the following formula: ， The final joint weight is obtained by combining the climate weight and the spatial weight, as shown in the following formula: ； 5) Spatiotemporal generalization formulas for phenological parameters: For target pixel Its phenological parameters Promoted as: ； 6) Substitute the generalized phenological parameters into the phenological curve formula: ， That is, a, b, and c use the parameters at (x0, y0).
9. The remote sensing spatiotemporal data fusion method based on phenological constraints and meteorological driving as described in claim 1, characterized in that, The data fusion result after meteorological driving correction in step S5 is shown in the following formula: ， This indicates the final fusion result; This indicates the index results after phenological constraints; △ This indicates the amount of vegetation index change caused by meteorological factors. This represents the meteorological correction weighting coefficient, which is determined based on the overall quality of the remote sensing image. The lower the image quality, the larger the coefficient value.