A multi-source data fusion method based on spatiotemporal dynamic triple matching analysis

Abstract

The present invention discloses a multi-source data fusion method based on spatiotemporal dynamic triple matching analysis, and relates to the field of information fusion. The method comprises obtaining three independent observation grid data sets for describing the same object; performing spatiotemporal matching on the three data sets, taking one of the data sets as a reference, and adjusting the spatiotemporal dynamic scales of the other two data sets by using a moving window combined with a triple matching analysis technique; then, based on the premise of minimizing the spatiotemporal dynamic random error of the fused data set, the least square method is used to solve the optimal spatiotemporal dynamic fusion weights of the three data sets, and the three scale-adjusted data sets are respectively subjected to spatiotemporal dynamic linear weighting and summed with their optimal fusion weights, and finally the fused data set is obtained. The present invention simultaneously takes into account the temporal and spatial variability of the random error of the observation data set, and the accuracy and quality of the obtained fused data result are improved compared with the single original data set.

Description

Technical Field

[0001] The invention belongs to the field of information fusion, and in particular relates to a multi-source data fusion method based on spatiotemporal dynamic triple matching analysis. Background Art

[0002] For any variable that needs to be observed, a single observation method always has various uncertainties and errors. By fusing data sets under different observation methods, the advantages of different data sources can be fully combined to reduce data uncertainty. For spatiotemporal dynamic variables, the traditional one-dimensional fusion method only considers the error variability of the observation data set in space (time), and assumes that the error of the data set in time (space) is constant. Therefore, considering the variability of the random error of the observation data set in time and space at the same time, and fusing different observation data sets in time and space is an effective method to obtain high-precision observation data sets. Summary of the invention

[0003] The present invention provides a multi-source data fusion method based on spatiotemporal dynamic triple matching analysis, thereby solving the above-mentioned problems in the prior art.

[0004] The specific technical solutions adopted by the present invention are as follows:

[0005] The present invention provides a multi-source data fusion method based on spatiotemporal dynamic triple matching analysis, comprising the following steps:

[0006] S1: Obtain three observation data sets used to describe the same object; the three observation data sets have the same spatiotemporal resolution and spatiotemporal coverage, are in raster format, and are independent of each other;

[0007] S2: Performing spatiotemporal matching on the three observation data sets obtained in S1 to obtain three spatiotemporal matching data sets; designating one of the spatiotemporal matching data sets as a reference data set, and dynamically adjusting the time series observation values ​​of the other two spatiotemporal matching data sets based on the time series observation values ​​of the reference data set to obtain two dynamically adjusted data sets; and calculating the random errors of the two dynamically adjusted data sets.

[0008] S3: Based on the least squares method, for each raster pixel of the reference dataset and the two dynamic scale adjustment datasets at each moment, the optimal fusion weights of the pixel values ​​of the reference dataset and the pixel values ​​of the two dynamic scale adjustment datasets are calculated respectively;

[0009] S4: For each raster pixel of the reference dataset and the two dynamic scale adjustment datasets at each moment, the pixel value of the reference dataset and the pixel value of the two dynamic scale adjustment datasets are multiplied by their respective optimal fusion weights, and the multiplication results are added to obtain the fused pixel value; then the final fused time series raster dataset is obtained, recorded as the fused dataset, and the minimum random error of the fused dataset is obtained.

[0010] Preferably, the three observation data sets are independent of each other, which means that the data sources and observation methods of the three observation data sets are different, the data sets do not contain each other, and the same other data sources are not used in the production process of the data sets.

[0011] Preferably, the step S2 is specifically as follows:

[0012] S21: Assume that the time resolution of the three space-time matching datasets is x, and the time series length is nx; for each corresponding grid pixel of the three space-time matching datasets, let a t 、b t 、c t are the observation values ​​of the three spatiotemporal matching data sets at time t, and θ t is the actual true value at the tth time; where t∈[1x, 2x, 3x, …, nx];

[0013] A moving window with a window size of 101x and a step size of x is used, with time m as the window center, m∈[51x, 52x, 53x, …, (n-50)x], and the observation values ​​a=[a m-50x , a m-49x , a m-48x , …, a m , …, a m+50x ]、b=[b m-50x , b m-49x , b m-48x , …, b m , …, b m+50x ]、c=[c m-50x , a m-49x , c m-48x , …, c m , …, c m+50x ], and the triple collocation analysis (TCA) method is used to establish the relationship between a, b and c and the actual true value of the time series θ = [θ m-50x ,θ m-49x ,θ m-48x ,…,θ m ,…,θ m+50xThe linear relationship between them is as follows:

[0014] a=α a θ+β a +ε a (1)

[0015] b=α b θ+β b +ε b (2)

[0016] c=α c θ+β c +ε c (3)

[0017] Among them, α a , α b and α c are the multiplicative systematic errors of a, b and c relative to θ, β a , β b and β c are the additive systematic errors of a, b and c relative to θ, ε a , ε b and ε c are the random errors of the time series of a, b and c relative to θ; and the above linear equation satisfies three assumptions: (1) the three random errors of the time series ε a , ε b and ε c The arithmetic mean of is 0, that is, (2)ε a , ε b and ε c The two are independent of each other, and the covariance between them is 0, that is, (σ is the covariance operator); (3)ε a , ε b and ε c are independent of θ and have a covariance of 0, that is,

[0018] S22: Calculate the covariance σ between a, b and c using formulas (1) to (3) ij and ε a , ε b and ε c The respective variances are as follows:

[0019]

[0020]

[0021] Among them, i, j, k ∈ [a, b, c], and i ≠ j ≠ k, σ 2is the variance operator;

[0022] S23: According to formula (4), taking the time series observation value a of the reference data set as the reference, calculate b at time m m and c m Relative to a m The scale factor rb m and rc m , as follows:

[0023]

[0024]

[0025] S24: According to the scale factor rb obtained in S23 m and rc m , dynamically rescale the time series observations b and c of the other two spatiotemporal matching data sets; that is, for each moment m, calculate the observation value b′ after dynamic rescaling m and c′ m , as follows:

[0026]

[0027]

[0028] Then, for each grid pixel, the time series observation values ​​of the two dynamic scale adjustment data sets after dynamic scale adjustment are obtained, b′=[b′ 51x , b′ 52x , b′ 53x ,…,b′ (n-50)x ] and c′=[c′ 51x , c′ 52x , c′ 53x ,…,c′ (n-50)x ];

[0029] S25: Combine formulas (2), (3), (8) and (9) to establish the linear relationship between b′ and c′ and the actual true value θ of the time series, and calculate the random error ε′ corresponding to b′ and c′ b and ε′ c , as follows:

[0030] b′=rb*θ+β′ b +ε′ b (10)

[0031] c′=rc*θ+β′ c +ε′ c (11)

[0032] ε b′=ε b rb 2 (12)

[0033] ε c ′=ε c rc 2 (13)

[0034] Where rb=[rb 51x , rb 52x , rb 53x ,…,rb (n-50)x ] and rc = [rc 51x , rc 52x , rc 53x ,…,rc (n-50)x ] is the time series scale coefficient, β′ b and β′ c is the additive systematic error after dynamic scaling.

[0035] Furthermore, the step S3 is specifically as follows:

[0036] S31: Let the time series fusion value μ = [μ m-50x , μ m-49x , μ m-48x ,…,μ m ,…,μ m+50x ] and the actual true value θ of the time series satisfy the linear relationship and three assumptions described in step S21, as follows:

[0037] μ=α u θ+β u +ε u (14)

[0038] Among them, α u is the multiplicative error of μ relative to θ, β u is the additive error of μ relative to θ, ε u is the random error of μ relative to θ in the time series, and the fusion value μ at the mth moment is obtained by linear weighting. m , m∈[51x, 52x, 53x, …, (n-50)x], as follows:

[0039] μ m =τ am a m +τ bm b′ m +τ cm c′ m (15)

[0040] τ am +τ bm +τ cm =1 (16)

[0041] Among them, τ am , τ bm and τ cm are respectively a at time m m , b′ m and c′ m The fusion weight of

[0042] S32: Combine formulas (1), (10), (11), (14) and (15) to establish the time series random error ε of the time series fusion value μ u The linear relationship between the time series random errors of the reference dataset and the two dynamically scaled datasets is as follows:

[0043] ε u =τ a ε a +τ b ε′ b +τ c ε′ c (17)

[0044] where τ a , τ b and τ c is the time series fusion weight [τ k(51x) , τ k(52x) , τ k(53x) ,…τ k((n-50)x )], k∈[a,b,c];

[0045] S33: According to formula (17) and assumption 2) in step S21, the time series random error variance of the time series fusion value μ is The time series fusion value μ should have the smallest random error ε u , so the cost function F is set according to formula (16), as follows:

[0046]

[0047] Based on the least squares method, the function F is minimized to obtain the optimal fusion weight, that is, a , τ a and τ c Find the partial derivative of F and set it equal to 0 to find the minimum value of F, as follows:

[0048]

[0049]

[0050] S34: Combine formulas (5) to (7), (12), (13), (18) to (20) to solve the optimal fusion weight τ of the time series a , τ b and τ c , as follows:

[0051]

[0052]

[0053]

[0054] Furthermore, the step S4 is specifically as follows:

[0055] S41: Combine formulas (6) to (9), (15), (17), (21) to (23) to obtain the time series fusion value μ of each pixel, and then obtain the final fused time series raster dataset. At this time, μ has the minimum time series random error

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] Based on the idea of ​​least squares method, the present invention proposes a spatiotemporal dynamic triple matching analysis technology, which can simultaneously consider the spatiotemporal and spatial variability of observed variables and their errors, and dynamically fuse three independent data sets from different sources that observe the same variable. The fused data set obtained by the method of the present invention has high accuracy and quality, and its spatiotemporal error and uncertainty are lower than those of a single data set. Therefore, the present invention has certain theoretical and practical significance and promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Box plots of the temporal correlation R between the fusion product (UnionProduct) and the three single products and the ground observed soil moisture (a), frequency distribution histogram of the difference between the temporal correlation of the fusion product and the temporal correlation of SMOS-IC (△R_SMOS-IC) (b), frequency distribution histogram of the difference between the temporal correlation of the fusion product and the temporal correlation of ASCAT (△R_ASCAT) (c), frequency distribution histogram of the difference between the temporal correlation of the fusion product and the temporal correlation of GLDAS (△R_GLDAS) (d);

[0059] Figure 2Figure 2 shows the spatial distribution of the difference between the random error of SMOS-IC and the random error of the fused product (a), the spatial distribution of the difference between the random error of ASCAT and the random error of the fused product (b), and the spatial distribution of the difference between the random error of GLDAS and the random error of the fused product (c), where the black area indicates that the fused product has a smaller random error. DETAILED DESCRIPTION

[0060] The present invention is further described and illustrated below in conjunction with the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0061] Example

[0062] The global land surface area is selected as the study area, and the SMOS (the Soil Moisture and Ocean Salinity)-IC (INRA-CESBIO) passive microwave remote sensing surface soil moisture dataset, the ASCAT (the Advanced Scatterometer) active microwave remote sensing surface soil moisture dataset and the GLDAS (the Global Data Assimilation System) surface model simulation surface soil moisture dataset from May 24, 2010 to November 30, 2020 are used. The spatiotemporal dynamic triple matching analysis technology of the present invention is used to fuse these three datasets. The specific steps are as follows:

[0063] Step 1) Data acquisition: Obtain the SMOS-IC dataset, ASCAT dataset and GLDAS dataset of the study area. The spatial resolution of SMOS-IC and GLDAS datasets is 0.25°, the spatial resolution of ASCAT dataset is 12.5km, the temporal resolution of SMOS-IC and ASCAT is 1 day, and the temporal resolution of GLDAS is 3 hours. SMOS-IC is a surface soil moisture product based on the L-band microwave radiometer carried by the SMOS satellite launched by the European Space Agency. It is obtained by inverting the surface radiation brightness temperature data. It can be obtained through the official website link https: / / ib.remote-sensing.inrae.fr / The second version of the product is used here. ASCAT is a surface soil saturation moisture product, based on the C-band synthetic aperture radar carried by the Metop meteorological satellite launched by the European Space Agency, and inverted using the radar backscatter coefficient. You can find it through the official website link https: / / hsaf.meteoam.it / soil-moisture.phpGet. The Climate Data Record 2019 version (abbreviated as H115 and 116) descending orbit dataset is used here. GLDAS is a surface model simulation product based on the Global Land Data Assimilation System, which provides key atmospheric and surface variable information from 2000 to the present. The surface soil moisture dataset from the Noah model 0-10 cm is used here, which can be found on the official website link https: / / daac.gsfc.nasa.gov / The formats of the three datasets are all NetCDF.

[0064] Step 2) Data preprocessing and spatiotemporal matching: Use matlab software to batch read the three datasets in NetCDF format and convert them into tiff format output; process the temporal resolution of the GLDAS dataset to 1 day, that is, accumulate and average the corresponding pixel values ​​on 9 GLDAS grids every day as the pixel value of the day; use the bilinear resampling method to process the spatial resolution of ASCAT to 25°; use the coordinate conversion tool provided by the Arcpy toolkit to batch convert the coordinate systems of the three datasets to the GCS_WGS_1984 geographic coordinate system; adjust the reference grids of the SMOS-IC and ASCAT datasets to be the same as the reference grid of GLDAS. Through the above operations, ensure that the format, spatiotemporal resolution, coordinate system and grid of the three datasets are the same for subsequent processing.

[0065] Step 3) Quality control: For the SMOS-IC dataset obtained by step 2), the product's built-in "Scene_Flags" and "TB-RMSE" fields are used to remove pixels with complex terrain, snow and ice areas, and areas with severe radio frequency interference to ensure the quality of the dataset to be fused. The removal condition is "Scene_Flags" > 1 or "TB-RMSE" > 8K. Similarly, for the ASCAT dataset obtained by step 2), the product's built-in "sm_noise", "ssf", and "proc_flag" fields are used to remove pixels with abnormal and large uncertain inversion values ​​and frozen area pixels. The removal condition is "sm_noise" ≥ 1000 or "ssf" ≥ 3 or "proc_flag" ≠ 0. For the GLDAS dataset obtained by step 2), the product's built-in surface temperature "Surface Skin Temperature" and snow depth "Snow The snow and ice pixels in the GLDAS dataset itself, SMOS-IC and ASCAT datasets are further removed by using the "Surface Skin Temperature" field. The removal condition is "Surface Skin Temperature" < 273.15K or "Snow Depth" > 0. For the SMOS-IC dataset processed by the above steps, the pixel values ​​outside the interval [0, 1] are assigned to null values. For the ASCAT and GLDAS datasets processed by the above steps, the pixel values ​​outside the interval [0, 100] are assigned to null values ​​to ensure that the pixel values ​​of all datasets are within a reasonable range.

[0066] Step 4) Dynamic scale adjustment: For each common pixel of the three data sets obtained in step 3), the common time series pixel values ​​of the three data sets are obtained to form a 3-row n-column matrix M (n is the length of the time series), where the 1st, 2nd and 3rd rows are the time series observation values ​​of SMOS-IC, ASCAT and GLDAS respectively. A moving window with a window size of 101 days and a step length of 1 day is used, and the kth column (51≤k≤n-50) in the matrix M is used as the window center. The submatrix m (M (1:3, k-50:k+50)) composed of the 3 rows and 101 columns (from the k-50th column to the k+50th column) in the matrix M within the window range is used to establish a linear relationship r using the triple matching analysis technique (Triple Collocation Analysis). i =α i θ+β i +ε i , where i∈[1, 2, 3]; r i is the vector m(i,:) consisting of the i-th row data of matrix m; α i Yes iThe multiplicative systematic error of the true soil moisture value θ relative to its corresponding time series; β i Yes i Additive systematic error relative to θ; ε i Yes i The random error of the time series relative to θ; and i≠j (σ is the covariance operator) and Calculate r according to the above equation i and r j The covariance of and ε i Variance Among them, k∈[1, 2, 3], and i≠j≠k, σ 2 is the variance operator. On this basis, taking the SMOS-IC time series observation value r1 (i=1) as a reference, the ASCAT and GLDAS observation values ​​r for each day are calculated. 2d (j=2) and r 3d (k=3) relative to the observed value r of SMOS-IC every day 1d The scale factor and Among them, d∈[k-50,k+50]. Then according to the calculation, we can get s 2d and 3d , calculate the rescaled ASCAT observations for each day and GLDAS observations Then, for each common grid pixel, the time series observations r′2 of ASCAT and r′3 of GLDAS after scale adjustment are obtained. Finally, combining all the above equations, the linear relationship between r′2 and r′3 and θ is established. p =s p *θ+β′ p +ε′ p , and calculate the time series random errors of ASCAT and GLDAS after rescaling Among them, p∈[2,3], s p is the time series scale coefficient and β′ p is the additive systematic error after scaling.

[0067] Step 5) Calculation of spatiotemporal dynamic fusion weights: Refer to step 4) and establish the linear equation of the time series fusion values ​​μ and θ, that is, μ = α u θ+β u +ε u , where α u is the multiplicative error of μ relative to θ; β u is the additive error of μ relative to θ; ε uis the random error of μ relative to θ in the time series. Using linear weighting, for each grid pixel, the fusion value μ for each day is obtained. d =τ 1d r 1d +τ 2d r′ 2d +τ 3d r′ 3d , where τ 1d +τ 2d +τ 3d =1,τ 1d , τ 2d and τ 3d Each day 1d , r′ 2d and r′ 3d Based on step 4) and the above, the time series random error variance of the fusion product (i.e., the time series fusion value) μ is established. The linear relationship between the time series random error variance of the three observation data sets Where τ1, τ2 and τ3 are the time series fusion weights of SMOS-IC, ASCAT and GLDAS respectively. The optimal fusion value μ for each day is d Should have the smallest random error ε ud , that is, ε u When there is a minimum value, the weight at this time is the optimal fusion weight, so let Combining all the above formulas, the fusion weight of SMOS-IC can be obtained ASCAT fusion weights and GLDAS fusion weight

[0068] Step 6) Spatiotemporal dynamic linear weighting: For each pixel, the ASCAT and GLDAS datasets after dynamic scale adjustment in step 4) and the SMOS-IC dataset obtained in step 3) are linearly weighted with their respective dynamic fusion weights obtained in step 5) to obtain the time series surface soil moisture fusion value of each pixel.

[0069] Step 7) Verification of the accuracy of the fused dataset: The ground station data from the International Soil Moisture Network (ISMN) was used to verify the accuracy of the fused dataset. The Pearson correlation coefficient R was calculated to characterize the temporal correlation between the product and the ground-observed soil moisture, reflecting the product's ability to capture the temporal dynamic changes of ground-observed soil moisture. At the same time, the spatial distribution of the average random errors of the fused dataset and the three original datasets was compared and analyzed.

[0070] like Figure 1As shown in (a), the temporal correlation R between the fusion product and the ground-based soil moisture is significantly higher than that of ASCAT, and slightly higher than that of SMOS-IC and GLDAS. Figure 1 (bd) shows the frequency distribution of the time correlation difference between the fused product and the single product. It can be seen that for the three single products, the time correlation difference is mostly greater than 0, that is, the time correlation of the fused product is higher than that of the single product at most ground stations, indicating that the ability of the fused product to capture the temporal dynamic changes of ground observed soil moisture is better than that of the single product. Figure 2 The spatial distribution of the random error difference between the fusion product and the single product is shown, where the black area indicates that the fusion product has a smaller random error, and the gray area indicates that the single product has a smaller random error. Compared with SMOS-IC, the fusion product has smaller random errors in southern and central America, southern South America, southern Africa and the Sahara Desert, and most of Australia. Overall, the global average value of the random error difference is -0.00048125, close to 0, indicating that the fusion product and SMOS-IC have similar random errors overall; compared with ASCAT and GLDAS, the fusion product has smaller random errors in most parts of the world, its uncertainty is significantly reduced, and its accuracy is significantly improved.

[0071] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A multi-source data fusion method based on spatiotemporal dynamic triple matching analysis, characterized in that: The following steps are involved: S1: Obtain three observation data sets used to describe the same object; the three observation data sets have the same spatiotemporal resolution and spatiotemporal coverage, are in raster format, and are independent of each other; S2: Performing spatiotemporal matching on the three observation data sets obtained in S1 to obtain three spatiotemporal matching data sets; designating one of the spatiotemporal matching data sets as a reference data set, and dynamically adjusting the time series observation values ​​of the other two spatiotemporal matching data sets based on the time series observation values ​​of the reference data set to obtain two dynamically adjusted data sets; and calculating the random errors of the two dynamically adjusted data sets. S3: Based on the least squares method, for each raster pixel of the reference dataset and the two dynamic scale adjustment datasets at each moment, the optimal fusion weights of the pixel values ​​of the reference dataset and the pixel values ​​of the two dynamic scale adjustment datasets are calculated respectively; S4: For each raster pixel of the reference dataset and the two dynamic scale adjustment datasets at each moment, the pixel value of the reference dataset and the pixel value of the two dynamic scale adjustment datasets are multiplied by their respective optimal fusion weights, and the multiplication results are added to obtain the fused pixel value; then the final fused time series raster dataset is obtained, recorded as the fused dataset, and the minimum random error of the fused dataset is obtained.

2. The multi-source data fusion method based on spatiotemporal dynamic triple matching analysis according to claim 1 is characterized in that: The three observation data sets are independent of each other in that the data sources and observation methods of the three observation data sets are different, the data sets do not contain each other, and the same other data sources are not used in the production process of the data sets.

3. The multi-source data fusion method based on spatiotemporal dynamic triple matching analysis according to claim 1 is characterized in that: The step S2 is specifically as follows: S21: Assume that the time resolution of the three space-time matching datasets is x, and the time series length is nx; for each corresponding grid pixel of the three space-time matching datasets, let a t 、b t 、c t are the observation values ​​of the three spatiotemporal matching data sets at time t, and θ t is the actual true value at the tth time; where t∈[1x, 2x, 3x, …, nx]; A moving window with a window size of 101x and a step size of x is used, with time m as the window center, m∈[51x, 52x, 53x, …, (n-50)x], and the observation values ​​a=[a m-50x , a m-49x , a m-48x , …, a m , …, a m+50x ]、b=[b m-50x , b m-49x , b m-48x , …, b m , …, b m+50x ]、c=[c m-50x , a m-49x , c m-48x , …, c m , …, c m+50x ], and the triple matching analysis method is used to establish the relationship between a, b and c and the actual true value of the time series θ = [θ m-50x ,θ m-49x ,θ m-48x ,…,θ m ,…,θ m+50x The linear relationship between them is as follows: a=a a θ+β a +e a (1) b=a b θ+β b +e b (2) c=a c θ+β c +e c (3) Among them, α a , α b and α c are the multiplicative systematic errors of a, b and c relative to θ, β a , β b and β c are the additive systematic errors of a, b and c relative to θ, ε a , ε b and ε c are the random errors of the time series of a, b and c relative to θ respectively; and the above linear equation satisfies three assumptions: 1) 2) σ is the covariance operator; 3) S22: Calculate the covariance σ between a, b and c using formulas (1) to (3) ij and ε a , ε b and ε c The respective variances are as follows: Among them, i, j, k ∈ [a, b, c], and i ≠ j ≠ k, σ 2 is the variance operator; S23: According to formula (4), taking the time series observation value a of the reference data set as the reference, calculate b at time m m and c m Relative to a m The scale factor rb m and rc m , as follows: S24: According to the scale factor rb obtained in S23 m and rc m , dynamically rescale the time series observations b and c of the other two spatiotemporal matching data sets; that is, for each moment m, calculate the observation value b′ after dynamic rescaling m and c′ m , as follows: Then, for each grid pixel, the time series observation values ​​of the two dynamic scale adjustment data sets after dynamic scale adjustment are obtained, b′=[b′ 51x , b′ 52x , b′ 53x ,…,b′ (n-50)x ] and c′=[c′ 51x , c′ 52x , c′ 53x ,…,c′ (n-50)x ]; S25: Combine formulas (2), (3), (8) and (9) to establish the linear relationship between b′ and c′ and the actual true value θ of the time series, and calculate the random error ε′ corresponding to b′ and c′ b and ε′ c , as follows: b′=rb*θ+β′ b +e′ b (10) c′=rc*θ+β′ c +e′ c (11) e b ′=e b rb 2 (12) e c ′=e c rc 2 (13) Where rb=[rb 51x , rb 52x , rb 53x ,…,rb (n-50)x ] and rc = [rc 51x , rc 52x , rc 53x ,…,rc (n-50)x ] is the time series scale coefficient, β′ b and β′ c is the additive systematic error after dynamic scaling.

4. The multi-source data fusion method based on spatiotemporal dynamic triple matching analysis according to claim 3 is characterized in that: The step S3 is specifically as follows: S31: Let the time series fusion value μ = [μ m-50x , μ m-49x , μ m-48x ,…,μ m ,…,μ m+50x ] and the actual true value θ of the time series satisfy the linear relationship and three assumptions described in step S21, as follows: μ = a u θ+β u +e u (14) Among them, α u is the multiplicative error of μ relative to θ, β u is the additive error of μ relative to θ, ε u is the random error of μ relative to θ in the time series, and the fusion value μ at the mth moment is obtained by linear weighting. m , m∈[51x, 52x, 53x, …, (n-50)x], as follows: m m =t am a m +t bm b′ m +t cm c′ m (15) t am +t bm +t cm =1 (16) Among them, τ am , τ bm and τ cm are respectively a at time m m , b′ m and c′ m The fusion weight of S32: Combine formulas (1), (10), (11), (14) and (15) to establish the time series random error ε of the time series fusion value μ u The linear relationship between the time series random errors of the reference dataset and the two dynamically scaled datasets is as follows: e u =t a e a +t b e' b +t c e' c (17) where τ a , τ b and τ c is the time series fusion weight [τ k(51x) , τ k(52x) , τ k(53x) ,…τ k((n-50)x) ], k∈[a,b,c]; S33: According to formula (17) and assumption 2) in step S21, the time series random error variance of the time series fusion value μ is The time series fusion value μ should have the smallest random error ε u , so the cost function F is set according to formula (16), as follows: Based on the least squares method, the function F is minimized to obtain the optimal fusion weight, that is, a , τ a and τ c Find the partial derivative of F and set it equal to 0 to find the minimum value of F, as follows: S34: Combine formulas (5) to (7), (12), (13), (18) to (20) to solve the optimal fusion weight τ of the time series a , τ b and τ c , as follows:

5. The multi-source data fusion method based on spatiotemporal dynamic triple matching analysis according to claim 4 is characterized in that: The step S4 is specifically as follows: S41: Combine formulas (6) to (9), (15), (17), (21) to (23) to obtain the time series fusion value μ of each pixel, and then obtain the final fused time series raster dataset. At this time, μ has the minimum time series random error

Images