A method for reconstructing multi-source precipitation based on multi-temporal-scale collaborative back-learning and its storage medium

Abstract

This invention proposes a multi-source precipitation reconstruction method and storage medium based on multi-temporal-scale collaborative back-learning, belonging to the field of hydrological and meteorological data processing and remote sensing big data fusion technology. It acquires daily dynamic meteorological features and static surface features of the study area, constructs single samples based on learning objectives, and integrates them into sample units according to time span. A U-Net encoding / decoding model is constructed, and the sample units are input into the model. After the decoder outputs reconstructed values ​​at 9km, 3km, and 1km resolutions, they are first aggregated along the time dimension, and then the loss functions at each temporal and spatial scale are calculated to construct the total loss. The Adam optimizer is used for iterative optimization until convergence to obtain the trained model. Precipitation reconstruction can be achieved by inputting the feature data of the year to be reconstructed. This method can overcome the spatiotemporal alignment constraints, achieve multi-resolution collaborative reconstruction, and solve the problem of multi-temporal-scale fusion.

Description

Technical Field

[0001] This invention belongs to the field of hydrological and meteorological data processing and remote sensing big data fusion technology, specifically involving a multi-source precipitation reconstruction method and storage medium based on multi-temporal and spatial scale collaborative reverse learning. Background Technology

[0002] With the continuous improvement of remote sensing technology and numerical simulation capabilities, the means of acquiring precipitation elements (such as daily precipitation and precipitation frequency) have become increasingly diversified. Currently, the mainstream precipitation data sources are mainly divided into three categories: 1) Satellite remote sensing products (such as the Global Precipitation Measurement Program multi-satellite joint inversion dataset IMERG and the multi-source weighted integrated precipitation dataset MSWEP), which can provide high-frequency observation services at the global scale, but have the shortcomings of low spatial resolution and limited spatial accuracy in complex terrain areas; 2) Reanalysis datasets (such as the European Centre for Medium-Range Weather Forecasts' fifth-generation land reanalysis dataset ERA5-Land), which have excellent temporal continuity and multivariate coordination, but their original spatial resolution is generally low (about 9 km), making it difficult to accurately characterize the local precipitation variability characteristics within the watershed; 3) Ground observation and fusion products, although they have high observation accuracy in local areas, generally suffer from spatial coverage discontinuity due to the limitation of ground station density.

[0003] The fusion and reconstruction of multi-source data is the core approach to obtaining reliable precipitation data. However, current multi-source precipitation data fusion and reconstruction technologies still face four key challenges: 1) Inconsistent spatiotemporal coverage: Traditional fusion methods (such as simple weighting, statistical downscaling, and conventional interpolation) usually require all input data sources to be completely aligned in the spatiotemporal dimensions. However, in practical applications, factors such as sensor lifespan, orbital variations, and sparse ground stations make it difficult to obtain a consistent input sequence across all time and space, resulting in a large amount of fragmented data that cannot be effectively utilized; 2) Conflicts in spatial resolution heterogeneity: Existing fusion models struggle to be compatible with heterogeneous spatial resolutions of different magnitudes (such as 9 km reanalysis data, 3 km regional fusion products, and 100 km regional fusion products) within a unified framework. (m) Static terrain constraints and ground observation data), forcibly unifying the resolution can easily lead to the loss of detailed features or the generation of computational artifacts; 3) Significant differences in time frequency, the time frequency span of precipitation products is large, covering various types such as daily scale, 8-day composite, and monthly average, traditional deep learning models are difficult to learn synchronously and use these target data with different observation frequencies for collaborative constraints under a unified backpropagation framework; 4) Weak physical driving correlation, existing fusion algorithms mostly focus on numerical approximation at the pure mathematical level, lacking in-depth modeling of the complex physical correlation between atmospheric dynamic background (such as temperature, humidity, air pressure) and local terrain features (such as slope, height).

[0004] In summary, developing a precipitation fusion reconstruction method that can break through the "input alignment" limitation, achieve multi-temporal and spatial scale collaborative reverse learning, and possess accurate spatial reconstruction capabilities has become a key issue that urgently needs to be addressed in the current field of hydrological and meteorological research. Summary of the Invention

[0005] This invention proposes a multi-source precipitation reconstruction method and storage medium based on multi-temporal and spatial scale collaborative reverse learning, which is used to solve the defects of existing technologies such as strict requirements for spatiotemporal alignment of multi-source precipitation data fusion, difficulty in compatibility with heterogeneous spatial resolutions and inconsistent time frequencies.

[0006] A multi-source precipitation reconstruction method based on multi-temporal-scale collaborative back learning includes the following steps:

[0007] S1: For the selected study area, obtain the daily dynamic meteorological characteristics and inherent static surface characteristics of the area within the historical interval, and construct a single sample in combination with the preset learning objectives; then, along the time axis, according to the set time span, integrate multiple consecutive single samples into sample units to obtain all sample units within the historical interval.

[0008] S2: Construct a network model for encoding and decoding architecture based on U-Net;

[0009] S3: Input each sample unit in step S1 into the model constructed in step S2. For a certain sample unit, when the model's decoder solves the reconstruction value corresponding to the resolutions of 9km, 3km and 1km, first aggregate the reconstruction value in the time dimension, and then calculate the loss function under the specific spatiotemporal resolution.

[0010] S4: Construct the total loss function of the model based on the loss function at a specific spatiotemporal resolution;

[0011] S5: Based on the total loss function constructed in step S4, the Adam optimizer is used to update and optimize the model parameters until the total loss function converges, and the trained model is obtained.

[0012] S6: By inputting the dynamic meteorological characteristics and static surface characteristics of the historical years that need to be reconstructed into the trained model, the model reconstructs the precipitation data of the historical years.

[0013] Furthermore, step S3 includes the following steps:

[0014] S3.1: For the m-th sample cell, m∈[1,N], and is an integer; given a calculated resolution of 9km, the reconstructed value of the cell at time step t is: The reconstructed value corresponding to 3km is: The reconstructed value corresponding to 1km is: ;

[0015] S3.2: Time-scale aggregation; including the following steps:

[0016] S3.2.1: Daily-scale aggregation; the formula is:

[0017]

[0018] In the formula, s∈{s1, s2, s3} represents the spatial scale, s1, s2 and s3 represent spatial scales of 9km, 3km and 1km respectively; i∈{i1, i2, i3}, i1, i2 and i3 represent the pixel indexes at spatial scales of 9km, 3km and 1km respectively. Represents the spatial scale s, the daily-scale aggregated value after aggregation at pixel index i and time step k1; The value before aggregation is represented by t; t is the time step index of the sample unit before aggregation, t∈{1,2,…,32}; k1 represents the daily time step index.

[0019] S3.2.2: 8-day scale aggregation; the formula is:

[0020]

[0021] In the formula, s represents the spatial scale, the pixel index, and the 8-day aggregated value at time step k2; k2 represents the 8-day time period grouping index after aggregation.

[0022] S3.2.3: Lunar-scale aggregation:

[0023]

[0024] In the formula, This represents the monthly scale aggregated value after aggregation at spatial scale s, pixel index i, and time step k3; k3 represents the monthly scale time step index after aggregation.

[0025] S3.3: Calculate the loss function for different precipitation data products; specifically including the following steps:

[0026] S3.3.1: Loss function for daily precipitation data products; the formula is:

[0027]

[0028] In the formula, M1 represents the loss function at spatial scale s for the m-th sample unit of the daily precipitation data product; M1 represents the daily time dimension, M1=32. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the daily-scale aggregated value after aggregation at pixel index i and time step k1; Represents the spatial scale s, the pixel index i, and the daily-scale learning target value at time step k1;

[0029] S3.3.2: Loss function for 8-day precipitation data products; the formula is:

[0030]

[0031] In the formula, M2 represents the loss function at spatial scale s for the m-th sample unit of the 8-day precipitation data product; M2 represents the 8-day time dimension, M2=4. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the 8-day aggregated value after aggregation at pixel index i and time step k2; Represents the spatial scale s, the pixel index i, and the 8-day learning target value at time step k2;

[0032] S3.3.3: Loss function for monthly precipitation data products; the formula is:

[0033]

[0034] In the formula, M3 represents the loss function at the spatial scale s in the m-th sample cell of the monthly precipitation data product; M3 represents the monthly time dimension, M3=1. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the monthly scale aggregated value after aggregation at pixel index i and time step k3; Represents the spatial scale s, the pixel index i, and the monthly scale learning target value at time step k3.

[0035] Furthermore, step S2 includes the following steps:

[0036] S2.1: Feature map E of the encoder's l-th stage output l The calculation formula is:

[0037]

[0038] In the formula, E l The output of the encoder at level l represents the feature map; MaxPool2d represents the max pooling operator; Conv2d represents a two-dimensional convolutional layer; ReLU represents the activation function; E l-1 This represents the feature map of the encoder's (l-1)th stage output.

[0039] S2.2: The bottleneck layer embeds a Transformer module to capture long-range spatial correlations at the whole watershed scale;

[0040] First, the feature map of the deepest layer of the encoder is flattened and superimposed with the positional encoding PE to obtain the initial feature sequence tensor; the formula is:

[0041]

[0042] In the formula, E bottom Represents the feature map of the deepest layer of the encoder; PE is the positional encoding; Z (0) Here, represents the initial feature sequence tensor; Flatten is the flattening operator.

[0043] Self-attention is then calculated using a two-layer Transformer encoder, with the core mechanism being:

[0044]

[0045] In the formula, Q, K, and V represent the query, key, and value matrices, respectively; W Q W K W V Represents the learnable weight matrix; This represents the scaling factor; softmax is the normalization function.

[0046] S2.3: The decoder recovers spatial resolution step by step using transposed convolutions and fuses multi-scale features from the encoder through skip connections; the formula is:

[0047]

[0048] In the formula, D l [*;*] represents the feature map output of the l-th level of the decoder; [*;*] represents the channel-dimensional concatenation operator; ConvTranspose2d represents the two-dimensional transpose convolution operator; D l+1 E is the feature map output by the (l+1)th stage of the decoder. l This represents the feature map of the encoder's level l output.

[0049] Furthermore, in step S4, the formula for the total loss function is:

[0050]

[0051] In the formula, Indicates the daily scale weight; Indicates the 8-day scale weight; Indicates the monthly scale weight.

[0052] Furthermore, Set to 0.4; Set to 0.05; Set to 0.01.

[0053] Furthermore, step S1 includes the following steps:

[0054] S1.1: Select the ERA5-Land dataset of the study area as dynamic meteorological features, and resample the dynamic meteorological features to a 1km target reconstruction resolution;

[0055] S1.2: Select the digital elevation model data and land use / cover data of the study area as static surface feature data, and resample them to a 1km target reconstruction resolution to maintain the same spatial resolution as the dynamic meteorological feature data;

[0056] S1.3: Define the learning objectives; the learning objectives are multidimensional precipitation data products, specifically including daily precipitation products, 8-day precipitation products, and monthly precipitation products;

[0057] S1.4: Construct a single sample; the single sample is composed of: Sample = {Dynamic meteorological characteristics; Static surface characteristics; Learning objective};

[0058] S1.5: Construct sample units; each sample unit is set to contain 32 consecutive daily samples to characterize the time change process of 4 consecutive 8-day cycles.

[0059] A storage medium storing a computer program that, when executed, performs a multi-source precipitation reconstruction method based on multi-temporal-scale collaborative back-learning.

[0060] The beneficial effects that can be achieved by adopting the above technologies are:

[0061] 1. Breaking through the limitations of spatiotemporal alignment: Changing the traditional fusion model from taking the data to be fused as input to setting it as the learning target, so that the model no longer depends on the complete spatiotemporal coverage of all source data, thus enabling effective use of fragmented or sparse precipitation observation data.

[0062] 2. Achieve multi-spatial resolution collaborative reconstruction: By constructing an encoding and decoding architecture with multi-level spatial resolution output, the model is guided to simultaneously capture the cross-scale features of macro-meteorological background and local topographic precipitation by using the reconstruction heads of different levels of the decoder and precipitation products of corresponding resolutions (such as 9km, 3km, 1km).

[0063] 3. Solving the problem of multi-timescale fusion: A time aggregation training method is proposed. By introducing an aggregation operator at the output end, the high-frequency daily scale reconstruction values ​​are aggregated to the time scale corresponding to the target product (such as 8 days or months), realizing the collaborative reverse fusion and physical constraints of precipitation information of multiple time scales. Attached Figure Description

[0064] Figure 1 This is the flowchart of this solution;

[0065] Figure 2 This is a scatter plot comparing various precipitation products, the model reconstruction method of this scheme, and the traditional TC reconstruction method; Figure 2 (a) in the figure is a scatter plot of TPMFD precipitation products; Figure 2 (b) in the figure is a scatter plot of MSWEP precipitation products; Figure 2 (c) in the figure is a scatter plot of IMERG precipitation products; Figure 2 (d) in the diagram is the scatter plot after the model of this scheme is reconstructed; Figure 2 (e) in the diagram is the scatter plot after reconstruction using the traditional TC method;

[0066] Figure 3 This is a comparison of precipitation time series reconstructed by the model in this scheme and reconstructed by the traditional TC scheme from 1956 to 2020.

[0067] Figure 4 This is a comparison chart of the spatial distribution of precipitation for various precipitation products, the model reconstruction method of this scheme, and the traditional TC reconstruction method; Figure 4 (a) in the figure is a spatial distribution map of precipitation for TPMFD precipitation products; Figure 4 (b) in the figure is a spatial distribution map of precipitation from the MSWEP precipitation product; Figure 4 (c) in the figure is the spatial distribution map of precipitation in the IMERG precipitation product; Figure 4 (d) in the figure is the spatial distribution map of precipitation reconstructed by the model of this method; Figure 4 (e) in the diagram is the spatial distribution of precipitation after reconstruction by the traditional TC. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example 1: A multi-source precipitation reconstruction method based on multi-temporal-scale collaborative back learning, comprising the following steps:

[0070] S1: For the selected study area, obtain the daily dynamic meteorological characteristics and inherent static surface characteristics of the area within the historical interval, and construct a single sample in combination with the preset learning objectives; then, along the time axis according to the set time span, integrate multiple consecutive single samples into sample units to obtain all sample units within the historical interval.

[0071] This embodiment takes the Yarlung Tsangpo River Basin as the study area and the precipitation data of the basin from 1979 to 2020 as the research object. Step S1 is further refined into the following sub-steps:

[0072] S1.1: The ERA5-Land dataset (European Centre for Medium-Range Weather Forecasts, fifth generation land reanalysis dataset) of the study area was selected as the dynamic meteorological features. The dynamic meteorological features include temperature, precipitation, radiation, relative humidity, evapotranspiration and soil moisture. At the same time, the dynamic meteorological features were resampled to a 1km target reconstruction resolution.

[0073] S1.2: Select the digital elevation model (DEM) data and land use / cover (LULC) data of the study area as static surface feature data, and resample them to a target reconstruction resolution of 1km to maintain the same spatial resolution as the dynamic meteorological feature data.

[0074] S1.3: Define learning objectives.

[0075] This embodiment uses multidimensional precipitation data products as the learning target, covering multiple spatial and temporal scales. The spatial resolution includes 9 km, 3 km and 1 km, and the temporal resolution includes daily, 8-day and monthly.

[0076] S1.4: Construct a single sample.

[0077] The composition of a single sample is as follows:

[0078] Sample = {Dynamic meteorological characteristics; Static surface characteristics; Learning objectives};

[0079] Specific examples are as follows:

[0080] If the current moment happens to be the start day of the month and is at an 8-day synthesis cycle node, then the learning target of this sample will simultaneously load {daily scale target, 8-day scale target, and monthly scale target}.

[0081] Taking January 1, 1979 as an example, the sample constructed on that day is: {dynamic meteorological characteristics of January 1, 1979; static surface characteristics; daily precipitation data products of January 1, 1979; 8-day precipitation data products of January 1, 1979; monthly precipitation data products of January 1, 1979};

[0082] If only daily-scale observations are available at the current moment, then the learning target will only load daily-scale precipitation products.

[0083] Taking January 2, 1979 as an example, the sample constructed for that day is: {dynamic meteorological characteristics of January 2, 1979; static surface characteristics; daily precipitation data product of January 2, 1979};

[0084] Following the above rules, all single samples within the historical intervals of the study area are constructed sequentially along the timeline.

[0085] S1.5: Construct sample units.

[0086] A sample unit is defined, containing 32 consecutive daily samples, which can correspond to 4 complete 8-day cycles, used to characterize the variation characteristics of longer time series. After the model outputs continuous daily-scale time series, it can be mapped to 8-day and monthly-scale statistical features through time aggregation to achieve consistent characterization across multiple time scales (see step S3).

[0087] Slide along the time axis with a fixed step size to combine 32 consecutive single samples into a sample unit, and finally construct a total of N sample units.

[0088] S2: Construct a network model for encoding and decoding architecture based on U-Net. This includes the following steps:

[0089] S2.1: The encoder extracts deep spatial features from a 1 km resolution input set through continuous two-dimensional convolution and downsampling operations.

[0090] Feature map E of the encoder's level l output l The calculation formula is:

[0091]

[0092] In the formula, E l The output of the encoder at level l is represented by the feature map; MaxPool2d represents the max pooling operator, used to achieve spatial downsampling; Conv2d is a two-dimensional convolutional layer, used to increase the depth of feature channels and capture local meteorological and physical correlations; ReLU is the activation function, used to introduce nonlinear mapping capabilities; E l-1 This represents the feature map of the encoder's (l-1)th stage output.

[0093] S2.2: The bottleneck layer embeds a Transformer module to capture long-distance spatial correlations at the whole watershed scale.

[0094] First, the feature map of the deepest layer of the encoder is flattened and superimposed with the positional encoding PE to obtain the initial feature sequence tensor; the formula is:

[0095]

[0096] In the formula, E bottom Represents the feature map of the deepest layer of the encoder; PE is the positional encoding; Z (0) is the initial feature sequence tensor; Flatten is the flattening operator.

[0097] Self-attention is then calculated using a two-layer Transformer encoder, with the core mechanism being:

[0098]

[0099] In the formula, Q, K, and V represent the query, key, and value matrices, respectively; W Q W K W V Represents the learnable weight matrix; represents the scaling factor; softmax is the normalization function; Attention is the attention function.

[0100] S2.3: The decoder uses transposed convolution (ConvTranspose2d) to recover spatial resolution step by step and fuses multi-scale features from the encoder through skip connections.

[0101] The formula is:

[0102]

[0103] In the formula, D l [*;*] represents the feature map output of the l-th level of the decoder; [*;*] represents the channel-dimensional concatenation operator; ConvTranspose2d represents the two-dimensional transpose convolution operator; D l+1 E is the feature map output by the (l+1)th stage of the decoder. l This represents the feature map of the encoder's level l output.

[0104] S3: Input each sample unit from step S1 into the model constructed in step S2. For a given sample unit, when the model's decoder reconstruction head calculates the reconstructed values ​​at resolutions of 9km, 3km, and 1km, first aggregate the reconstructed values ​​along the time dimension, and then calculate the loss function at each specific spatiotemporal resolution. "Specific spatiotemporal resolution" refers to the combined scale of spatial and temporal resolutions corresponding to the learning target precipitation data product. Different learning target precipitation data products have their own inherent spatial resolution attributes, such as 9, 3, and 1km, and temporal resolution attributes, such as daily, 8-day, and monthly.

[0105] Specifically, the following steps are included:

[0106] S3.1: For the m-th sample cell, m∈[1,N], and is an integer; given a calculated resolution of 9km, the reconstructed value of the cell at time step t is: Similarly, the reconstructed value corresponding to 3km is: The reconstructed value corresponding to 1km is: ;

[0107] S3.2: Time-scale aggregation. This includes the following sub-steps:

[0108] S3.2.1: Daily-scale aggregation;

[0109] The formula is:

[0110]

[0111] In the formula, s∈{s1, s2, s3} represents the spatial scale, which is the spatial resolution; s1, s2, and s3 represent spatial scales of 9km, 3km, and 1km, respectively; i∈{i1, i2, i3} represents the pixel index at the corresponding spatial scale, i1, i2, and i3 represent the pixel index at spatial scales of 9km, 3km, and 1km, respectively. Represents the spatial scale s, the daily-scale aggregated value after aggregation at pixel index i and time step k1; represents the reconstructed value before aggregation; t is the time step index of the sample unit before aggregation, t∈{1,2,…,32}; k1 represents the daily-scale time step index after aggregation.

[0112] S3.2.2: 8-day scale aggregation;

[0113]

[0114] In the formula, This represents the spatial scale s, the pixel index i, and the 8-day scale aggregated value at time step k2; k2 represents the 8-day time period grouped index after aggregation.

[0115] S3.2.3: Lunar-scale aggregation:

[0116]

[0117] In the formula, This represents the monthly scale aggregated value after aggregation at spatial scale s, pixel index i, and time step k3; k3 represents the monthly scale time step index after aggregation.

[0118] S3.3: Calculate the loss function for different precipitation data products.

[0119] S3.3.1: Loss function for daily precipitation data products:

[0120]

[0121] In the formula, M1 represents the loss function at spatial scale s for the m-th sample unit of the daily precipitation data product; M1 represents the daily time dimension, M1=32. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the daily-scale aggregated value after aggregation at pixel index i and time step k1; Represents the spatial scale s, the pixel index i, and the daily-scale learning target value at time step k1;

[0122] S3.3.2: Loss function for 8-day precipitation data products:

[0123]

[0124] In the formula, represents the loss function at spatial scale s for the m-th sample unit of the 8-day precipitation data product; M2 represents the 8-day time dimension size, M2=4; This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the 8-day aggregated value after aggregation at pixel index i and time step k2; Represents the spatial scale s, the pixel index i, and the 8-day learning target value at time step k2;

[0125] S3.3.3: Loss function for monthly precipitation data products:

[0126]

[0127] In the formula, M3 represents the loss function at the spatial scale s in the m-th sample cell of the monthly precipitation data product; M3 represents the monthly time dimension, M3=1. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the monthly scale aggregated value after aggregation at pixel index i and time step k3; Represents the spatial scale s, the pixel index i, and the monthly scale learning target value at time step k3.

[0128] S4: The total loss function of the model is used for backpropagation.

[0129] The formula for the total loss function is:

[0130]

[0131] In the formula, This indicates the daily scale weight, set to 0.4; This indicates an 8-day scale weight, set to 0.05; This indicates the monthly scale weight, set to 0.01.

[0132] S5: Based on the total loss function constructed above, the Adam optimizer is used to update and optimize the model parameters, with an initial learning rate of 10. -4 The entire model training cycle is set to 50 epochs; the model training is completed when the total loss function converges, and the trained model is obtained.

[0133] S6: By inputting the dynamic meteorological characteristics and static surface characteristics of the historical years that need to be reconstructed into the trained model, the model reconstructs the precipitation data of the historical years.

[0134] Example 2: A storage medium storing a computer program that executes a multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning according to Example 1.

[0135] Comparison of reconstruction results from different methods:

[0136] Following the method in Example 1, the experiment was divided into two phases: model training and accuracy validation. The training phase used data from 1979 to 2010, with learning objectives including daily observations from MSWEP, IMERG, TPMFD, and basic meteorological stations, at spatial resolutions of 9 km, 3 km, and 1 km, respectively. The validation phase used data from 2011 to 2020 to evaluate model performance. To further demonstrate model performance, a comparative analysis was simultaneously conducted using the traditional Triple Collocation (TC) method.

[0137] Figure 2 This paper presents a comparison of the station accuracy of the model reconstruction results from this proposed scheme with that of the target precipitation products (TPMFD, MSWEP, IMERG) and the traditional TC reconstruction method. Experimental results show that the correlation coefficient (R) of the model reconstruction results from this scheme reaches 0.731, significantly higher than that of the target precipitation products and the TC reconstruction method (both R values ​​are below 0.54). Simultaneously, its root mean square error (RMSE) is reduced to 2.566 mm, the lowest among all schemes. These data demonstrate that this scheme, through multi-source data fusion, reconstructs a precipitation product with superior performance.

[0138] like Figure 3As shown, taking the Shigatse station as an example, the time series evolution of reconstructed precipitation is plotted. The results show that this scheme uses multi-source precipitation products as a unified learning objective for modeling, without relying on overlapping intervals of various data sources in the time dimension. This enables cross-time period information learning and generalization, supporting continuous, long-term precipitation series reconstruction output (continuous precipitation data is available within the 1956-2020 period). In contrast, the traditional TC method uses multi-source products as input variables, limited by the common time coverage of different data sources. It requires taking the intersection of their time series before modeling, resulting in a significantly limited time length of the final reconstruction result, making it difficult to meet the needs of long-term time series analysis and applications (precipitation data for the 1956-2000 period is missing).

[0139] like Figure 4 As shown, taking the multi-year average daily precipitation from 2016 to 2020 as an example, the spatial distribution characteristics of different data sources and reconstruction results are compared. The results show that this method, by constructing a multi-level resolution reconstruction head structure, can simultaneously learn information representations at different scales within a unified framework. This allows the model to depict both large-scale meteorological backgrounds and capture fine spatial differences driven by local topography, achieving collaborative modeling of cross-scale features. Furthermore, by introducing high-resolution topographic features as auxiliary information, the model can further enhance its response to local control factors such as topographic relief, slope aspect, and elevation changes, resulting in a clearer detailed structure and a more reasonable spatial gradient in the spatial distribution of precipitation. In contrast, traditional TC methods typically require unified spatial resolution processing of multi-source data. During resampling, this inevitably smooths or weakens the original high-frequency spatial information, leading to the loss of local detailed features and making it difficult to accurately characterize the spatial heterogeneity of precipitation under complex topographic conditions.

[0140] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning, characterized in that, Includes the following steps: S1: For the selected study area, obtain the daily dynamic meteorological characteristics and inherent static surface characteristics of the area within the historical interval, and construct a single sample in combination with the preset learning objectives; then, along the time axis, according to the set time span, integrate multiple consecutive single samples into sample units to obtain all sample units within the historical interval. S2: Construct a network model for encoding and decoding architecture based on U-Net; S3: Input each sample unit in step S1 into the model constructed in step S2. For a certain sample unit, when the model's decoder solves the reconstruction value corresponding to the resolutions of 9km, 3km and 1km, first aggregate the reconstruction value in the time dimension, and then calculate the loss function under the specific spatiotemporal resolution. S4: Construct the total loss function of the model based on the loss function at a specific spatiotemporal resolution; S5: Based on the total loss function constructed in step S4, the Adam optimizer is used to update and optimize the model parameters until the total loss function converges, and the trained model is obtained. S6: By inputting the dynamic meteorological characteristics and static surface characteristics of the historical years that need to be reconstructed into the trained model, the model reconstructs the precipitation data of the historical years.

2. The multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning according to claim 1, characterized in that, Step S3 includes the following steps: S3.1: For the m-th sample cell, m∈[1,N], and is an integer; given a calculated resolution of 9km, the reconstructed value of the cell at time step t is: The reconstructed value corresponding to 3km is: The reconstructed value corresponding to 1km is: ; S3.2: Time-scale aggregation; including the following steps: S3.2.1: Daily-scale aggregation; the formula is: ； In the formula, s∈{s1, s2, s3} represents the spatial scale, s1, s2 and s3 represent spatial scales of 9km, 3km and 1km respectively; i∈{i1, i2, i3}, i1, i2 and i3 represent the pixel indexes at spatial scales of 9km, 3km and 1km respectively. Represents the spatial scale s, the daily-scale aggregated value after aggregation at pixel index i and time step k1; The value before aggregation is represented by t; t is the time step index of the sample unit before aggregation, t∈{1,2,…,32}; k1 represents the daily time step index. S3.2.2: 8-day scale aggregation; the formula is: ； In the formula, s represents the spatial scale, the pixel index, and the 8-day aggregated value at time step k2; k2 represents the 8-day time period grouping index after aggregation. S3.2.3: Lunar-scale aggregation: ； In the formula, This represents the monthly scale aggregated value after aggregation at spatial scale s, pixel index i, and time step k3; k3 represents the monthly scale time step index after aggregation. S3.3: Calculate the loss function for different precipitation data products; specifically including the following steps: S3.3.1: Loss function for daily precipitation data products; the formula is: ； In the formula, M1 represents the loss function at spatial scale s for the m-th sample unit of the daily precipitation data product; M1 represents the daily time dimension, M1=32. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the daily-scale aggregated value after aggregation at pixel index i and time step k1; Represents the spatial scale s, the pixel index i, and the daily-scale learning target value at time step k1; S3.3.2: Loss function for 8-day precipitation data products; the formula is: ； In the formula, M2 represents the loss function at spatial scale s for the m-th sample unit of the 8-day precipitation data product; M2 represents the 8-day time dimension, M2=4. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the 8-day aggregated value after aggregation at pixel index i and time step k2; Represents the spatial scale s, the pixel index i, and the 8-day learning target value at time step k2; S3.3.3: Loss function for monthly precipitation data products; the formula is: ； In the formula, M3 represents the loss function at the spatial scale s in the m-th sample cell of the monthly precipitation data product; M3 represents the monthly time dimension, M3=1. This represents the total number of valid pixels at a spatial scale s. Represents the spatial scale s, the monthly scale aggregated value after aggregation at pixel index i and time step k3; Represents the spatial scale s, the pixel index i, and the monthly scale learning target value at time step k3.

3. The multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning according to claim 1, characterized in that, Step S2 includes the following steps: S2.1: Feature map E of the encoder's l-th stage output l The calculation formula is: ； In the formula, E l The output of the encoder at level l represents the feature map; MaxPool2d represents the max pooling operator; Conv2d represents a two-dimensional convolutional layer; ReLU represents the activation function; E l-1 This represents the feature map output of the encoder at level l-1; S2.2: The bottleneck layer embeds a Transformer module to capture long-range spatial correlations at the whole watershed scale; First, the feature map of the deepest layer of the encoder is flattened and superimposed with the positional encoding PE to obtain the initial feature sequence tensor; the formula is: ； In the formula, E bottom Represents the feature map of the deepest layer of the encoder; PE is the positional encoding; Z (0) Here, represents the initial feature sequence tensor; Flatten is the flattening operator. Self-attention is then calculated using a two-layer Transformer encoder, with the core mechanism being: ； In the formula, Q, K, and V represent the query, key, and value matrices, respectively; W Q W K W V Represents the learnable weight matrix; This represents the scaling factor; softmax is the normalization function. S2.3: The decoder recovers spatial resolution step by step using transposed convolutions and fuses multi-scale features from the encoder through skip connections; the formula is: ； In the formula, D l [*;*] represents the feature map output of the l-th level of the decoder; [*;*] represents the channel-dimensional concatenation operator; ConvTranspose2d represents the two-dimensional transpose convolution operator; D l+1 E is the feature map output by the (l+1)th stage of the decoder. l This represents the feature map of the encoder's level l output.

4. The multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning according to claim 2, characterized in that, In step S4, the formula for the total loss function is: ； In the formula, Indicates the daily scale weight; Indicates the 8-day scale weight; Indicates the monthly scale weight.

5. The multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning according to claim 4, characterized in that, Set to 0.4; Set to 0.05; Set to 0.

01.

6. The multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning according to claim 2, characterized in that, Step S1 includes the following steps: S1.1: Select the ERA5-Land dataset of the study area as dynamic meteorological features, and resample the dynamic meteorological features to a 1km target reconstruction resolution; S1.2: Select the digital elevation model data and land use / cover data of the study area as static surface feature data, and resample them to a 1km target reconstruction resolution to maintain the same spatial resolution as the dynamic meteorological feature data; S1.3: Define the learning objectives; the learning objectives are multidimensional precipitation data products, specifically including daily precipitation products, 8-day precipitation products, and monthly precipitation products; S1.4: Construct a single sample; the single sample is composed of: Sample = {Dynamic meteorological characteristics; Static surface characteristics; Learning objectives}; S1.5: Construct sample units; each sample unit is set to contain 32 consecutive daily samples to characterize the time change process of 4 consecutive 8-day cycles.

7. A storage medium, characterized in that, The storage medium contains a computer program that, when executed, performs a multi-source precipitation reconstruction method based on multi-temporal-scale collaborative reverse learning as described in any one of claims 1-6.

Images