GConvAttnLSTM multi-source remote sensing rainfall estimation method and system fusing spatial-temporal characteristics

Abstract

The invention discloses a spatial-temporal feature fused GConvAttnLSTM multi-source remote sensing rainfall estimation method and system, and the method comprises the steps: firstly carrying out the spatial resampling and quality control preprocessing of multi-source remote sensing rainfall data, and extracting rainfall window data in a fixed space neighborhood with a site as the center; secondly, constructing sliding time sequence samples of multiple consecutive days, and forming a four-dimensional data structure with multi-product space-time characteristics; then, designing and training a deep neural network model fusing image convolution, an attention mechanism and a long-short-term memory network, respectively extracting spatial correlation features and a time evolution mode of the multi-source rainfall data, and realizing collaborative modeling and fusion output; and finally, performing error evaluation on the fusion result through actual measurement data of the ground meteorological station. According to the method, the precision and robustness of multi-source remote sensing rainfall estimation can be effectively improved, and the method is suitable for large-range and long-time-sequence remote sensing rainfall estimation tasks.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of multi-source data fusion, in particular to a GConvAttnLSTM multi-source remote sensing precipitation estimation method and system fusing space-time features. BACKGROUND

[0002] Ground meteorological observation stations can directly measure precipitation accumulation, and the observation results are considered to be reliable basis for quantitative estimation of precipitation; however, the ground meteorological observation stations are sparsely distributed in space and time, and the resolution is low, and the uncertainty of the interpolation processing method limits the precision of the precipitation data. Remote sensing precipitation products based on infrared, microwave and other sensors have the advantage of wide-range observation, and can better capture the space-time distribution information of precipitation; however, these satellite precipitation products usually have systematic bias and uncertainty, and it is difficult to meet the demand of high-precision precipitation estimation alone. Multi-source remote sensing precipitation products have complementarity, so by fusing multi-source precipitation data, the advantages of each product can be integrated, and the shortcomings of a single data source can be made up, so as to improve the precision and reliability of precipitation estimation. This multi-source precipitation fusion method is expected to obtain a precipitation field with high space-time resolution and close to the true precipitation in value, and to provide more accurate data support for hydro-meteorological analysis and disaster warning. Existing researches show that the fusion method based on deep learning significantly improves the precipitation estimation effect, for example, the space-time fusion method combining convolutional neural network (CNN) and long short-term memory network (LSTM) can more fully excavate the space-time features of precipitation data, and improve the precipitation estimation precision. SUMMARY

[0003] The application provides a GConvAttnLSTM multi-source remote sensing precipitation estimation method and system fusing space-time features.

[0004] In order to achieve the above application purposes, the technical scheme adopted by the application is as follows:

[0005] A GConvAttnLSTM multi-source remote sensing precipitation estimation method fusing space-time features, the method comprises the following steps:

[0006] Step S1: multi-source remote sensing precipitation data preprocessing, uniformly resampling, projecting and converting and quality controlling a plurality of remote sensing precipitation products;

[0007] Step S2: extracting an n*n neighborhood precipitation window centered on a station or a pixel, and constructing a space window sample; constructing a space-time precipitation sequence sample of T consecutive days according to day sliding, and forming a three-dimensional tensor structure;

[0008] Step S3: input the spatiotemporal precipitation sequence sample into the GConvAttnLSTM model, extract spatial features through graph convolution, extract temporal features through LSTM, and predict precipitation by fusing attention mechanism;

[0009] Step S4: output the fused precipitation result, and evaluate the accuracy in combination with the ground meteorological observation station data.

[0010] Further, the step S1 comprises:

[0011] S101: acquire multi-source remote sensing precipitation data with consistent time coverage, and eliminate abnormal values and missing dates;

[0012] S102: unify the coordinate systems and spatial resolutions of each data source;

[0013] S103: select dates common to all data sources to generate sample construction sets.

[0014] Further, the step S2 comprises:

[0015] S201: acquire the coordinates of the corresponding grid pixels of the ground meteorological station;

[0016] S202: extract n×n pixels to form a spatial window centered on the station pixel;

[0017] S203: perform window extraction on each data source to form D×n×n-dimensional precipitation segments, where D is the number of remote sensing data sources;

[0018] S204: repeat the above operation at each time step and each station pixel, and store it as the original spatial sample data.

[0019] Further, the step S2 further comprises time sliding sample construction, specifically as follows:

[0020] S205: time sort the spatial window samples according to date for each station pixel;

[0021] S206: construct time sequence samples with a time length of T days in a sliding window manner;

[0022] S207: each sample is a T×D×n×n four-dimensional tensor, representing the precipitation information of all data sources within the spatial window for consecutive T days;

[0023] S208: the label corresponding to each input sequence is the fused precipitation value of the target position on the Tth day.

[0024] Further, the GConvAttnLSTM model operation in the step S3 comprises:

[0025] S401: input X t∈R T×D×n×n Performing graph convolution:

[0026]

[0027] wherein, is the spatial representation after graph convolution, X t is the current spatio-temporal remote sensing precipitation input, is the graph structure adjacency matrix based on geographical adjacency, and GCN is the graph convolutional neural network operation.

[0028] S402: Enhance features through spatial attention mechanism:

[0029]

[0030] wherein, A t is the attention weight matrix, π is the spatial position, d is the feature dimension scaling factor, and δ is the temperature coefficient, represents pixel-by-pixel multiplication, is the spatial enhanced input after graph convolution and attention weighting;

[0031] S403: Input the sequence into the LSTM, update the memory state according to formulas (3)-(6), as follows:

[0032] ① Forget gate:

[0033]

[0034] ② Input gate and candidate memory:

[0035]

[0036] ③ Memory state update:

[0037]

[0038] ④ Output gate and current hidden state:

[0039]

[0040] wherein, h t-1 is the previous time hidden state, C t-1 is the previous time memory cell state, f t is the forget gate, i t is the input gate, is the candidate memory state, C t is the current memory state, o t is the output gate, h tfor the current hidden state (final output), is the sigmoid activation function, tanh is the hyperbolic tangent activation function, W, b are the network trainable weights and bias parameters.

[0041] S404: Apply time attention weights a to the LSTM output sequence t :

[0042]

[0043] where W α is the weight matrix of the time attention mechanism;

[0044] After weighting, the global time-aware features are aggregated

[0045] S405: The fully connected layer outputs the precipitation estimation value:

[0046]

[0047] where, is the global feature representation after time attention aggregation. FC represents the fully connected layer operation;

[0048] S406: Train the model using the mean square error loss function:

[0049]

[0050] where is the precipitation fusion estimation result generated by the model at the current time, y i is the true reference value of the precipitation, which is derived from the spatial interpolation result of the precipitation records of the ground meteorological observation station, and N is the total number of samples in the training / validation set.

[0051] Further, the step S4 comprises:

[0052] S501: Input the test sample into the model to output the precipitation estimation value

[0053] S502: Use the ground meteorological observation station data as the true value y i ;

[0054] S503: Calculate the RMSE, unbiased root mean square error (ubRMSE), MAE, and Pearson correlation coefficient (PCC) precision indicators.

[0055] S504: Compare the original remote sensing product to evaluate the fusion effect.

[0056] Further, the model training in step S3 uses a sliding window mechanism, and the input is X t ∈R T×D×n×nFour-dimensional tensor, where T=7 days, D=3 types of remote sensing data sources, n=9 (9x9 spatial window).

[0057] The application further discloses a GConvAttnLSTM multi-source remote sensing precipitation estimation system fusing space-time features.

[0058] The data preprocessing module is used for space-time standardization of multi-source remote sensing precipitation data, including removing outliers and missing dates, unifying coordinate systems and spatial resolutions, and screening a common date set of multi-source data.

[0059] The space-time sample construction module is used for structured feature extraction of precipitation data, including generating an n*n neighborhood spatial window centered on a site or a pixel, constructing a sliding time series sample of consecutive T days, associating a fusion precipitation value of a target position on the Tth day, and outputting a four-dimensional tensor with a dimension of TxDxn*n.

[0060] The deep learning fusion module is used for collaborative modeling of multi-modal precipitation features, including extracting spatial neighborhood dependency features through graph convolution, enhancing key area response through a spatial attention mechanism, modeling time evolution rules through an LSTM network, aggregating key time period features through a time attention mechanism, and outputting a precipitation fusion estimation value through a fully connected layer.

[0061] The precision verification and output module is used for quantitative evaluation and business output of fusion results, including accessing ground meteorological observation station true value data, calculating RMSE / ubRMSE / MAE / PCC precision indicators, generating an improved quantitative report of original products and fusion results, and supporting geographic spatial rendering visualization.

[0062] The application further discloses a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the GConvAttnLSTM multi-source remote sensing precipitation estimation method when executing the program.

[0063] The application further discloses a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the GConvAttnLSTM multi-source remote sensing precipitation estimation method.

[0064] Compared with the prior art, the application has the following advantages:

[0065] It has obvious advantages in the field of multi-source remote sensing precipitation data fusion, adopts the GConvAttnLSTM model structure integrating graph convolution (GConv), long short-term memory network (LSTM) and spatial-temporal attention mechanism, realizes efficient modeling of spatial neighborhood dependence relationship and time sequence evolution characteristics of precipitation data, and has good regional expansion; the method is based on the design of TxDxnxn sliding samples, constructs the spatial-temporal collaborative representation of multi-source remote sensing precipitation sequence, does not rely on station data for modeling, only introduces the measured precipitation value as a reference label in supervised training, and has good regional expansion; through the attention mechanism of spatial and temporal dimensions, the response ability of the model to key areas and key periods is enhanced, and the fusion result is significantly better than the original precipitation product in terms of accuracy, stability and spatial continuity, and has good generalization, adaptability and engineering deployability. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 is a flowchart of a GConvAttnLSTM multi-source remote sensing precipitation estimation method fusing spatial-temporal characteristics according to an embodiment of the present application;

[0067] Figure 2 is a GConvAttnLSTM model structure diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will be further described in detail according to the drawings and examples.

[0069] As shown in Figure 1 , the present application provides a GConvAttnLSTM multi-source remote sensing precipitation estimation method fusing spatial-temporal characteristics, which comprises:

[0070] Step S1: multi-source remote sensing precipitation data preprocessing, uniformly resampling, projecting and converting and quality control on inputted various remote sensing precipitation products, to ensure data consistency;

[0071] Step S2: spatial-temporal and sliding sample construction. First, construct a spatial window sample, take a station or a pixel as the center, extract an n*n neighborhood precipitation window from each product, generate a multi-source grid sample; second, construct a sliding time sequence sample, construct a continuous T-day spatial-temporal precipitation sequence sample according to day sliding, form a three-dimensional tensor structure;

[0072] Step S3: sample data division. Divide the sample data into a training set, a validation set and a test set;

[0073] Step S4: GConvAttnLSTM model modeling and training, inputting the spatio-temporal sample into the model, extracting spatial features by graph convolution, extracting time features by LSTM, and fusing attention mechanism for precipitation prediction;

[0074] Step S5: fusion precipitation result output and precision evaluation, outputting the fusion result, and using independent ground meteorological observation station data for precision verification and performance evaluation.

[0075] The step S1 comprises:

[0076] This step is used to ensure the consistency of various remote sensing precipitation data products in spatial resolution, coordinate system and time range, specifically including:

[0077] Step S101: obtaining multi-source remote sensing precipitation data, requiring consistent time coverage, and removing outliers from the original data, discarding missing or incomplete dates;

[0078] In this embodiment, CHIRPS, PERSIANN-CDR and MSWEP three types of daily scale precipitation products are downloaded, and the time range covers from 1983 to 2021; secondly, each data source is arranged in a unified format (GeoTIFF or NetCDF);

[0079] Step S102: projecting each data source to a unified coordinate system and resampling to a unified spatial resolution;

[0080] Firstly, the three types of data are projected to the 2000 national geodetic coordinate system using the ArcMap batch processing tool;

[0081] Secondly, it is resampled to a grid with a spatial resolution of 0.1°;

[0082] Finally, remove invalid values (such as -9999) and perform normalization to generate standard input data.

[0083] Step S103: selecting dates that exist in all data sources at the same time to generate a qualified date set for sample construction.

[0084] The step S2 comprises:

[0085] This step first extracts the spatial neighborhood sample composed of precipitation pixels within a fixed range with the precipitation estimation target position as the center, specifically including:

[0086] S201: obtaining the coordinates of the spatial position of the ground meteorological observation station (referred to as "station pixel") corresponding to the grid pixel in the study area;

[0087] First, determine the 9x9 neighborhood spatial window corresponding to each station pixel;

[0088] Secondly, extract the precipitation value of the three types of remote sensing products in the space every day;

[0089] Finally, form a daily 3x9x9 space tensor.

[0090] S202: Extract n*n pixels centered on the site pixel to form a spatial window;

[0091] First, set the sliding window length T=7 days;

[0092] Secondly, construct a continuous sample sequence by sliding by day;

[0093] Finally, organize into a tensor X∈R 7×3×9×9 .

[0094] S203: At each time step, perform window extraction on each data source to form a Dxnxn-dimensional precipitation segment, where D is the number of remote sensing data sources;

[0095] First, cache the tail data after each year of data processing is completed;

[0096] Secondly, use the cache to construct the next year's sample;

[0097] Finally, ensure that the sliding window time sequence continuity is not interrupted.

[0098] S204: Repeat the above operation at each time step and each site pixel to store as raw spatial sample data.

[0099] First, unify the sample structure and encode the meta information;

[0100] Secondly, pack into Python's NumPy array format;

[0101] Finally, organize and store by year.

[0102] To capture the temporal variation of precipitation, construct a time sliding window sample to form a spatio-temporal collaborative input data, which includes:

[0103] S205: Time sort the spatial window samples by date for each site pixel;

[0104] S206: Use a sliding window approach to build a time series sample with a time length of T days;

[0105] S207: Each sample is a TxDxnxn four-dimensional tensor, representing the precipitation information of all data sources within the spatial window for consecutive T days;

[0106] S208: The label corresponding to each input sequence is the fused precipitation value of the target location on the Tth day;

[0107] The step S3 comprises:

[0108] All samples are divided into training set, validation set and test set in time sequence, which are respectively used for model training, parameter optimization and precision evaluation.

[0109] Firstly, the data from 1983 to 2007 is taken as the training set;

[0110] Secondly, the data from 2008 to 2012 is set as the validation set;

[0111] Finally, the data from 2013 to 2021 is reserved as the test set.

[0112] The step S4 comprises:

[0113] The core of the application is the GConvAttnLSTM model structure, which combines graph convolution, LSTM and attention mechanism to mine the spatial correlation and time dynamic pattern in the precipitation data, and specifically comprises:

[0114] S401: Firstly, the sliding window mechanism is adopted to stack the D*n*n multi-source remote sensing precipitation spatial window of continuous T days and each day along the time axis to form a tensor input X t ∈R T×D×n×n .

[0115] Then, the input sample X t ∈R T×D×n×n of each day is subjected to graph convolution operation to extract high-order features between spatial neighborhoods.

[0116] As shown in formula (1):

[0117]

[0118] Among them, is the spatial representation after graph convolution, X t is the current remote sensing precipitation input, is the graph structure adjacency matrix based on geographical adjacency, and GCN is the graph convolution neural network operation.

[0119] S402: The graph convolution spatial features are compressed into daily scale vectors through adaptive pooling operation, and the spatial attention mechanism is introduced to calculate the attention weight A t , and the key area is enhanced to compress the spatial features into daily scale vectors.

[0120] As shown in formula (2):

[0121]

[0122] Among them, A tis the attention weight matrix, π is the spatial position, d is the feature dimension scaling factor, δ is the temperature coefficient (controls the smoothness of softmax), represents the per-pixel multiplication, is the spatial enhanced input after graph convolution and attention weighting.

[0123] S403: T spatial attention enhanced sequences of daily scale are input into the LSTM network to capture the temporal evolution process of precipitation.

[0124] The key update mechanism is shown in formulas (3)-(6):

[0125] ① Forget gate:

[0126]

[0127] ② Input gate and candidate memory:

[0128]

[0129] ③ Memory state update:

[0130]

[0131] ④ Output gate and current hidden state:

[0132]

[0133] where h t-1 is the hidden state at the previous time, C t-1 is the memory cell state at the previous time, f t is the forget gate, i t is the input gate, is the candidate memory state, C t is the current memory state, o t is the output gate, h t is the current hidden state (final output), σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, W, b are network trainable weights and bias parameters.

[0134] S404: Based on the LSTM output sequence {h t-T+1 ,...,h t}, introduce time attention weight α t , highlight the features of key precipitation periods.

[0135] as shown in formula (7):

[0136]

[0137] The weighted and aggregated global time-aware features

[0138] S405: Aggregated representation of time attention The input fully connected layer generates the precipitation fusion estimation result of the current time.

[0139] As shown in equation (8):

[0140]

[0141] S406: Model parameter training using the training set and hyperparameter tuning using the validation set. The loss function is usually the mean squared error (MSE).

[0142] As shown in equation (9):

[0143]

[0144] where is the model-generated precipitation fusion estimation result of the current time, y i is the real reference value of precipitation, derived from the spatial interpolation result of ground meteorological observation station precipitation records, and N is the total number of samples in the training / validation set.

[0145] In this embodiment, S4 constructs a hybrid deep network with graph convolution structure, spatial-temporal attention mechanism and LSTM memory capability, realizes multi-source remote sensing precipitation estimation, as shown in equation (10). Figure 2

[0146] Step 4.1: Spatial graph convolution modeling.

[0147] First, construct a 3x9x9 input for each day into a graph structure; second, apply graph convolution to extract spatial neighborhood correlation features;

[0148] Finally, output the spatial embedding vector for subsequent processing.

[0149] Step 4.2: Spatial attention mechanism.

[0150] First, calculate the importance weight of each pixel;

[0151] Second, enhance the response of key areas through attention allocation;

[0152] Finally, optimize the spatial feature expression capability.

[0153] Step 4.3: Time series modeling.

[0154] First, concatenate the T=7-day spatial embedding vectors;

[0155] ​Secondly, it is fed into the LSTM network to model time evolution;

[0156] Finally, long-term dependency information is captured.

[0157] Step 4.4: Temporal Attention Mechanism.

[0158] First, calculate the attention score output at each moment;

[0159] Second, highlight the characteristic contributions of key periods;

[0160] Finally, strengthen the model's ability to discriminate time series changes. Step 4.5: Output layer and loss function.

[0161] First, the fully connected layer outputs a single-day precipitation estimate;

[0162] Secondly, MSE is used as the loss function;

[0163] Finally, the Adam optimizer is used for parameter iteration. Step 4.6: Train the control policy.

[0164] First, use the validation set to monitor model convergence;

[0165] Secondly, the early stopping mechanism avoids overfitting;

[0166] Finally, use GPU training to improve efficiency.

[0167] Further, the step S5 includes:

[0168] Finally, the fused precipitation value is output and the effect is verified, including:

[0169] S501: Input the spatiotemporal sliding samples of the test phase into the trained GConvAttnLSTM model and output the precipitation estimate at each moment

[0170] S502: Introduce the precipitation record of the ground meteorological observation station corresponding to the test set station pixel as the true reference value of precipitation y i , perform error evaluation on the model output;

[0171] S503: Based on the test set and y iAs the basis, the Root Mean Square Error (RMSE), Unbiased Root Mean Squared Error (ubRMSE), Mean Absolute Error (MAE) and Pearson correlation coefficient (PCC) indicators are used to quantify the model accuracy;

[0172] S504: Compare the estimation results of the fusion model with the original remote sensing precipitation product output, and quantitatively evaluate the improvement effect and performance advantage of GConvAttnLSTM in spatio-temporal modeling and fusion accuracy, and verify its actual application value in the multi-source precipitation fusion task.

[0173] In this embodiment, S5 evaluates the model performance in the test set and compares the fusion effect with the original data source.

[0174] Step 5.1: Generate fusion precipitation estimation value.

[0175] First, load the test sample into the model;

[0176] Second, output the corresponding estimated precipitation value;

[0177] Step 5.2: Accuracy verification analysis.

[0178] First, take the ground meteorological observation precipitation data of the test set site pixel as the reference true value;

[0179] Second, calculate the error indicators such as RMSE, ubRMSE, MAE and PCC;

[0180] Finally, evaluate the model accuracy.

[0181] Step 5.3: Compare the original data source.

[0182] First, select the original CHIRPS, MSWEP and PERSIANN-CDR product data as the comparison object;

[0183] Second, analyze point by point in the test set;

[0184] Finally, verify that the result after fusion is better than any single source product.

[0185] Step 5.4: Visual output display.

[0186] First, convert the prediction results to a visual layer;

[0187] Second, perform dynamic rendering on the ArcGIS platform;

[0188] Finally, for regional precipitation fusion analysis and decision support.

[0189] In another embodiment of the present application, a GConvAttnLSTM multi-source remote sensing precipitation estimation system fusing spatio-temporal features is provided, which can be used to implement the GConvAttnLSTM multi-source remote sensing precipitation estimation method described above, and specifically comprises:

[0190] The data preprocessing module is used for spatio-temporal standardization of multi-source remote sensing precipitation data, including removing outliers and missing dates, unifying coordinate systems and spatial resolutions, and screening a common date set of multi-source data.

[0191] The spatio-temporal sample construction module is used for structured feature extraction of precipitation data, including generating an n x n neighborhood spatial window centered on a site or a pixel, constructing a sliding time series sample of consecutive T days, associating a fusion precipitation true value of a target position on the Tth day, and outputting a four-dimensional tensor with a dimension of T x D x n x n.

[0192] The deep learning fusion module is used for collaborative modeling of multi-modal precipitation features, including extracting spatial neighborhood dependency features through graph convolution, enhancing key area response through spatial attention mechanism, modeling time evolution law through LSTM network, aggregating key time period features through time attention mechanism, and outputting precipitation fusion estimation value through a fully connected layer.

[0193] The precision verification and output module is used for quantitative evaluation and business output of fusion results, including accessing ground meteorological observation station true value data, calculating RMSE / ubRMSE / MAE / PCC precision indicators, generating an improved quantitative report of original products and fusion results, and supporting geographic spatial rendering visualization.

[0194] In another embodiment of the present application, a terminal device is provided, which comprises a processor and a memory, the memory being configured to store a computer program, the computer program comprising program instructions, and the processor being configured to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are particularly suitable for loading and executing one or more instructions to implement a corresponding method flow or a corresponding function; the processor in the embodiments of the present application can be used for the operation of the GConvAttnLSTM multi-source remote sensing precipitation estimation method fusing spatio-temporal features.

[0195] In another embodiment of the present application, a memory medium is provided, specifically a computer readable memory medium (Memory), which is a memory device in the terminal device, and is configured to store programs and data. It can be understood that the computer readable memory medium herein can include an internal storage medium in the terminal device, and of course can also include an expansion storage medium supported by the terminal device. The computer readable memory medium provides a storage space, and the storage space stores an operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer readable memory medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory.

[0196] The one or more instructions stored in the computer readable memory medium can be loaded and executed by the processor to implement the corresponding steps of the GConvAttnLSTM multi-source remote sensing precipitation estimation method fusing spatio-temporal features in the above embodiments; the one or more instructions in the computer readable memory medium are loaded and executed by the processor.

[0197] Those skilled in the art will appreciate that embodiments of the present application can be devised for a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0198] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 means for performing the function specified by the flowchart illustrations and / or block diagrams block or blocks.

[0199] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 means for performing the function specified by the flowchart illustrations and / or block diagrams block or blocks.

[0200] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 means for performing the function specified by the flowchart illustrations and / or block diagrams block or blocks.

[0201] Those skilled in the art will appreciate that the embodiments described herein are illustrative of the application and are not meant to limit the scope of the application. Variations and modifications of the embodiments disclosed herein can be made based on the technical teachings of the present disclosure, and that such modifications do not depart from the scope of the present application.

Claims

1. A GConvAttnLSTM multi-source remote sensing precipitation estimation method integrating spatiotemporal features, characterized by: The method comprises: Step S1: Preprocessing of multi-source remote sensing precipitation data, unified resampling, projection conversion and quality control of various remote sensing precipitation products; Step S2: Extract n×n neighborhood precipitation windows centered on the station or pixel to construct spatial window samples; construct spatiotemporal precipitation sequence samples for T consecutive days by sliding on a daily basis to form a three-dimensional tensor structure; Step S3: Input the spatiotemporal precipitation sequence samples into the GConvAttnLSTM model, extract spatial features through graph convolution, extract temporal features through LSTM, and integrate the attention mechanism to predict precipitation; Step S4: Output the fused precipitation result and evaluate the accuracy by combining it with the data from the ground meteorological observation station.

2. The method according to claim 1, wherein: The step S1 comprises: S101: Acquire multi-source remote sensing precipitation data with consistent temporal coverage, and remove outliers and missing dates; S102: Unify the coordinate systems and spatial resolutions of each data source; S103: Filter the dates common to all data sources to generate a sample construction set.

3. The method according to claim 1, wherein: The step S2 comprises: S201: Obtain the coordinates of the grid pixels corresponding to the ground meteorological station; S202: extracting n×n pixels with the station pixel as the center to form a spatial window; S203: Perform window extraction on each data source to form D×n×n dimensional precipitation segments, where D is the number of remote sensing data sources; S204: Repeat the above operation at each time step and each station pixel, and store them as original spatial sample data.

4. The method according to claim 1, wherein: The step S2 also includes time sliding sample construction, which is as follows: S205: Temporally sorting the spatial window samples by date for each station pixel; S206: Using a sliding window method, construct a time series sample with a time length of T days; S207: Each sample is a T×D×n×n four-dimensional tensor, representing the precipitation information of all data sources within the spatial window within T consecutive days; S208: The label corresponding to each input sequence is the fused precipitation value of the target location on the Tth day.

5. The method according to claim 1, wherein: The GConvAttnLSTM model operation in step S3 includes: S401: Daily input X t ∈R T×D×n×n Perform graph convolution: in, is the spatial representation after graph convolution, X t is the remote sensing precipitation input of the current time and space, is a graph structure adjacency matrix based on geographic adjacency, and GCN is a graph convolutional neural network operation; S402: Enhance features through spatial attention mechanism: Among them, A t is the attention weight matrix, π is the spatial position, d is the feature dimension scaling factor, δ is the temperature coefficient, represents pixel-by-pixel multiplication, It is the spatially enhanced input after graph convolution and attention weighting; S403: Sequence Enter LSTM and update the memory state according to formulas (3)-(6), as follows: ① Forget Gate: ② Input gate and candidate memory: ③Memory status update: ④Output gate and current hidden state: Among them, h t-1 is the hidden state at the previous moment, C t-1 is the state of the memory unit at the previous moment, f t For the forget gate, i t is the input gate, is the candidate memory state, C t is the current memory state, o t is the output gate, h t is the current hidden state, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, W,b are the network trainable weights and bias parameters; S404: Apply temporal attention weight α to the LSTM output sequence t : Among them, W α The weight matrix representing the temporal attention mechanism; After weighted aggregation, global time-aware features are obtained S405: The fully connected layer outputs the precipitation estimate: in, It is the global feature representation after temporal attention aggregation, and FC represents the fully connected layer operation; S406: Use mean square error loss function to train the model: in Generate the current precipitation fusion estimation result for the model, y i is the true reference value of precipitation, which comes from the spatial interpolation result of precipitation records at ground meteorological observation stations, and N is the total number of samples in the training / validation set.

6. The method according to claim 1, wherein: The step S4 comprises: S501: Input test samples to the model and output precipitation estimates S502: Using ground meteorological observation station data as reference truth value y i ; S503: Calculate RMSE, unbiased root mean square error ubRMSE, MAE and Pearson correlation coefficient PCC accuracy indicators; S504: Compare with the original remote sensing products and evaluate the fusion effect.

7. The method according to claim 1, wherein: In step S3, the model training adopts a sliding window mechanism, and the input is X t ∈R T×D×n×n Tensor, where T = 7 days, D = 3 types of remote sensing data sources, and n = 9.

8. A GConvAttnLSTM multi-source remote sensing precipitation estimation system integrating spatiotemporal features, characterized by: The system can be used to implement the method according to any one of claims 1 to 7, specifically including: The data preprocessing module is used for the spatiotemporal standardization of multi-source remote sensing precipitation data, including the removal of outliers and missing dates, the unification of coordinate systems and spatial resolutions, and the screening of common date sets for multi-source data. The spatiotemporal sample construction module is used to extract structured features from precipitation data. It generates an n×n neighborhood spatial window centered on a station or pixel, constructs a sliding time series sample for T consecutive days, and associates the fused precipitation value at the target location on the Tth day. The output is a four-dimensional tensor with dimensions of T×D×n×n. The deep learning fusion module is used for collaborative modeling of multimodal precipitation features, including extracting spatial neighborhood dependency features through graph convolution, enhancing key area responses through spatial attention mechanisms, modeling temporal evolution patterns through LSTM networks, aggregating key period features through temporal attention mechanisms, and outputting precipitation fusion estimates through fully connected layers. The accuracy verification and output module is used for quantitative evaluation and operational output of fusion results, including accessing true value data from ground meteorological observation stations, calculating RMSE / ubRMSE / MAE / PCC accuracy indicators, generating improved quantitative reports of original products and fusion results, and supporting geospatial rendering and visualization.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and capable of being run on the processor, wherein when the processor executes the program, the method according to one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images