Structure-aware-driven long-time-sequence point-surface data fusion method and device

Abstract

The invention relates to the technical field of spatial information processing and data fusion, and discloses a long-time-sequence point-surface data fusion method and equipment driven by structure perception, and the method comprises the steps: extracting spatial structure information in rainfall products, elevation model products and slope data through a spatial feature extraction module; time sequence change features are modeled in combination with a sequence feature analysis module, so that the nonlinear relation between product data and actual measurement sites is comprehensively captured, and high-quality enhanced site data is generated. And finally, fusing actual measurement with an enhanced site, realizing high-precision conversion from long-time-sequence sparse observation data to continuous spatial information, effectively improving the spatial coverage capability and dynamic response capability of the data, and providing more continuous, accurate and credible data support for disaster monitoring and early warning of the mountain drainage basin.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spatial information processing and data fusion, and in particular to a structure perception driven long time series point surface data fusion method and device. BACKGROUND

[0002] In the field of geographic information processing, environmental monitoring and weather simulation, observation data is often collected in the form of stations. However, due to the limitations of terrain conditions, economic costs and operation and maintenance, the stations are generally sparse, uneven and even partially missing in space. This non-homogeneity not only limits the spatial coverage of observation data, but also poses a serious challenge to subsequent spatial analysis and modeling. Especially in the construction of high spatial resolution surface data (such as precipitation, temperature or land deformation, etc.), the traditional interpolation method often leads to a decrease in prediction accuracy and a blurring of the distribution of ground objects due to the inability to fully capture spatial heterogeneity and local feature changes, thereby affecting the reliability of the fine simulation and decision support system.

[0003] To improve the expression ability of sparse observation data, existing research attempts to introduce multi-source auxiliary data such as remote sensing images, digital elevation models (DEM) and meteorological reanalysis products into modeling, which serves as background information to improve the spatial continuity and rationality of the prediction results. However, this method has the following shortcomings: on the one hand, it fails to fully utilize the information support of observation points, resulting in a lack of enhanced modeling effect; on the other hand, for non-observation areas, there is a lack of effective completion mechanism, making it difficult to accurately restore the data in the missing areas. In particular, in mountainous areas and other regions with complex natural environment, these problems are more prominent. For example, in the monitoring of water and electricity disaster factors in a watershed, station data is one of the core information sources. However, due to the difficulty of construction in mountainous areas, complex environment and other reasons, the layout of observation stations is sparse and uneven, making it difficult to accurately capture key disaster factors such as sudden and local heavy rain, thereby restricting the identification and early warning ability of regional disaster risk.

[0004] Therefore, it is a key problem to be solved to improve the spatial continuity and dynamic adaptability of station data and to construct a high-precision precipitation expression model. SUMMARY

[0005] The present application provides a structure perception driven long time series point surface data fusion method and device, which aims to solve the problem of poor spatial continuity and dynamic adaptability of station data in the prior art, and the lack of high-precision precipitation expression model.

[0006] To achieve the above-mentioned purpose, the present application realizes the technical scheme as follows:

[0007] In a first aspect, the present application provides a structure perception driven long time series point surface data fusion method, comprising:

[0008] constructing a spatio-temporal training sample and a spatio-temporal test sample by using daily precipitation observation data of existing meteorological stations, daily precipitation product data and digital elevation model product data;

[0009] constructing an enhanced station generation model, the enhanced station generation model comprising a spatial feature enhancement perception module, a spatial feature extraction module and a sequence feature analysis module;

[0010] inputting the n-th group of data in the spatio-temporal training sample into the spatial feature enhancement perception module to obtain enhanced features, and processing the enhanced features through the spatial feature extraction module and the sequence feature analysis module to obtain predicted daily precipitation data;

[0011] determining an optimal enhanced station generation model based on the daily precipitation data and the spatio-temporal test sample;

[0012] obtaining enhanced station daily precipitation data based on the daily precipitation product data, the digital elevation model product data and the optimal enhanced station generation model;

[0013] merging the existing meteorological station daily precipitation observation data and the enhanced station daily precipitation data to obtain expanded mixed station daily precipitation data, and accumulating the expanded mixed station daily precipitation data on a monthly basis to obtain mixed station monthly precipitation data;

[0014] accumulating the daily precipitation product data on a monthly basis to obtain monthly precipitation product data, generating array points according to the spatial resolution Resolution of the monthly precipitation product data, extracting the values of the monthly precipitation product data into the array points to obtain point monthly precipitation product data;

[0015] mainly using the point monthly precipitation product data and supplemented by the mixed station monthly precipitation data, under the constraint of the spatial resolution Resolution, using a set algorithm to complete the spatial conversion of dense point monthly precipitation data to continuous surface monthly precipitation data, and realizing point-surface fusion.

[0016] In a second aspect, the present application further provides a computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of the first aspect when executing the computer program.

[0017] Advantages:

[0018] ​This invention provides a structure-aware-driven long-term point-area data fusion method. First, a two-dimensional Pascal array sensing mechanism is introduced through a spatial feature enhancement sensing module to construct an enhancement matrix with local weighting characteristics. This facilitates subsequent structured spatial enhancement of key areas in remote sensing images. Then, adaptive spatial importance analysis highlights the changing characteristics of complex terrain areas. Next, a spatial feature extraction module extracts spatial structure information from precipitation products, elevation model products, and slope data. Combined with a sequence feature analysis module, time-series variation characteristics are modeled to comprehensively capture the nonlinear relationship between product data and measured stations, generating high-quality enhanced station data. Finally, the measured and enhanced station data are fused to achieve high-precision conversion of long-term sparse observation data into continuous spatial information, effectively improving the spatial coverage and dynamic response capabilities of the data. This provides more continuous, accurate, and reliable data support for disaster monitoring and early warning in mountainous watersheds. Attached Figure Description

[0019] Figure 1 This is a flowchart of a structure-aware driven long-time-series point-area data fusion method provided by a preferred embodiment of the present invention. Detailed Implementation

[0020] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0021] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of the invention.

[0022] Unless otherwise specified, all raw materials, reagents, instruments and equipment used in this invention can be purchased from the market or prepared by existing methods.

[0023] Please see Figure 1 This application provides a structure-aware driven long-term time-series point-area data fusion method, including:

[0024] Step 1: Construct spatiotemporal training samples and spatiotemporal test samples using existing daily precipitation observation data, daily precipitation product data, and digital elevation model product data from meteorological stations;

[0025] Step 2: Construct an enhanced site generation model, which includes a spatial feature enhancement perception module, a spatial feature extraction module, and a sequence feature analysis module;

[0026] Step 3: Extract the data from the nth group of the spatiotemporal training samples. The enhanced features are obtained by inputting them into the spatial feature enhancement perception module. After the enhanced features are processed by the spatial feature extraction module and the sequence feature analysis module, the predicted daily precipitation data is obtained.

[0027] Step 4: Determine the optimal enhanced station generation model based on daily precipitation data and spatiotemporal test samples;

[0028] Step 5: Obtain enhanced station daily precipitation data based on daily precipitation product data, digital elevation model product data, and the best-performing enhanced station generation model;

[0029] Step 6: Combine the existing daily precipitation observation data from meteorological stations with the enhanced daily precipitation data from stations to obtain the expanded mixed daily precipitation data. Then, accumulate the expanded mixed daily precipitation data month by month to obtain the mixed monthly precipitation data.

[0030] Step 7: Accumulate the daily precipitation product data month by month to obtain monthly precipitation product data, generate array points according to the spatial resolution of the monthly precipitation product data; extract the values ​​of the monthly precipitation product data into the array points to obtain point-like monthly precipitation product data.

[0031] Step 8: Using point-based monthly precipitation product data as the main source and mixed station monthly precipitation data as a supplement, under the constraint of spatial resolution, a set algorithm is used to complete the spatial transformation of dense point-based precipitation data into continuous areal precipitation data, thereby achieving point-area fusion.

[0032] It is worth explaining that in this application, x represents the serial number; Data represents the obtained training sample data; gpm represents the data in the GPM product; dem represents the data in the DEM (Digital Elevation Model); and slope represents the slope data. These three represent data from three different sources.

[0033] The aforementioned structure-aware-driven long-term point-area data fusion method first introduces a two-dimensional Pascal array sensing mechanism through a spatial feature enhancement sensing module to construct an enhancement matrix with local weighting characteristics. This facilitates subsequent structured spatial enhancement of key areas in remote sensing images. Then, adaptive spatial importance analysis highlights the changing characteristics of complex terrain areas. Subsequently, a spatial feature extraction module extracts spatial structure information from precipitation products, elevation model products, and slope data. Combined with a sequence feature analysis module, time-series variation characteristics are modeled to comprehensively capture the nonlinear relationship between product data and measured stations, generating high-quality enhanced station data. Finally, the measured and enhanced station data are fused to achieve high-precision conversion of long-term sparse observation data into continuous spatial information, effectively improving the spatial coverage and dynamic response capabilities of the data. This provides more continuous, accurate, and reliable data support for disaster monitoring and early warning in mountainous watersheds.

[0034] Furthermore, the specific implementation of step 1 includes the following sub-steps:

[0035] Step 1-1: Calculate slope information: Read the DEM data Z using the GDAL third-party library in Python. For any location DEM data Z(a,b), obtain its 9-neighbor DEM data Z' in a 3×3 window. 3×3 , where a represents the row number and b represents the column number.

[0036]

[0037] First, calculate the row direction gradient S. a :

[0038]

[0039] Where Δa is the spatial resolution of the pixel in the row direction.

[0040] Then, calculate the column directional gradient S. b :

[0041]

[0042] Where Δb is the spatial resolution of the pixel in the row direction.

[0043] Finally, calculate the slope (in degrees):

[0044]

[0045] Steps 1-2, Data Preprocessing: It is worth noting that the daily precipitation observation data from the meteorological stations are in vector format.

[0046] First, daily precipitation data is read using the OGR third-party library in Python. in, This represents the precipitation data for the i-th station on day t. Let represent the precipitation data for the i-th station on day t. The upper limit of t is determined by the number of stations i. The upper limit of t is determined by the number of days of data t.

[0047] Then, according to Data Station The latitude and longitude coordinates {Lng} of the i-th station are stored in the middle. i ,Lat i Get the row and column numbers of precipitation product data, DEM data, and slope data for day t. By row and column number respectively Centered on the target area, and with a set width of 32 pixels and a height of 32 pixels, obtain the corresponding precipitation product slice data:

[0048]

[0049] DEM tile data

[0050] In the formula, This represents the daily precipitation product slice data for the i-th station on day t. This represents the digital elevation model product slice data for the i-th station.

[0051] slope slice data

[0052] Among them, Data Station Data gpm These consist of daily precipitation observation data from meteorological stations and precipitation product slice data, totaling i×t items; Data gpm-i Data slope-i These are the precipitation product slice data and the slope slice data of the i-th station, respectively.

[0053] Steps 1-3: Dataset Partitioning Station Data gpm The data is for each day, and its quantity is i×t; Data gpm-i Data slope-i The quantity is i (spatial dimension), and the Data... gpm-i Data slope-i By replicating t copies to achieve alignment and expansion on the time scale, we obtain:

[0054]

[0055] In the formula, This represents the elevation model product slice data for the i-th site on day t. This represents the slope slice data of the i-th station on the t-th day.

[0056] {Data Station Data gpm Data dem Data slope The spatiotemporal samples D were obtained by randomly dividing the samples in a 7:3 ratio. Train and D Test D Train As spatiotemporal training samples, D Test This is a spatiotemporal test sample.

[0057] In this application, by calculating the slope of the DEM and extracting precipitation products, DEM and slope slices, the precise spatial and temporal alignment of multi-source data is achieved. The static features of the terrain and the dynamic information of daily precipitation are integrated into a complete spatiotemporal sample. The scientific nature of the model training and the generalization ability of the prediction are ensured by dividing the training set and the test set, thereby providing a reliable and information-rich input basis for subsequent enhancement of site generation and interpolation.

[0058] It is worth noting that the spatial feature enhancement perception module of the enhanced site generation model includes a two-dimensional Pascal array enhancement layer, a layer normalization layer, an importance analysis layer, a first convolutional layer, and a Sigmoid activation function. The spatial feature extraction module consists of a second convolutional layer, a first ReLU activation function, a first max pooling layer, a third convolutional layer, a first ReLU activation function, and a first max pooling layer. Specifically, the second convolutional layer has 3 input channels, 16 output channels, a kernel size of 3, a stride of 1, and padding of 1; the third convolutional layer has 16 input channels, 32 output channels, a kernel size of 3, a stride of 1, and padding of 1. The sequence feature analysis module includes a first linear layer, a Long Short-Term Memory (LSTM) layer, and a second linear layer. Specifically, the first linear layer has an input size of 32 and an output size of 32; the LSTM layer has an input size of 32 and an output size of 64; and the second linear layer has an input size of 64 and an output size of 1.

[0059] In the specific working process, the two-dimensional Pascal array reinforcement layer constructs a 4×4 weight coefficient matrix W using Pascal triangles. yang According to the properties of Pascal's Triangle:

[0060]

[0061] Then, W yang Dividing by the maximum value yields the normalized weight coefficient matrix:

[0062]

[0063] Based on spatiotemporal sample D Train In the nth slice of data A weight coefficient matrix W, consisting entirely of 1s, is generated with a width of 32 pixels and a height of 32 pixels. one The dimensions are width × Height.

[0064] W' yang Add to W one At the center 4×4 position, the two-dimensional Pascal array reinforcement coefficient matrix was obtained. Will and Multiplying them together yields enhanced precipitation product slice data.

[0065] Enhanced precipitation product slice data Spatiotemporal test sample D Train In the nth group of data Enhanced hybrid data is obtained by splicing data along channels and then passing it through a layer normalization layer.

[0066] The importance analysis layer includes maximum, average, and minimum value calculations. This will enhance the analysis of mixed data. The maximum value is calculated according to the channel dimension to obtain the feature. The features are obtained by calculating the average value according to the channel dimension. The minimum value is calculated according to the channel dimension to obtain the feature. Then, the features are concatenated according to the channel dimension to obtain the importance mixture feature.

[0067] Then, the importance of mixed features is combined. The input is fed into the second convolutional layer for feature extraction, resulting in spatial features F. 2 The second convolutional layer has 9 input channels, 3 output channels, a kernel size of 1, and a stride of 1.

[0068] Next, the spatial feature F 2 The feature probability map F is obtained by using the Sigmoid activation function. 2-1 .

[0069] Finally, the feature probability map F 2-1 With enhanced hybrid data Element-wise multiplication yields the enhanced feature F3.

[0070] Furthermore, the enhanced feature F3 is input into the second convolutional layer to obtain spatial features. Then spatial features The input is fed into the first ReLU activation function to obtain spatial features. Next, spatial features Spatial features are obtained after the first max pooling layer.

[0071] Finally, spatial features The input is fed into the third convolutional layer to obtain spatial features. Then, spatial features The input is fed into the second ReLU activation function to obtain spatial features. Next, spatial features Spatial features are obtained after the second maximum pooling layer.

[0072] In one example, The dimensions are 16×32×32. The dimensions are 16×16×16. The dimensions are 32×16×16. The dimensions are 32×8×8.

[0073] Furthermore, spatial features Expanding along the width and height dimensions yields the sequence features. Its size is 32×64. Then, the sequence features... By exchanging channels and expanding the width and height dimensions, sequence features are obtained. Its dimensions are 64×32. Next, sequence features... The input is fed into the first linear layer to obtain sequence features. Next, sequence features The input is fed into an LSTM layer to obtain sequence features. sequence features The average value is calculated along the channel dimension and then input into the second linear layer to obtain the predicted daily precipitation data F. P .

[0074] In one example, The dimensions are 32×32. The dimensions are 64×64, F P The size is 1.

[0075] Furthermore, the specific implementation of step 4 includes the following sub-steps:

[0076] Step 4-1: Calculate the predicted daily precipitation data F using the mean square loss function MSELoss. P With the xth group of data The losses between them.

[0077] Step 4-2: Iteratively train the augmented site generation model using the Adam optimizer and MSELoss. After training, use D... Test The optimized augmented site generation model was evaluated and compared. The parameter weights with the smallest root mean square value were selected as the best-performing augmented site generation model. This ensures efficient and stable training while selecting model parameters with the strongest generalization ability and most reliable prediction results, consistent with the evaluation metrics.

[0078] Furthermore, the specific implementation of step 5 includes the following sub-steps:

[0079] Step 5-1: Randomly generate N stations to be augmented within the spatial range of daily precipitation product data and DEM product data. p , where p∈{1,2,3,…,N}. According to Station p The latitude and longitude coordinates are calculated from the row and column numbers of daily precipitation product data, DEM product data, and slope data, according to the width = 32 pixels and the height = 32 pixels, to obtain the Station. p Corresponding slice data In the formula, This represents the precipitation product slice data on day t for the p-th proposed enhancement site. This represents the elevation model product slice data for the p-th site to be enhanced. This represents the slope slice data of the p-th station to be reinforced.

[0080] Step 5-2, The best-performing augmented site generation model is used to obtain daily precipitation data for N augmented sites.

[0081] Furthermore, the specific implementation of step 8 includes the following sub-steps:

[0082] Step 8-1: Extract point-based monthly precipitation product data The coordinates of each point are used to obtain coordinate data C. gpm Extract monthly precipitation data from mixed stations. The coordinates of each point are used to obtain coordinate data C. Station .

[0083] Then, the coordinate data is standardized to obtain the standardized coordinates C'. gpm and C' Station .

[0084]

[0085] Next, the precipitation data is normalized to obtain the normalized point-based monthly precipitation product data, Data'. gpm Normalized mixed site monthly precipitation data Station .

[0086]

[0087] In subsequent interpolation:

[0088]

[0089] Step 8-2: Calculate the parameters of the covariance model using the least squares method:

[0090]

[0091] Wherein, C(x) i ,x j ) represents x i and x j The covariance between them, x i and x j All represent known points, where i and j take values ​​of 1, 2, ..., n; and i ≠ j. C0 is the gap effect, C1 and a are the fitting parameters for the covariance, and ||x i -x j || represents x i and x j The Euclidean distance between them.

[0092] Step 8-3: For each unobserved point x0 and n known points {x1, x2, x3, ..., x...} n First, construct an n×n covariance matrix to represent the covariance of the known points:

[0093]

[0094] Then, construct an n×1 vector to represent the covariance between known points and unobserved points, specifically C(x i ,x0), where i=1,2,…,n.

[0095]

[0096] Next, the solution is obtained using the weight calculation formula:

[0097]

[0098] λ=C -1 ·C0;

[0099] In the formula, λ={λ1,λ2,…,λ iThe} represents the weighting coefficients calculated based on the covariance model. These are used to measure the influence of each known observation point on the target interpolation point.

[0100] Step 8-4: Interpolate for each unobserved point to obtain the interpolation result Z(x0) in the normalized space. Reconstruct the interpolation result into raster data according to the specified spatial resolution.

[0101]

[0102] In the formula, λ i Z(x) represents the weights calculated based on the covariance model. i ) represents the normalized observation (from GPM or site data).

[0103] In this system, raster data is in matrix form, and each element (point) in the matrix is ​​considered an unknown point. This formula can be used to extrapolate the results for any unknown point. Once all points are calculated, the entire interpolated raster data is obtained.

[0104] In this implementation, by weighted averaging of the observations of known points, the optimal estimate of unobserved points can be obtained under the constraint of minimum variance unbiasedness. This ensures that the interpolation results fully utilize neighboring information while guaranteeing the accuracy and reliability of spatial prediction. It can fully utilize multi-source observation information at a uniform scale, obtaining optimal weights through covariance constraints, thereby achieving stable, highly generalizable, and physically consistent precipitation interpolation results.

[0105] The method of this application will be described and verified below with a specific experiment:

[0106] The study area, encompassing a county in a certain city and its surrounding region (96.374°W, 33.684°N – 102.766°W, 28.696°N), was selected. Daily precipitation data from November 2018 (GPM) and daily precipitation observation data from 10 meteorological stations within the region were used as experimental data. The proposed method was compared with Kriging interpolation and a point-area fusion method based on co-Kriging. Experimental results show that the root mean square error (RMSE) of the proposed method is reduced to 5.392 mm, significantly better than the comparative methods. This indicates that the proposed method has a higher ability to characterize the spatial distribution of precipitation under sparse observation conditions, effectively improving fusion accuracy and spatial continuity, and has strong adaptability and promotional value for precipitation estimation in complex terrain areas.

[0107]

[0108] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method. This computer device can implement various embodiments of the structure-aware driven long-time-series point-area data fusion method described above, and can achieve the same beneficial effects, which will not be elaborated here.

[0109] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A structure-aware driven long-term time-series point-area data fusion method, characterized in that, include: Spatiotemporal training samples and spatiotemporal test samples are constructed using existing daily precipitation observation data from meteorological stations, daily precipitation product data, and digital elevation model product data. An enhanced site generation model is constructed, which includes a spatial feature enhancement perception module, a spatial feature extraction module, and a sequence feature analysis module. The nth group of data in the spatiotemporal training samples The enhanced features are obtained by inputting them into the spatial feature enhancement sensing module. These enhanced features are then processed by the spatial feature extraction module and the sequence feature analysis module to obtain the predicted daily precipitation data. For daily precipitation product data; For digital elevation model product data, Based on elevation model product data The calculated slope data; The optimal enhanced station generation model was determined based on daily precipitation data and spatiotemporal test samples. Enhanced station daily precipitation data are obtained based on daily precipitation product data, digital elevation model product data, and the best-performing enhanced station generation model. The daily precipitation data of existing meteorological stations are combined with the daily precipitation data of enhanced stations to obtain the expanded mixed station daily precipitation data. The expanded mixed station daily precipitation data are then accumulated month by month to obtain the mixed station monthly precipitation data. The daily precipitation product data is accumulated month by month to obtain the monthly precipitation product data. Array points are generated according to the spatial resolution of the monthly precipitation product data. The values ​​of the monthly precipitation product data are extracted into the array points to obtain the point-like monthly precipitation product data. Using primarily point-based monthly precipitation product data, supplemented by mixed station monthly precipitation data, and under the constraint of spatial resolution, a set algorithm is employed to complete the spatial transformation of dense point-based precipitation data into continuous areal precipitation data, thereby achieving point-area fusion.

2. The structure-aware driven long-time-series point-area data fusion method according to claim 1, characterized in that, The construction of spatiotemporal training and test samples using existing daily precipitation observation data from meteorological stations, daily precipitation product data, and digital elevation model product data includes: Calculate slope information: Read the digital elevation model (DEM) product data. For any location's DEM product data Z(a,b), obtain the 9-neighbor DEM product data Z using a 3×3 window. 3×3 Where 'a' represents the row index and 'b' represents the column index, as follows: In the formula, Z represents the digital elevation model product data; Calculate the gradient S in the row direction a ,as follows: Where Δa represents the spatial resolution of the pixel in the row direction; Calculate the gradient S in the column direction b ,as follows: Where Δb is the spatial resolution of the pixel in the row direction; Calculate the slope as follows: The daily precipitation data is as follows: in, This represents the precipitation data for the i-th station on the t-th day, where i takes the values ​​1, 2, ...; and t takes the values ​​1, 2, ...; According to Data Station The latitude and longitude coordinates {Lng} of the i-th station are stored in the middle. i ,Lat i } Obtain the row and column numbers of precipitation product data, DEM data, and slope data for day t. By row and column number respectively Centered on a target area, with a width of 32 pixels and a height of 32 pixels, the corresponding precipitation product slice data is obtained. gpm as follows: DEM tile data In the formula, This represents the daily precipitation product slice data for the i-th station on day t. This represents the digital elevation model product slice data for the i-th site; The slope slice data is as follows: Among them, Data Station Data gpm These consist of daily precipitation observation data from meteorological stations and precipitation product slice data, totaling i×t items; Data gpm-i Data slope-i Let each be a slice of precipitation product data and a slice of slope data for the i-th station, respectively. gpm-i Data slope-i By replicating t copies to achieve alignment and expansion on the time scale, we obtain: In the formula, This represents the elevation model product slice data for the i-th site on day t. This represents the slope slice data of the i-th station on day t. {Data Station Data gpm Data dem Data slope The spatiotemporal training samples D are randomly divided according to a set ratio. Train and spacetime test sample D Test .

3. The structure-aware driven long-time-series point-area data fusion method according to claim 1, characterized in that, The spatial feature enhancement perception module includes a two-dimensional Pascal array enhancement layer, a layer normalization layer, an importance analysis layer, a first convolutional layer, and a Sigmoid activation function.

4. The structure-aware driven long-time-series point-area data fusion method according to claim 3, characterized in that, The nth group of data in the spatiotemporal training samples The input to the spatial feature enhancement perception module yields enhanced features, including: The two-dimensional Pascal array enhancement layer constructs a 4×4 weight coefficient matrix W using Pascal triangles. yang as follows: W yang The normalized weight coefficient matrix is ​​obtained by dividing by its maximum value as follows: Based on spatiotemporal sample D Train In the nth slice of data A weight coefficient matrix W, consisting entirely of 1s, is generated with a width of 32 pixels and a height of 32 pixels. one The dimensions are Width × Height; W ' yang Add to W one At the center 4×4 position, the two-dimensional Pascal array reinforcement coefficient matrix was obtained. Will and Multiplying them together yields enhanced precipitation product slice data. Enhanced precipitation product slice data Spacetime test sample D Train In the nth group of data Enhanced hybrid data is obtained by splicing data along channels and then passing it through a layer normalization layer. The importance analysis layer will enhance the analysis of mixed data. The maximum value is calculated according to the channel dimension to obtain the feature. The features are obtained by calculating the average value according to the channel dimension. The minimum value is calculated according to the channel dimension to obtain the feature. Then, the features are concatenated according to the channel dimension to obtain the importance mixture feature. Mixed features of importance The input is fed into the first convolutional layer for feature extraction, resulting in spatial features F. 2 ; Spatial features F 2 The feature probability map F is obtained by using the Sigmoid activation function. 2-1 ; The feature probability map F 2-1 With enhanced hybrid data Element-wise multiplication yields the enhanced feature F3.

5. The structure-aware driven long-time-series point-area data fusion method according to claim 1, characterized in that, The spatial feature extraction module includes a second convolutional layer, a first ReLU activation function, a first max pooling layer, a third convolutional layer, a first ReLU activation function, and a first max pooling layer; the sequence feature analysis module includes a first linear layer, a long short-term memory (LSTM) network layer, and a second linear layer.

6. The structure-aware driven long-time-series point-area data fusion method according to claim 5, characterized in that, The process of processing the enhanced features through a spatial feature extraction module and a sequence feature analysis module to obtain predicted daily precipitation data includes: The enhanced feature F3 is input into the first convolutional layer to obtain the spatial features. Then spatial features The input is fed into the first ReLU activation function to obtain spatial features. spatial features Spatial features are obtained after the first max pooling layer. spatial features The input is fed into the second convolutional layer to obtain spatial features. spatial features The input is fed into the second ReLU activation function to obtain spatial features. spatial features Spatial features are obtained after the second maximum pooling layer. spatial features Expanding along the width and height dimensions yields the sequence features. sequence features By exchanging channels and expanding the width and height dimensions, sequence features are obtained. sequence features The input is fed into the first linear layer to obtain sequence features. sequence features The input is fed into the LSTM layer of a Long Short-Term Memory (LSTM) network to obtain sequence features. sequence features The average value is calculated along the channel dimension and then input into the second linear layer to obtain the predicted daily precipitation data F. P .

7. The structure-aware driven long-time-series point-area data fusion method according to claim 1, characterized in that, The optimal enhanced station generation model, determined based on daily precipitation data and spatiotemporal test samples, includes: The mean square loss function is used to calculate the loss between predicted daily precipitation data and daily precipitation observation data from meteorological stations. Based on the loss, the gradient backpropagation principle is used to train an enhanced station generation model, resulting in an optimized enhanced station generation model. After training, spatiotemporal test samples are used to evaluate and compare the performance of the optimized enhanced station generation model. The parameter weight with the smallest root mean square value is selected as the best-performing enhanced station generation model.

8. The structure-aware driven long-time-series point-area data fusion method according to claim 1, characterized in that, The enhanced station daily precipitation data, obtained based on daily precipitation product data, digital elevation model product data, and the best-performing enhanced station generation model, includes: N stations to be augmented are randomly generated within the spatial range of daily precipitation product data and digital elevation model product data. p Where p∈{1,2,3,…,N}; According to Station p The latitude and longitude coordinates of N stations to be augmented are calculated based on the row and column numbers of daily precipitation product data, digital elevation model product data, and slope data. Stations are then obtained according to a width of 32 pixels and a height of 32 pixels. p The corresponding slice data is as follows: In the formula, This represents the precipitation product slice data on day t for the p-th proposed enhancement site. This represents the elevation model product slice data for the p-th site to be enhanced. This represents the slope slice data for the p-th station to be reinforced; Slice data corresponding to N sites to be enhanced The best-performing augmented site generation model is used to obtain daily precipitation data for N augmented sites.

9. The structure-aware driven long-time-series point-area data fusion method according to claim 1, characterized in that, The method primarily uses point-based monthly precipitation data, supplemented by mixed station-based monthly precipitation data. Under the constraint of spatial resolution, a set algorithm is employed to spatially transform dense point-based precipitation data into continuous areal precipitation data, achieving point-area fusion. This includes: Extracting point-based monthly precipitation product data The coordinates of each point are used to obtain coordinate data C. gpm Extract monthly precipitation data from mixed stations. The coordinates of each point are used to obtain coordinate data C. Station ; For coordinate data C gpm and C Station After standardization, the standardized coordinates C' are obtained. gpm and C' Station ,as follows: Normalize the precipitation data to obtain normalized point-based monthly precipitation product data. gpm Normalized mixed site monthly precipitation data Station as follows: In subsequent interpolation: Calculate the parameters of the covariance model using the least squares method: Wherein, C(x) i ,x j ) represents x i and x j The covariance between them, x i and x j All represent known points, where i and j take values ​​of 1, 2, ..., n; and i ≠ j. C0 is the gap effect, C1 and a are the fitting parameters for the covariance, and ||x i -x j || represents x i and x j Euclidean distance between them; For each unobserved point x0 and n known points {x1, x2, x3, ..., xn} n First, construct an n×n covariance matrix to represent the covariance of the known points as follows: Then, construct an n×1 vector to represent the covariance between known points and unobserved points, specifically C(x i (x0), where i = 1, 2, ..., n, as follows: Solve using the weight calculation formula: λ=C -1 ·C0; In the formula, λ={λ1,λ2,…,λ i } represents the weighting coefficients calculated based on the covariance model; Interpolate for each unobserved point to obtain the interpolation result Z(x0) in the normalized space. Reconstruct the interpolation result into raster data according to the specified spatial resolution, as follows: In the formula, λ i Z(x) represents the weights calculated based on the covariance model. i ) represents the normalized observation value.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes a computer program, it implements the steps of any one of the methods of claims 1 to 9 above.

Images