Rainfall data processing method and device based on multi-source data fusion

Abstract

The invention provides a rainfall data processing method and device based on multi-source data fusion, and the method comprises the steps: inputting multi-satellite fusion remote sensing rainfall observation data, terrain elevation data and latitude and longitude data of a target region at a target time interval into a rainfall data downscaling model, carrying out feature fusion, channel attention enhancement and space attention enhancement processing on the model, and outputting downscaling image data of which the resolution is higher than that of remote sensing rainfall observation data; and correcting the downscaling image data according to the site rainfall observation data corresponding to the plurality of observation sites in the target area at the target time interval to obtain target rainfall observation data of the target area at the target time interval. According to the method, the capturing capability and the multi-modal fusion capability of the model for rainfall distribution information in remote sensing data can be improved, the capturing capability of the model for complex terrains and rainfall characteristics can be improved, and then the expression precision and the application reliability of the rainfall distribution information in the generated high-resolution image can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of data downscaling, and in particular to a precipitation data processing method and device based on multi-source data fusion. Background Art

[0002] Soil erosion assessment is the basis for soil erosion prevention and control and an important way to evaluate the benefits of soil and water conservation. Hydraulic erosion is the main type of soil erosion, and its main driving factor is rainfall erosion. Rainfall erosivity is a dynamic indicator for evaluating the potential ability of rainfall to cause soil erosion. Especially for areas with vast areas, complex internal terrain and obvious spatial heterogeneity of precipitation distribution, it is necessary to calculate rainfall erosivity, which requires high-precision and high-resolution precipitation data. At present, although the development of satellite remote sensing technology and meteorological observation networks has provided many data sources, there are still some challenges in achieving accurate assessment of regional precipitation erosivity.

[0003] At present, the lack of spatial resolution is one of the key challenges faced by remote sensing precipitation data in refined applications, and spatial downscaling is an important technical means to solve this problem. Downscaling is the process of refining large-scale remote sensing data into smaller-scale remote sensing data and converting low-resolution data into high-resolution data. Its core goal is to mine the spatial characteristics of precipitation data through algorithms and models, thereby improving the resolution and application capabilities of the data.

[0004] However, in the existing precipitation data processing method, the SRCNN super-resolution model is usually used to process precipitation data to obtain high-resolution data. However, the SRCNN super-resolution model has a poor ability to capture precipitation distribution information, complex terrain and precipitation characteristics. Therefore, the accuracy and application reliability of the generated high-resolution data are difficult to meet the detail fidelity and accuracy requirements of the precise calculation of soil erosion processes such as water erosion and ablation. Summary of the invention

[0005] In view of this, the embodiments of the present application provide a precipitation data processing method and device based on multi-source data fusion to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present application provides a precipitation data processing method based on multi-source data fusion, comprising:

[0007] Inputting the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement and space attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data, and outputs downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data;

[0008] The downscaled image data is corrected according to the station precipitation observation data corresponding to each of the multiple observation stations in the target area at the target time interval to obtain the target precipitation observation data of the target area at the target time interval.

[0009] In some embodiments of the present application, the step of correcting the downscaled image data according to the precipitation observation data of each of the plurality of observation stations in the target area corresponding to the target time interval to obtain the target precipitation observation data of the target area at the target time interval includes:

[0010] Determine, according to the downscaled image data, the residuals corresponding to the precipitation observation data of each observation station in the target area corresponding to the target time interval;

[0011] Based on the inverse distance weighted average interpolation method, the residuals corresponding to the precipitation observation data of each of the stations are interpolated to obtain a corresponding residual grid map;

[0012] Based on the residual grid map, residual correction is performed on the downscaled image data with preset correction constraints to obtain target precipitation observation data of the target area at a target time interval; wherein the correction constraints include: performing residual correction only on points where precipitation exists in the downscaled image data, and if, after the residual correction is performed on the downscaled image data, there is a point in the downscaled image data with a precipitation value less than 0, setting the precipitation value of the point to 0.

[0013] In some embodiments of the present application, the precipitation data downscaling model includes: a generator constructed based on a channel attention mechanism and a spatial attention mechanism, the generator and a discriminator forming a generative adversarial network;

[0014] Correspondingly, before inputting the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and latitude and longitude data corresponding to the target area at the target time interval into the preset precipitation data downscaling model, it also includes:

[0015] Obtain the pre-processed multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to different regions at each historical time interval;

[0016] Constructing data samples, wherein each of the data samples includes pre-processed multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to a region in a historical time interval;

[0017] Using each of the data samples to perform at least one round of iterative training on the generative adversarial network, and in each round of training, inputting the downscaled image data output by the generator into the discriminator and optimizing the generator according to the determination result output by the discriminator;

[0018] The generator in the generative adversarial network that has completed iterative training is determined as a precipitation data downscaling model.

[0019] In some embodiments of the present application, the generator includes a multimodal input processing module, an attention enhancement module, a residual module and an upsampling reconstruction module connected in sequence; the attention enhancement module includes a channel attention enhancement unit and a spatial attention enhancement unit connected in series;

[0020] The multimodal input processing module is used to perform feature fusion processing on the multi-satellite fused remote sensing precipitation observation data, terrain elevation data and latitude and longitude data corresponding to the target area at the target time interval to obtain a corresponding fusion feature map;

[0021] The channel attention enhancement unit is used to generate a channel attention enhancement feature map corresponding to the fusion feature map based on the channel attention mechanism;

[0022] The spatial attention enhancement unit is used to generate the spatial and channel attention enhancement feature maps corresponding to the channel attention enhancement feature map based on the spatial attention mechanism;

[0023] The residual module is used to perform gradient stabilization processing on the spatial and channel attention enhancement feature maps;

[0024] The upsampling reconstruction module is used to upsample the spatial and channel attention enhancement feature maps after gradient stabilization processing to obtain downscaled image data with a resolution higher than that of the multi-satellite fusion remote sensing precipitation observation data.

[0025] In some embodiments of the present application, the multimodal input processing module includes: a precipitation image feature extraction unit, a terrain feature extraction unit, and a feature stitching unit;

[0026] The precipitation image feature extraction unit is used to perform channel dimension-upgrading feature extraction on the multi-satellite fusion remote sensing precipitation observation data corresponding to the target area at the target time interval based on the first convolution layer and the PReLU activation function in sequence, so as to obtain precipitation observation image feature data corresponding to the multi-satellite fusion remote sensing precipitation observation data;

[0027] The terrain feature extraction unit includes three downsampling layers connected in sequence, and the downsampling layers are used to perform channel dimension up-conversion feature extraction on the terrain elevation data corresponding to the target area at the target time interval and the longitude data and latitude data in the longitude and latitude data in sequence based on the second convolution layer and the PReLU activation function, so that the terrain feature extraction unit outputs the terrain elevation feature data corresponding to the terrain elevation data, the longitude feature data corresponding to the longitude data, and the latitude feature data corresponding to the latitude data respectively;

[0028] The feature stitching layer is used to perform feature fusion on the precipitation observation image feature data, the terrain elevation feature data, the longitude feature data and the latitude feature data in the channel dimension to obtain a corresponding fusion feature map.

[0029] In some embodiments of the present application, the channel attention enhancement unit includes: a global average pooling layer, a global maximum pooling layer, a shared fully connected layer, a first Sigmoid activation layer and a first channel weight addition unit;

[0030] The global average pooling layer is used to extract the global average feature map of the fused feature map;

[0031] The global maximum pooling layer is used to extract the global maximum feature map of the fused feature map;

[0032] The shared fully connected layer is used to perform nonlinear transformation on the global average feature map and the global maximum feature map based on the sequentially connected ReLU activation function and the channel compression unit to obtain a corresponding fully connected feature map;

[0033] The first Sigmoid activation layer is used to generate a channel weight map corresponding to the fully connected feature map within a preset interval based on the Sigmoid function, wherein the preset interval is [0, 1];

[0034] The first channel weight addition unit is used to perform channel weight addition on the channel weight map and the fusion feature map to obtain a corresponding channel attention enhancement feature map.

[0035] In some embodiments of the present application, the spatial attention enhancement unit includes: an average pooling layer, a maximum pooling layer, a channel splicing layer, a third convolutional layer, a second Sigmoid activation layer, and a second channel weight addition unit;

[0036] The average pooling layer is used to extract the average response map of the channel attention enhanced feature map;

[0037] The maximum pooling layer is used to extract the maximum response map of the channel attention enhancement feature map;

[0038] The channel splicing layer is used to splice the average response map and the maximum response map along the channel dimension based on the ReLU activation function and the channel compression unit connected in sequence to obtain a corresponding feature splicing map;

[0039] The third convolutional layer is used to learn the spatial weight distribution of the feature splicing graph to obtain the corresponding spatial feature graph;

[0040] The second Sigmoid activation layer is used to generate a spatial weight map corresponding to the spatial feature map based on a Sigmoid function;

[0041] The second channel weight addition unit is used to perform channel weight addition on the spatial weight map and the channel attention enhancement feature map to obtain the corresponding spatial and channel attention enhancement feature map.

[0042] Another aspect of the present application provides a precipitation data processing device based on multi-source data fusion, comprising:

[0043] A precipitation data downscaling module is used to input the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement and space attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data, and outputs downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data;

[0044] The downscaled image correction module is used to correct the downscaled image data according to the station precipitation observation data corresponding to each of the multiple observation stations in the target area at the target time interval, so as to obtain the target precipitation observation data of the target area at the target time interval.

[0045] The third aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the precipitation data processing method based on multi-source data fusion when executing the computer program.

[0046] A fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the precipitation data processing method based on multi-source data fusion.

[0047] The fifth aspect of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the precipitation data processing method based on multi-source data fusion.

[0048] The precipitation data processing method based on multi-source data fusion provided in the present application inputs the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement and spatial attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data, and outputs downscaled image data with a resolution higher than the multi-satellite fusion remote sensing precipitation observation data; the downscaled image data is corrected according to the precipitation observation data of the station corresponding to the target time interval of multiple observation stations in the target area, so as to obtain the target precipitation observation data of the target area at the target time interval, which can improve the precipitation data downscaling model's ability to capture precipitation distribution information in remote sensing data and multimodal fusion ability, can improve the precipitation data downscaling model's ability to capture complex terrain and precipitation characteristics, can improve the expression accuracy and application reliability of precipitation distribution information in the generated high-resolution image, and thus can provide more effective data support for the accurate calculation of soil erosion processes such as water erosion and ablation.

[0049] Additional advantages, purposes, and features of the present application will be partially described in the following description, and will become partially apparent to those skilled in the art after studying the following, or may be learned from the practice of the present application. The purposes and other advantages of the present application can be achieved and obtained by the structures specifically pointed out in the specification and the drawings.

[0050] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present application are not limited to the above specific description, and the above and other purposes that can be achieved by the present application will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings described herein are used to provide a further understanding of the present application, constitute a part of the present application, and do not constitute a limitation of the present application. The components in the drawings are not drawn to scale, but are only for the purpose of illustrating the principles of the present application. In order to facilitate the illustration and description of some parts of the present application, the corresponding parts in the drawings may be enlarged, that is, they may become larger relative to other components in the exemplary device actually manufactured according to the present application. In the drawings:

[0052] Figure 1 This is a first flow chart of a precipitation data processing method based on multi-source data fusion in one embodiment of the present application.

[0053] Figure 2 This is a schematic diagram of the processing flow of multi-satellite fusion remote sensing precipitation observation data GPM with a resolution of 0.1° HR in an example of this application.

[0054] Figure 3 This is a second flow chart of a precipitation data processing method based on multi-source data fusion in one embodiment of the present application.

[0055] Figure 4 This is a general architecture diagram of a generator in one embodiment of the present application.

[0056] Figure 5 Detailed architecture diagram of a precipitation data downscaling model in one embodiment of the present application.

[0057] Figure 6 Schematic diagram of the structure of a multimodal input processing module in one embodiment of the present application.

[0058] Figure 7 This is a schematic diagram of the structure of a channel attention enhancement unit in one embodiment of the present application.

[0059] Figure 8 This is a structural diagram of a spatial attention enhancement unit in one embodiment of the present application.

[0060] Fig. 9 Schematic diagram of the structure of an identifier in one embodiment of the present application.

[0061] Fig.10 It is a structural schematic diagram of a precipitation data processing device based on multi-source data fusion in one embodiment of the present application.

[0062] Fig.11 This is a structural schematic diagram of a precipitation product generation system in an application example of the present application.

[0063] Fig.12 This is a schematic diagram of the overall processing flow of a precipitation product generation system in an application example of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the implementation modes and the accompanying drawings. Here, the illustrative implementation modes and descriptions of the present application are used to explain the present application, but are not intended to limit the present application.

[0065] It should also be noted here that in order to avoid obscuring the present application due to unnecessary details, only the structures and / or processing steps closely related to the scheme according to the present application are shown in the accompanying drawings, while other details that are not very relevant to the present application are omitted.

[0066] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0067] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0068] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0069] It should be noted that there are three main ways to obtain precipitation observation information: ground rain gauges, ground-based radars, and satellite remote sensing detection information. The precipitation data from ground rain gauges and ground-based radars are relatively accurate and are generally considered to be "true values", but compared with satellite remote sensing detection data, their spatial distribution is uneven. They are suitable for observing local regional precipitation, but it is difficult to obtain more accurate precipitation information for large regions or the world.

[0070] Satellite remote sensing can conduct continuous detection over a large area of ​​space and obtain global precipitation data over a large area. For example, the TRMM satellite is the first meteorological satellite dedicated to observing tropical and subtropical precipitation. It provides free global precipitation data from 50°N to 50°S and multiple time intervals, making up for the lack of precipitation information in areas without data around the world. GPM is a follow-up satellite precipitation program to TRMM, which can provide a new generation of global rain and snow observation data within 3 hours. GPM precipitation products are more accurate and more applicable than previous satellite precipitation products, and can better reflect the temporal and spatial characteristics of precipitation. However, due to the limitations of the physical principles and algorithms of satellite precipitation inversion, the accuracy of its precipitation inversion is still relatively low.

[0071] Precipitation data based on a single source have their own advantages and disadvantages. How to effectively combine the advantages of precipitation data from different sources and develop multi-source precipitation fusion technology has become the mainstream trend in the research and development of high-precision precipitation products in recent years. For regions with vast areas, complex internal terrain and obvious spatial heterogeneity of precipitation distribution, precipitation data is required to be not only high-resolution, but also able to capture the spatial changes at the microscale in the region. The current mainstream global precipitation product GPM remote sensing precipitation data cannot meet this demand in terms of spatial resolution and accuracy. Therefore, it is necessary to study the downscaling model of precipitation data in such regions, combining GPM remote sensing precipitation data and ground rain gauge observation data to obtain more accurate and detailed precipitation information. If high-precision and high-resolution precipitation data can be generated by fusing terrain feature data, ground station data and GPM remote sensing precipitation data, the accuracy of regional soil and water loss assessment will be further improved, providing more targeted measures and decision-making support for soil and water conservation work. At present, research on daily downscaling of regional precipitation remote sensing data is still relatively scarce, and the existing traditional methods have certain limitations when processing precipitation data in this region. Therefore, it is particularly important to carry out downscaling research on regional daily precipitation remote sensing data.

[0072] The current mainstream remote sensing precipitation products are GPM precipitation data and TRMM precipitation data. In terms of spatial resolution, the spatial resolution of TRMM precipitation data is 0.25°*0.25°, and the spatial resolution of GPM data is 0.1°*0.1°. In terms of accuracy, for areas located in high latitudes, TRMM data has large errors in high latitudes due to the physical characteristics of its satellite. Although GPM data can accurately obtain precipitation information within 65° north and south latitude of the earth compared to TRMM data, there is still a big gap in accuracy between it and ground observation station data. Although the accuracy of ground observation station data is high enough, it is not continuous in space. The existing precipitation products cannot meet the needs of the project in terms of resolution and accuracy. Therefore, the goal is to downscale the daily-scale GPM remote sensing data with a resolution of 0.1°*0.1°, and fuse and correct it with the station data to generate higher resolution and more accurate remote sensing data.

[0073] Precipitation data includes ground station data and satellite remote sensing data. Satellite remote sensing data is obtained by inverting microwave / infrared sensors carried on meteorological satellites (such as GPM, TRMM, FY-4A). It has the advantages of complete spatial coverage and high temporal and spatial resolution, but there are problems of sensor errors and cloud interference. Commonly used data sets include:

[0074] CMORPH (CPC MORPHing technique): integrates multi-satellite microwave observations with a temporal resolution of 30 minutes and a spatial resolution of 8 km.

[0075] IMERG (Integrated Multi-satellitE Retrievals): GPM core product, providing high-resolution precipitation with a time resolution of 0.1°×0.1° and 30 minutes.

[0076] PERSIANN-CDR: An inversion algorithm based on artificial neural network, covering the 60°S-60°N area, with a daily-scale data spatial resolution of 0.25°.

[0077] The lack of spatial resolution is one of the key challenges faced by remote sensing precipitation data in refined applications, and spatial downscaling is an important technical means to solve this problem. Downscaling is the process of refining large-scale remote sensing data into smaller-scale remote sensing data and converting low-resolution data into high-resolution data. Its core goal is to mine the spatial characteristics of precipitation data through algorithms and models, thereby improving the resolution and application capabilities of the data. Depending on the method, downscaling technology can be mainly divided into the following categories:

[0078] 1. Interpolation method

[0079] Interpolation is one of the earliest traditional methods used in downscaling. It estimates unknown areas through known low-resolution data points to generate high-resolution data. This method includes a variety of interpolation algorithms, such as nearest neighbor interpolation, bilinear interpolation, bicubic interpolation, etc. Although the calculation is simple and efficient, the interpolation method tends to ignore the spatial heterogeneity of precipitation distribution, and the generated data often lacks fine spatial features and has limited applicability.

[0080] 2. Multi-parameter statistical downscaling

[0081] Multi-parameter statistical downscaling achieves downscaling by introducing environmental variables related to precipitation (such as topography, vegetation index, temperature, etc.) and establishing statistical models to analyze the relationship between these variables and precipitation. This method uses linear regression, stepwise regression or other statistical models to refine and optimize the spatial distribution of precipitation in combination with environmental factors. Its advantage is that it can improve the downscaling accuracy by combining multi-source data, but the effect of the model depends on the quality of the environmental factors and the expression ability of the model, and it is usually weak in the performance of nonlinear relationships.

[0082] 3. Multi-parameter physical downscaling

[0083] The multi-parameter physical downscaling method is based on physical models and simulates the spatial variation of precipitation by modeling physical factors such as topography, atmospheric dynamics and climate characteristics. For example, the distribution characteristics of precipitation are simulated by using the lifting effect of topography on airflow. This method has a solid theoretical basis and is suitable for studying areas with significant topographic influences, but it has high computational complexity and requires high model construction and parameter settings.

[0084] 4. Spatiotemporal fusion downscaling

[0085] The spatiotemporal fusion downscaling method uses time series data and spatial distribution laws to downscale remote sensing data by integrating information from time and space dimensions. This method can jointly optimize time and space and is particularly suitable for the processing and analysis of long time series data. By utilizing the spatiotemporal coupling characteristics, spatiotemporal fusion downscaling can better capture the dynamic changes of precipitation, but its computational cost is high and requires high complexity of the model algorithm.

[0086] 5. Dynamic downscaling

[0087] The dynamical downscaling method is based on numerical meteorological models (such as regional climate models) and refines large-scale climate data through nested simulations. For example, the output of the global climate model (GCM) can be downscaled to higher-resolution data through the regional climate model (RCM). Dynamical downscaling can combine observational data and meteorological process simulations at the same time, and performs well in processing data over a large range and over a long time span, but its computational complexity is huge and it is suitable for high-precision scenarios that require combined physical process research.

[0088] 6. Super-resolution downscaling

[0089] Super-resolution downscaling is a new downscaling method based on deep learning technology. It converts low-resolution remote sensing data into high-resolution data through deep learning models such as convolutional neural networks (CNN) and generative adversarial networks (GAN). This method simulates the idea of ​​super-resolution processing of computer images, can extract complex nonlinear features, and generate high-resolution precipitation data with fine structures. Compared with traditional methods, super-resolution downscaling has significant advantages in capturing the spatial details and heterogeneity of precipitation. However, the accuracy and physical interpretability of super-resolution downscaling results still need to be improved.

[0090] A type of experimental scheme similar to the present application adopts a two-stage processing method of "convolutional neural network downscaling + site data correction". In the convolutional neural network downscaling stage, the SRCNN super-resolution model is used to process precipitation data, improve the spatial resolution of remote sensing precipitation data, and generate more detailed high-resolution data. In this stage, environmental factors such as altitude information, longitude and latitude information, and land cover type are introduced as auxiliary inputs. In the site data correction stage, the generated high-resolution precipitation data is compared and analyzed with the ground observation data, and the residual of each observation site is calculated (i.e., the generated value minus the site observation value). The site residual is used to generate a residual grid map through the interpolation method, and then the interpolated residual grid is superimposed and corrected with the high-resolution generated data. However, the downscaling model used in this method has poor ability to capture precipitation distribution information, complex terrain and precipitation characteristics, so the accuracy and application reliability of the generated high-resolution data are difficult to meet the detailed fidelity and accuracy requirements of the precise calculation of soil erosion processes such as water erosion and ablation.

[0091] Based on this, in order to solve the problems existing in the existing precipitation data processing methods, such as the poor ability of the downscaling model to capture precipitation distribution information, complex terrain and precipitation characteristics, which makes it difficult for the accuracy and application reliability of the generated high-resolution data to meet the detail fidelity and accuracy requirements of the precise calculation of soil erosion processes such as water erosion and ablation, the embodiments of the present application respectively provide a precipitation data processing method based on multi-source data fusion, a precipitation data processing device based on multi-source data fusion for executing the precipitation data processing method based on multi-source data fusion, a physical device, a computer-readable storage medium and a computer program product.

[0092] The details are described in detail through the following examples.

[0093] Based on this, the embodiment of the present application provides a precipitation data processing method based on multi-source data fusion, which can be implemented by a precipitation data processing device based on multi-source data fusion. Figure 1 The precipitation data processing method based on multi-source data fusion specifically includes the following contents:

[0094] Step 100: Input the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement, and spatial attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data, and outputs downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data.

[0095] In one or more embodiments of the present application, the target area refers to the regional scope for a certain area, and its size can be artificially set according to the actual application requirements and the collection scope of multi-satellite fusion remote sensing precipitation observation data; the target time interval can be set to a number of seconds, minutes, hours or days, etc., which can be set specifically according to the actual application requirements.

[0096] The precipitation data used in the embodiments of the present application mainly include two types of observation sources: site precipitation observation data and multi-satellite fusion remote sensing precipitation observation data. The two types of data complement each other in terms of temporal and spatial resolution and coverage, and can provide multi-dimensional information support for precipitation downscaling modeling.

[0097] The multi-satellite fusion remote sensing precipitation observation data in the embodiment of the present application can use GPM remote sensing precipitation observation data, namely GPM IMERG (Integrated Multi-satellite Retrievals for GPM), which is one of the core products of the Global Precipitation Measurement Program (GPM) jointly developed by NASA and JAXA. The data integrates microwave and infrared remote sensing observations from multiple satellites (including GPM core satellites and other cooperative satellites) and integrates ground rain gauge data. Through algorithm calibration and optimization, precipitation estimates covering the global 60°S-60°N range are generated. Its spatial resolution is 0.1°×0.1° and its temporal resolution is 30 minutes.

[0098] It should be noted that the terrain elevation data and longitude and latitude data used in the embodiments of the present application are environmental factor data, and other parameters such as coastline distance can be added to the environmental factor data. However, in order to maintain the continuity of terrain features while improving image resolution and to minimize the application and training complexity of the precipitation data downscaling model, the embodiments of the present application only select terrain elevation data and longitude and latitude data as environmental factor data.

[0099] It is understandable that the terrain elevation data can select the digital elevation model data DEM. DEM is a digital data model used to represent the terrain on the earth's surface. It stores terrain height information through regular grids or irregular triangulated networks, and each pixel corresponds to a specific altitude value. DEM only contains elevation data of exposed terrain, and does not include surface covers such as buildings and vegetation. It is often used in remote sensing analysis, geographic information systems (GIS), hydrological research, environmental monitoring and other fields. DEM data can be obtained through satellite remote sensing, aerial photography, laser radar, etc., and is processed by interpolation or filtering to improve accuracy. As an additional input feature, DEM can provide terrain constraint information for the model, so that the super-resolution generation network can maintain the consistency of terrain features when improving the image resolution.

[0100] In step 100, the downscaling model used in this application has stronger multimodal fusion capabilities. It not only introduces environmental factors such as DEM and longitude and latitude information, but also enables the precipitation data downscaling model to dynamically adjust the weight relationship between various environmental factors and precipitation characteristics by constructing multimodal input, channel attention mechanism and spatial attention mechanism. This design can more accurately capture the impact of environmental factors on precipitation distribution during the generation process and avoid feature loss caused by simple linear superposition.

[0101] Specifically, the embodiment of the present application introduces a channel attention mechanism and a spatial attention mechanism, and performs feature enhancement in a serial manner. The channel attention mechanism can automatically learn the importance distribution of different environmental factors and precipitation data channels, and the spatial attention mechanism can focus more on capturing information about local precipitation areas. This design can further enhance the downscaling model's ability to capture complex terrain and precipitation features, and improve the detail fidelity and accuracy of the generated results.

[0102] Step 200: Correcting the downscaled image data according to the precipitation observation data corresponding to each of the plurality of observation sites in the target area at the target time interval, so as to obtain the target precipitation observation data of the target area at the target time interval.

[0103] The observation site can be a meteorological site on the ground. The site precipitation observation data can be directly measured by the rain gauge at the meteorological site. It has the characteristics of strong temporal continuity and high observation accuracy, and is usually used as the "ground truth" for precipitation data verification. The original site data is distributed in the form of discrete points, and the spatial coverage is limited by the site density. In particular, there is a problem of spatial discontinuity in complex terrain areas (such as mountainous areas and plateaus).

[0104] That is to say, the embodiment of the present application is divided into two stages: step 100 and step 200. The first stage studies a general precipitation data downscaling model to refine the spatial resolution of the GPM precipitation data from 0.1°×0.1° to a higher resolution. In this stage, a deep learning model can be used as a precipitation data downscaling model to construct a general precipitation data downscaling model suitable for most areas. This model not only improves the spatial resolution of remote sensing data, but also optimizes the downscaling results by introducing auxiliary environmental variables such as DEM data and longitude and latitude information, so that the generated high-resolution data can well capture the precipitation distribution information in the region.

[0105] The second stage of processing integrates and corrects the precipitation data to further improve the accuracy of the data. In this stage, the GPM data is integrated and corrected with the ground observation station data, combining the high accuracy of the station data and the spatial coverage advantage of the GPM data to generate a more accurate precipitation product.

[0106] In one example, see Figure 2 , the multi-satellite fusion remote sensing precipitation observation data GPM with a resolution of 0.1°HR, the terrain elevation data DEM with a resolution of 0.0125°HR, and the longitude data and latitude data are respectively input into the precipitation data downscaling model for downscaling model processing, so that the precipitation data downscaling model outputs precipitation data of 0.0125°HR with a resolution increased by 8 times (i.e., downscaled image data); then the station data (i.e., station precipitation observation data) and the precipitation data with a resolution of 0.0125°HR are fused and corrected, thereby obtaining 0.0125° high-resolution and high-precision precipitation data (i.e., target precipitation observation data) with higher accuracy.

[0107] The two-stage processing is closely related in logic. The first-stage general precipitation data downscaling model effectively improves the resolution of GPM data, providing more refined basic data for the second-stage fusion correction; and the second-stage fusion correction further improves the accuracy of the general downscaling results, and ultimately provides higher-quality precipitation products. Finally, based on these two-stage processing methods, the embodiment of the present application can build a high-quality precipitation product generation system, which can perform full-process processing from downscaling, fusion correction to final result output of precipitation data, and provide data support for the accurate calculation of soil erosion processes such as water erosion and ablation.

[0108] From the above description, it can be seen that the precipitation data processing method based on multi-source data fusion provided in the embodiment of the present application can improve the precipitation data downscaling model's ability to capture precipitation distribution information in remote sensing data and its multimodal fusion capability, can improve the precipitation data downscaling model's ability to capture complex terrain and precipitation characteristics, can improve the expression accuracy and application reliability of precipitation distribution information in the generated high-resolution images, and can provide more effective data support for the precise calculation of soil erosion processes such as water erosion and ablation.

[0109] In order to improve the accuracy of precipitation products while ensuring temporal and spatial coverage, many studies in recent years have integrated information from various data sources and proposed a variety of technologies for fusing satellite remote sensing data and ground observation data. These methods include geographically weighted regression (GWR), optimal interpolation, Bayesian estimation, Kalman filter algorithm, probability density function matching (PDFMatching), machine learning, etc. Among them, geographically weighted regression can capture local changes in space through weighted processing, which is suitable for spatial autocorrelation of precipitation data; optimal interpolation methods (such as Kriging interpolation) perform spatial interpolation through the covariance matrix to fill in the missing areas of data; Bayesian estimation introduces prior knowledge to fuse multi-source observations, but the algorithm complexity is relatively high; Kalman filter algorithm has a significant effect on modeling time series data, but has limited performance for nonlinear precipitation processes; probability density function matching adjusts the probability distribution of satellite or radar precipitation to be consistent with ground observations, and corrects system deviations from a global perspective; machine learning methods (such as deep learning) can capture complex nonlinear relationships between multi-source data, but are prone to overfitting in areas with sparse data.

[0110] Compared with the above methods, residual fusion correction has unique advantages: by calculating the residuals between generated data and observed data, and using interpolation to generate residual fields, the non-stationarity of precipitation in time and space can be dynamically reflected, and the correction results are transparent and easy to trace. Residual correction does not require the preset of complex physical models and has a higher tolerance for data sparsity. At the same time, this application optimizes the residual fusion correction in combination with the physical characteristics of actual precipitation, and uses this method to further process the downscaled data.

[0111] Based on this, in order to further improve the effectiveness, rationality and reliability of the downscaled image data corrected by the site precipitation observation data, so as to further improve the authenticity and application reliability of the target precipitation observation data, in a precipitation data processing method based on multi-source data fusion provided in an embodiment of the present application, see Figure 3 , step 200 in the precipitation data processing method based on multi-source data fusion specifically includes the following contents:

[0112] Step 210: Determine the residuals corresponding to the precipitation observation data of each observation station in the target area at the target time interval according to the downscaled image data.

[0113] Specifically, the generated high-resolution downscaled image data and the station precipitation observation data corresponding to the target time interval of each observation station in the target area are processed into the same geographic coordinate system. On this basis, the residual value of each observation station is calculated, and the residual is defined as the difference between the generated value (i.e., the downscaled image data) and the station observation value (i.e., the station precipitation observation data). Through the calculation of the residual, the error distribution of the generated data at each observation station is determined.

[0114] Step 220: interpolating the residuals corresponding to the precipitation observation data of each of the stations based on the inverse distance weighted average interpolation method to obtain a corresponding residual grid map.

[0115] Specifically, the inverse distance weighted average interpolation method refers to IDW interpolation, which is a weighted average interpolation method based on the inverse of distance. Its basic assumption is that spatially adjacent data points have similar properties, that is, the correlation between data points decreases as the distance increases. For the points to be interpolated, the IDW method will give these known points different weights according to their distance from the surrounding known points. The closer the distance, the greater the weight.

[0116]

[0117] (1) Represents the interpolation result at the position to be estimated x.

[0118] (2) z i Represents the observed value of known point i.

[0119] (3)d i Represents the distance between the estimated position x and the known point i.

[0120] (4) p represents the power parameter of the distance weight, which is usually 2. The larger the value of p, the greater the influence of points close to each other on the interpolation.

[0121] IDW interpolation is suitable for data with strong spatial autocorrelation. Precipitation data has obvious spatial autocorrelation, that is, the precipitation conditions in adjacent areas are often similar. IDW interpolation takes advantage of this feature and effectively retains the spatial distribution characteristics of precipitation data by weighting adjacent points. At the same time, compared with other complex interpolation methods (such as Kriging interpolation or Bayesian estimation), IDW has lower computational complexity, is easy to understand and implement, and is especially suitable for rapid processing of large-scale data.

[0122] Based on this, in step 220, the residuals of the observation points are interpolated into a complete residual grid map using the IDW interpolation method. IDW interpolation distributes the residuals of closer observation points to the surrounding area to a greater extent by weighting the distance, thereby generating a residual map with reasonable spatial distribution. Specifically, the interpolation process is as follows:

[0123] For each grid pixel value, calculate its distance w to all residual points i ;

[0124]

[0125] The weights are assigned using the inverse distance weight formula, where di is the distance from the ith observation station to the target grid, and p is the weight index (here the value is 2);

[0126] The interpolation value Z of the target grid is calculated using the weighted average formula:

[0127]

[0128] where z i is the residual value of the ith observation point, n is the total number of observation sites, and can be set according to the actual number of observation sites in the target area. In an example, the total number of observation sites can be set to no less than 100.

[0129] Step 230: Based on the residual grid map, residual correction is performed on the downscaled image data with preset correction constraints to obtain target precipitation observation data of the target area at a target time interval; wherein the correction constraints include: residual correction is performed only on points where precipitation exists in the downscaled image data, and if, after residual correction is performed on the downscaled image data, there is a point in the downscaled image data with a precipitation value less than 0, the precipitation value of the point is set to 0.

[0130] Specifically, in the residual correction process, considering the physical characteristics of the actual precipitation distribution, the following constraints are set:

[0131] Correction is only performed on areas with precipitation in the original raster data (i.e., downscaled image data): avoid inserting residual values ​​in areas without precipitation (grids with precipitation of 0) to ensure the physical rationality of the correction results.

[0132] Prevent negative precipitation: During correction, if the interpolation residual correction results in a precipitation value less than 0, it is forced to be corrected to 0. This process is achieved through the following formula:

[0133] Z 校正 =max(Z 生成 +Z 残差 ,0)

[0134] Where Z 生成 is the original generated data, Z 残差 is the residual value obtained by interpolation.

[0135] Through the constrained calculation of residual correction, the final corrected high-resolution precipitation data (i.e., target precipitation observation data) are obtained. The corrected data retains the spatial distribution characteristics of the original generated data, while the residual correction significantly reduces the systematic error.

[0136] That is to say, in the site correction stage, the embodiment of the present application calculates and corrects the residuals, and proposes to perform residual correction only in the precipitation areas in the original high-resolution data, while no residuals are artificially added in the non-precipitation areas. This strategy avoids the introduction of false precipitation data in precipitation-free areas due to interpolation errors, ensuring the physical rationality of the correction results. The present application also adds a rationality check on the corrected precipitation data during the correction process to ensure that negative precipitation values ​​will not appear after correction, which is more in line with the actual precipitation situation and avoids the impact of abnormal precipitation data.

[0137] In order to further improve the precipitation data downscaling model's ability to capture precipitation distribution information in remote sensing data and its multimodal fusion capability, in a precipitation data processing method based on multi-source data fusion provided in an embodiment of the present application, the precipitation data downscaling model includes: a generator constructed based on a channel attention mechanism and a spatial attention mechanism, and the generator and a discriminator constitute a generative adversarial network.

[0138] Specifically, the Generative Adversarial Network (GAN) is an unsupervised learning model. GAN generates high-quality data through adversarial training between two neural networks, the Generator and the Discriminator. Due to its outstanding performance in image generation, data enhancement, and super-resolution, GAN has become an important technology in the field of image generation models.

[0139] The generative adversarial network consists of two core modules: the generator and the discriminator. Through adversarial training, the two promote each other's improvement and form a game relationship. The goal of the generator is to receive random noise (such as standard normal distribution) and generate realistic high-resolution precipitation images through neural networks. In this application, the generator converts low-resolution precipitation data into high-resolution images through multi-layer convolutional networks and residual modules. The goal of the discriminator: The goal of the discriminator is to distinguish whether the input image comes from real observation data or a high-resolution image generated by the generator. The discriminator extracts the features of the input image through a convolutional neural network and outputs a probability indicating the authenticity of the input data.

[0140] Due to the physical limitations of current remote sensing satellites, the resolution of satellite remote sensing precipitation data cannot be improved. The mainstream approach is to improve the resolution of precipitation products through downscaling models or algorithms. Traditional downscaling models and algorithms mainly include physical downscaling and dynamic downscaling, which usually rely on large-scale climate models (such as global climate models), require high computational costs, are relatively complex in structure, and require complex geographical and meteorological knowledge. The downscaling problem in the field of remote sensing is somewhat similar to super-resolution processing in the field of computer image processing. Some scholars even use super-resolution deep learning methods to solve the downscaling problem. However, the high-resolution precipitation data generated in this way does not actually meet the requirements. Although it has good detailed information, it does not take into account other environmental variables and the complex relationship of precipitation in the entire meteorological system. The precipitation data represented by these detailed information is inaccurate and not interpretable. At present, traditional downscaling deep learning models tend to treat downscaling problems as ordinary super-resolution problems, and rarely combine the characteristics of precipitation data to further improve the model structure.

[0141] Therefore, the goal of this stage is to build a deep learning downscaling model that integrates terrain features, and increase the spatial resolution of GPM precipitation data by 8 times, from 0.1° to 0.0125°. By solving the limitation of traditional super-resolution methods that rely only on precipitation data, a generative adversarial network (TFGAN) based on terrain feature fusion is proposed, which integrates DEM terrain elevation data, longitude and latitude information, and distance from the coastline as environmental factor data. It also dynamically adjusts the correlation between terrain information and precipitation characteristics by incorporating channel attention mechanism and spatial attention mechanism, enhances feature learning of local precipitation areas, and solves the problems of poor interpretability and detail distortion of existing deep learning downscaling models. The spatial resolution of GPM precipitation data is increased by 8 times, from 0.1° to 0.0125°. While improving the resolution, the root mean square error and mean absolute error of precipitation data are reduced, and indicators such as structural similarity and correlation index are improved to meet the application needs of precipitation data such as erosion simulation and agricultural drought monitoring.

[0142] In the past, research on neural network models for downscaling problems mostly focused on convolutional neural network models. Traditional convolutional neural network models often cannot extract precipitation data features well when solving downscaling problems. The generative adversarial network, through continuous adversarial training of the generator and the discriminator, can make the generated images closer and closer to the real images in this process.

[0143] SRGAN is a super-resolution generative adversarial network that can restore low-resolution (LR) images to high-resolution (HR) images. Compared with the traditional super-resolution method based on mean square error (MSE), SRGAN adopts the generative adversarial network (GAN) framework, which makes the generated images closer to the real high-resolution images in terms of details and perceptual quality. The network is effective in solving the problem of computer image super-resolution, so the SRGAN model is considered to be applied to the downscaling problem. SRGAN mainly consists of two core parts. The generator is responsible for generating super-resolution images from low-resolution images. The discriminator is used to distinguish the generated super-resolution images from the real high-resolution images. However, SRGAN can only extract the characteristics of precipitation data itself and can only optimize precipitation data at the image level. In order to better learn the spatial variation law and terrain influence of precipitation data, this paper introduces environmental variables such as DEM elevation data and longitude and latitude information as auxiliary information of terrain features into the model, integrates channel attention mechanism and spatial attention mechanism as attention enhancement modules to improve feature learning ability, and constructs a generative adversarial network precipitation data downscaling model based on terrain feature fusion-TFGAN model (Topography-Fusion GAN).

[0144] For the corresponding Figure 3 The precipitation data processing method based on multi-source data fusion further specifically includes the following contents before step 100:

[0145] Step 010: Obtain the pre-processed multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to different regions at each historical time interval.

[0146] Step 020: construct each data sample, wherein each of the data samples contains pre-processed multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to a region in a historical time interval.

[0147] Step 030: Use each of the data samples to perform at least one round of iterative training on the generative adversarial network, and in each round of training, input the downscaled image data output by the generator into the discriminator and optimize the generator according to the judgment result output by the discriminator.

[0148] Step 040: Determine the generator in the generative adversarial network that has completed iterative training as a precipitation data downscaling model.

[0149] In other words, compared to directly using the traditional SRCNN super-resolution model, this application proposes a generative adversarial network (TFGAN) that integrates multiple improved modules, which achieves a stronger ability to capture precipitation distribution information and has stronger multimodal fusion capabilities. In the generator model, this application not only introduces environmental factors such as terrain (DEM) and longitude and latitude information, but also enables the model to dynamically adjust the weight relationship between various environmental factors and precipitation characteristics by constructing a multimodal input module and a channel attention mechanism and a spatial attention mechanism. This design can more accurately capture the impact of environmental factors on precipitation distribution during the generation process and avoid feature loss caused by simple linear superposition.

[0150] In order to further improve the application effectiveness and reliability of the precipitation data downscaling model, in a precipitation data processing method based on multi-source data fusion provided in an embodiment of the present application, see Figure 4 , the generator includes a multimodal input processing module 1, an attention enhancement module 2, a residual module 3 and an upsampling reconstruction module 4 connected in sequence; the attention enhancement module 2 includes a channel attention enhancement unit 21 and a spatial attention enhancement unit 22 connected in series;

[0151] The multimodal input processing module 1 is used to perform feature fusion processing on the multi-satellite fused remote sensing precipitation observation data, terrain elevation data and latitude and longitude data corresponding to the target area at the target time interval to obtain a corresponding fusion feature map;

[0152] The channel attention enhancement unit 21 is used to generate a channel attention enhancement feature map corresponding to the fusion feature map based on the channel attention mechanism;

[0153] The spatial attention enhancement unit 22 is used to generate the spatial and channel attention enhancement feature maps corresponding to the channel attention enhancement feature map based on the spatial attention mechanism;

[0154] The residual module 3 is used to perform gradient stabilization processing on the spatial and channel attention enhancement feature maps;

[0155] The upsampling reconstruction module 4 is used to upsample the spatial and channel attention enhancement feature maps after gradient stabilization processing to obtain downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data.

[0156] In order to further improve the application effectiveness and reliability of the multimodal input processing module, in a precipitation data processing method based on multi-source data fusion provided in an embodiment of the present application, see Figure 5 and Figure 6 , the multimodal input processing module 1 includes: a precipitation image feature extraction unit 11, a terrain feature extraction unit 12 and a feature stitching unit 13;

[0157] The precipitation image feature extraction unit 11 is used to perform channel dimension-upgrading feature extraction on the multi-satellite fusion remote sensing precipitation observation data corresponding to the target area at the target time interval based on the first convolutional layer and the PReLU activation function in sequence, so as to obtain precipitation observation image feature data corresponding to the multi-satellite fusion remote sensing precipitation observation data.

[0158] The terrain feature extraction unit 12 includes three downsampling layers connected in sequence, and the downsampling layers are used to perform channel dimension upscaling feature extraction on the terrain elevation data corresponding to the target area at the target time interval and the longitude data and latitude data in the longitude and latitude data based on the second convolutional layer and the PReLU activation function, so that the terrain feature extraction unit outputs the terrain elevation feature data corresponding to the terrain elevation data, the longitude feature data corresponding to the longitude data, and the latitude feature data corresponding to the latitude data.

[0159] The feature stitching layer 13 is used to perform feature fusion on the precipitation observation image feature data, the terrain elevation feature data, the longitude feature data and the latitude feature data in the channel dimension to obtain a corresponding fusion feature map.

[0160] Among them, the first convolution layer is a 9×9 convolution layer, and the second convolution layer is a 3×3 convolution layer. Specifically, the multimodal input processing module receives low-resolution (LR) images (i.e., multi-satellite fusion remote sensing precipitation observation data), DEM data, latitude data, and longitude data as input. The multi-satellite fusion remote sensing precipitation observation data is upgraded from 1 channel to 64 channels through a 9×9 convolution layer, and the PReLU activation function is connected after the convolution operation to enhance the nonlinear feature extraction capability. At the same time, the DEM data, latitude, and longitude are respectively extracted through independent terrain feature extraction units. These data are subjected to three layers of 3×3 convolution operations, each layer is downsampled with a step size of 2, and the spatial dimension is gradually compressed. After each convolution, the ReLU activation function is used to improve the feature extraction effect. After that, the precipitation observation image feature data, the terrain elevation feature data, the longitude feature data, and the latitude feature data are feature fused in the channel dimension in the feature splicing layer to form a fusion feature map containing 256 channels.

[0161] In order to further improve the application effectiveness and reliability of the channel attention enhancement unit, in a precipitation data processing method based on multi-source data fusion provided in an embodiment of the present application, see Figure 5 and Figure 7The channel attention enhancement unit 21 includes: a global average pooling layer 211, a global maximum pooling layer 212, a shared fully connected layer 213, a first Sigmoid activation layer 214 and a first channel weight addition unit 215.

[0162] The global average pooling layer 211 is used to extract the global average feature map of the fused feature map.

[0163] The global maximum pooling layer 212 is used to extract the global maximum feature map of the fused feature map.

[0164] The shared fully connected layer 213 is used to perform nonlinear transformation on the global average feature map and the global maximum feature map based on the ReLU activation function and the channel compression unit connected in sequence to obtain a corresponding fully connected feature map.

[0165] The first Sigmoid activation layer 214 is used to generate a channel weight map corresponding to the fully connected feature map within a preset interval based on the Sigmoid function, wherein the preset interval is [0, 1].

[0166] The first channel weight addition unit 215 is used to perform channel weight addition on the channel weight map and the fusion feature map to obtain a corresponding channel attention enhancement feature map.

[0167] Specifically, the attention enhancement module that applies the attention mechanism to the spliced ​​256-channel feature map adjusts the feature weight distribution in channels and space. The attention enhancement module applies both the channel attention mechanism and the spatial attention mechanism. Channel attention learns the importance of different features, while spatial attention strengthens the model's feature learning of key areas of the image.

[0168] The channel attention module first extracts global information from the input features, using the average pooling and maximum pooling dual paths working in parallel. The average pooling path calculates the global average of each channel, representing the overall feature distribution of the channel, while the maximum pooling dual path captures local significant features. Then the shared fully connected layer (including ReLU activation and 16x channel compression) learns the nonlinear dependencies between channels. Finally, the Sigmoid function generates a channel weight map in the range of 0-1 to reflect the importance of each channel. The introduction of the channel attention mechanism can strengthen the correlation between the environmental factor channel and the LR image texture channel. For example, in mountainous terrain scenes, the weight ratio of the dem elevation model data related channels can be significantly improved. For the input fusion feature map X1∈R C×H×W :

[0169] w avg =σ(W 2 (δ(W 1 (AvgPool(X1)))))

[0170] w max =σ(W 2 (δ(W 1 (MaxPool(X1)))))

[0171]

[0172]

[0173] Among them, w avg Global average feature map; w max Global maximum feature map; W 1 , W 2 is the weight of the fully connected layer, δ is ReLU, σ is Sigmoid, is the channel-level multiplication; AvgPool(X) represents the global average pooling for the fused feature map; (MaxPool(X)) represents the global maximum pooling for the fused feature map; W channel is the channel weight map; X out Enhance feature maps for channel attention.

[0174] In order to further improve the application effectiveness and reliability of the channel attention enhancement unit, in a precipitation data processing method based on multi-source data fusion provided in an embodiment of the present application, see Figure 5 and 8 , the spatial attention enhancement unit 22 includes: an average pooling layer 221, a maximum pooling layer 222, a channel splicing layer 223, a third convolutional layer 224, a second Sigmoid activation layer 225 and a second channel weight addition unit 226;

[0175] The average pooling layer 221 is used to extract the average response map of the channel attention enhancement feature map;

[0176] The maximum pooling layer 222 is used to extract the maximum response map of the channel attention enhancement feature map;

[0177] The channel splicing layer 223 is used to splice the average response map and the maximum response map along the channel dimension based on the ReLU activation function and the channel compression unit connected in sequence to obtain a corresponding feature splicing map;

[0178] The third convolutional layer 224 is used to learn the spatial weight distribution of the feature splicing map to obtain the corresponding spatial feature map;

[0179] Among them, the third convolutional layer is implemented using a 7×7 large convolution kernel.

[0180] The second Sigmoid activation layer 225 is used to generate a spatial weight map corresponding to the spatial feature map based on the Sigmoid function;

[0181] The second channel weight addition unit 226 is used to perform channel weight addition on the spatial weight map and the channel attention enhancement feature map to obtain the corresponding spatial and channel attention enhancement feature map.

[0182] Specifically, the spatial attention enhancement unit 22 is mainly used for modeling geographic spatial structures. By splicing the average response map and the maximum response map along the channel dimension as a dual-channel input model, the average pooled feature map represents the overall response intensity of each position, and the maximum pooled feature map highlights the significant features of each position. At the same time, the 7×7 large convolution kernel is used to learn the spatial weight distribution with a wide receptive field, which can capture long-distance spatial dependencies. Finally, Sigmoid generates a spatial weight map, focusing on important areas (such as important geographical features such as mountains and rivers), and can capture the geographical feature information of these areas.

[0183] X avg =Mean(X2,dim=1)

[0184] X max =Max(X2,dim=1)

[0185] M spatial =σ(f 7×7 ([X avg ,X max ]))

[0186] X out =X2⊙M spatial

[0187] Among them, X avg represents the average response graph; X max represents the maximum response map; X2 represents the channel attention enhancement feature map; dim represents the dimension, and dim equals 1, which means that this is an operation in the channel dimension; M spatial Represents the spatial feature map; X out Represents the spatial and channel attention enhanced feature map; f7×7 is the 7×7 convolutional layer, and ⊙ is the spatial position multiplication.

[0188] The two attention mechanisms adopt a serial approach, first adjusting the importance distribution of feature channels through channel weights, and then using spatial weights to focus on key geographical areas. This processing enables the model to suppress the inter-channel interference noise introduced by multi-source feature fusion during downscaling and reconstruction, while maintaining the spatial consistency of terrain details and visual features. Compared with the traditional attention mechanism, this design innovatively implements joint attention adjustment after the multimodal feature fusion layer, so that the channel enhancement effect guided by DEM elevation data and the spatial focusing of geographic prior constraints form a synergy, effectively solving the dual challenges of feature alignment and structure preservation in geospatial data super-resolution.

[0189] After the fusion feature map passes through the attention enhancement module, it enters six series-connected residual modules. Each residual module consists of two layers of 3×3 convolution operations. Each convolution is followed by a BatchNorm normalization layer, and then the nonlinear feature representation capability is enhanced through the PReLU activation function. The residual module part uses a skip connection to directly add the initial input features to the final convolution output, retaining the initial feature information of the precipitation data and avoiding the gradient vanishing problem. This allows the model to increase the network depth while ensuring the stability of the training and the network expression capability.

[0190] Finally, there is the upsampling reconstruction module, which can be composed of multiple PixelShuffle modules. Each PixelShuffle module contains a 3×3 convolution operation, which expands the number of channels and improves the spatial resolution by rearranging pixels. After each upsampling, the PReLU activation function is connected to maintain the nonlinear feature capability. The final upsampling module uses a 9×9 convolution operation to compress the feature map to a single channel, output a single-channel high-resolution precipitation image, and normalize the output value through the tanh activation function. The final output value is in the range of [0, 1].

[0191] See also Fig. 9 The discriminator adopts a deep convolutional network design. The input layer includes 3×3 convolutions. LeakyReLU (negative slope 0.2) is still used to introduce nonlinearity. The subsequent layers include 3×3 convolutions, BN layers, and LeakyReLU activations. Batch normalization is inserted after each convolution operation to accelerate convergence, alleviate the problem of gradient outliers, and improve training stability. Then, adaptive average pooling is used at the end to compress the spatial dimension, and two-level 1×1 convolutions are used to achieve feature compression (512→1024→1). Finally, in the model training stage, the probability of a real image or a fake image is output through sigmoid.

[0192] From the software level, the present application also provides a precipitation data processing device based on multi-source data fusion for executing all or part of the precipitation data processing method based on multi-source data fusion, see Fig.10The precipitation data processing device based on multi-source data fusion specifically includes the following contents:

[0193] The precipitation data downscaling module 10 is used to input the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement and space attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data, and outputs downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data;

[0194] The downscaled image correction module 20 is used to correct the downscaled image data according to the station precipitation observation data corresponding to each of the multiple observation stations in the target area at the target time interval, so as to obtain the target precipitation observation data of the target area at the target time interval.

[0195] The embodiment of the precipitation data processing device based on multi-source data fusion provided in the present application can be specifically used to execute the processing flow of the embodiment of the precipitation data processing method based on multi-source data fusion in the above-mentioned embodiment. Its functions are not repeated here, and reference can be made to the detailed description of the above-mentioned precipitation data processing method based on multi-source data fusion embodiment.

[0196] The precipitation data processing device based on multi-source data fusion performs part of the precipitation data processing based on multi-source data fusion in a server or client device. The selection can be made based on the processing capability of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed in the client device, the client device may also include a processor for specific processing of precipitation data processing based on multi-source data fusion.

[0197] The client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side, and other implementation scenarios may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, or a server cluster consisting of multiple servers, or a server structure of a distributed device.

[0198] The server and the client device may communicate with each other using any suitable network protocol, including network protocols that have not yet been developed on the date of filing this application. The network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Of course, the network protocols may also include, for example, RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer) protocols used on top of the above protocols.

[0199] From the above description, it can be seen that the precipitation data processing device based on multi-source data fusion provided in the embodiment of the present application can improve the precipitation data downscaling model's ability to capture precipitation distribution information in remote sensing data and its multimodal fusion capability, can improve the precipitation data downscaling model's ability to capture complex terrain and precipitation characteristics, can improve the expression accuracy and application reliability of precipitation distribution information in the generated high-resolution images, and can provide more effective data support for the precise calculation of soil erosion processes such as water erosion and ablation.

[0200] On this basis, an application example of this application also provides a precipitation product generation system, see Fig.11 The precipitation product generation system is a complete solution that applies the TFGAN downscaling model and fusion correction algorithm mentioned above, integrating data preprocessing and precipitation product generation, aiming to provide high-precision and high-resolution precipitation data support for related applications. The system covers the full process of multi-source data input, processing, downscaling generation and data correction, which can effectively meet the needs of multiple fields under complex meteorological environments.

[0201] The main goal of the data preprocessing module is to clean, process and convert the original input data to ensure that it meets the requirements of the subsequent precipitation product generation module. The core functions include data loading, data resampling and data format conversion. The data loading function supports loading local multi-source precipitation, DEM and other data. It can automatically identify and adapt to various data formats (such as raster data and station data). At the same time, it provides integrity checking functions to ensure the integrity and absence of data. The data resampling function is aimed at auxiliary data that does not meet the resolution requirements. The system supports multiple resampling methods such as bilinear interpolation and bicubic interpolation. The data format conversion function supports conversion between multiple data formats (such as from NC4 to GeoTIFF).

[0202] The precipitation product generation module is the core of the system. It is mainly responsible for generating high-resolution and high-precision precipitation data products. It includes a two-stage processing flow: downscaling processing and fusion correction processing. This module includes functions such as model management, downscaling processing, and fusion correction processing. The model management function supports the loading, switching and updating of multiple generation models, supports the automatic loading of model configurations, and allows users to choose the best model to call in different scenarios. The downscaling processing function uses the generative adversarial network (TFGAN) model based on terrain feature fusion, which can upgrade low-resolution precipitation data to high resolution and significantly improve data details. The model has the ability to process large-scale and long-time span data while ensuring the efficiency and accuracy of calculations. The fusion correction processing function supports the fusion correction of satellite remote sensing data and ground station observation data. The system performs data fusion correction on the generated precipitation products to improve the accuracy and reliability of the products.

[0203] See also Fig.12 ,The overall processing flow of the precipitation product generation system is as follows:

[0204] First, the format, content, resolution and integrity of precipitation files, DEM files, longitude and latitude files and site files are batch checked through the data loading module, and loaded into the system to prepare for subsequent processing. The format checks whether the file extension meets the requirements to prevent users from entering the wrong file. In terms of content, the main check is whether the data is abnormal. For example, the precipitation data checks whether there are negative values ​​or values ​​beyond the reasonable range, the DEM data checks whether the terrain value is consistent with the actual value (such as the height exceeds the range), and the site data checks whether the site number exists and whether the precipitation value is abnormal. In terms of resolution, it mainly checks whether the environmental factor data (DEM, longitude and latitude) is the resolution required by the model (for example, the initial precipitation data resolution is 1x1, and the resolution needs to be increased to 8x8, so the model needs to input environmental factor data with a resolution of 8x8). At the same time, if there are problems with the format, content, resolution and integrity of the file, the output prompt will remind the user and record it in the log.

[0205] Then load the data into the abnormally formatted files (mainly GPM precipitation data, the initial data format is NetCDF), and use the data format conversion function to batch convert the file format to the required GeoTiff format. The converted GeoTIFF file uses the latitude and longitude coordinates, spatial range, grid resolution and other information extracted from the original file by default. The geographic projection coordinate system type allows the user to manually specify it. Use GDAL to convert the format of raster data, and call GDALTranslate to convert the file to the target format. Provide a user interface that allows the user to select the input file and output folder. Support traversing the folder selected by the user and converting all files that meet the specified format in sequence.

[0206] Continue to load data into the wrong resolution file, and use the data resampling function to batch resample the files to the required resolution. Support the user to set the pixel size as a resolution parameter, such as 0.01°. Use GDAL's gdalwarp tool to resample. Provide a user interface to allow the user to select input files and output folders. Iterate through all files in the input folder and resample each file in turn. Also provide unified logging to report processing progress and error files.

[0207] After the data preprocessing is successful, the model management module is used to load the model file in the PyTorch.pt format, and the model's JSON configuration file (modelconfig.json) is used to describe the model's related data (such as name, input / output size, etc.). It also supports users to switch to different models through the selection list, and also supports users to upload a new version of the model file to replace the model file and update the version number in the configuration file.

[0208] After the model is loaded, downscaling begins. Use GDAL to read precipitation data and environmental factor data in GeoTIFF format, extract necessary metadata (such as resolution, spatial range, and projection), and check whether the precipitation data and environmental factor data are in the same geographic range. If not, automatically crop them to align them. Use Torch C++ (LibTorch) to load the .pth model file. Combine precipitation data and environmental factor data as model input for downscaling. Use the GDAL library to save output data and retain the original geographic information. Provide a user interface that allows users to select input files and output folders, traverse multiple precipitation files and environmental factor files in the folder, downscale all precipitation files, and generate high-resolution precipitation data.

[0209] After the downscaling process, the fusion correction process begins. The generated high-resolution precipitation data and the station observation data are unified into the same geographic coordinate system. For each station, the precipitation data of the corresponding date of the station is read through the date of the precipitation data, and the residual is calculated. The residual is interpolated into a complete raster map using the inverse distance weighting method. Then the correction begins, only the precipitation area is corrected, and negative precipitation is prevented, and finally the output corrected high-resolution precipitation data is generated.

[0210] The embodiment of the present application also provides an electronic device, which may include a processor, a memory, a receiver and a transmitter, wherein the processor is used to execute the precipitation data processing method based on multi-source data fusion mentioned in the above embodiment, wherein the processor and the memory may be connected via a bus or other means, such as by bus connection. The receiver may be connected to the processor and the memory via wired or wireless means.

[0211] The processor may be a central processing unit (CPU). The processor may 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 gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0212] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the precipitation data processing method based on multi-source data fusion in the embodiment of the present application. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, realizing the precipitation data processing method based on multi-source data fusion in the above method embodiment.

[0213] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0214] The one or more modules are stored in the memory, and when executed by the processor, perform the precipitation data processing method based on multi-source data fusion in the embodiment.

[0215] In some embodiments of the present application, the user equipment may include a processor, a memory, and a transceiver unit, which may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected through a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0216] As an implementation method, the functions of the receiver and the transmitter in the present application can be considered to be implemented through a transceiver circuit or a dedicated chip for transceiver, and the processor can be considered to be implemented through a dedicated processing chip, a processing circuit or a general chip.

[0217] As another implementation method, it is possible to use a general-purpose computer to implement the server provided in the embodiment of the present application, that is, to store the program code for implementing the functions of the processor, receiver, and transmitter in a memory, and the general-purpose processor implements the functions of the processor, receiver, and transmitter by executing the code in the memory.

[0218] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the precipitation data processing method based on multi-source data fusion are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0219] The embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the aforementioned precipitation data processing method based on multi-source data fusion.

[0220] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is specifically performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier.

[0221] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0222] In the present application, features described and / or illustrated for one embodiment may be used in the same manner or in a similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0223] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A precipitation data processing method based on multi-source data fusion, characterized in that: include: Inputting the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement and space attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data, and outputs downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data; The downscaled image data is corrected according to the station precipitation observation data corresponding to each of the multiple observation stations in the target area at the target time interval to obtain the target precipitation observation data of the target area at the target time interval.

2. The precipitation data processing method based on multi-source data fusion according to claim 1 is characterized in that: The step of correcting the downscaled image data according to the precipitation observation data of each observation station in the target area corresponding to the target time interval to obtain the target precipitation observation data of the target area in the target time interval includes: Determine, according to the downscaled image data, the residuals corresponding to the precipitation observation data of each observation station in the target area corresponding to the target time interval; Based on the inverse distance weighted average interpolation method, the residuals corresponding to the precipitation observation data of each of the stations are interpolated to obtain a corresponding residual grid map; Based on the residual grid map, residual correction is performed on the downscaled image data with preset correction constraints to obtain target precipitation observation data of the target area at a target time interval; wherein the correction constraints include: performing residual correction only on points where precipitation exists in the downscaled image data, and if, after the residual correction is performed on the downscaled image data, there is a point in the downscaled image data with a precipitation value less than 0, setting the precipitation value of the point to 0.

3. The precipitation data processing method based on multi-source data fusion according to claim 1 is characterized in that: The precipitation data downscaling model includes: a generator constructed based on a channel attention mechanism and a spatial attention mechanism, wherein the generator and a discriminator form a generative adversarial network; Correspondingly, before inputting the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and latitude and longitude data corresponding to the target area at the target time interval into the preset precipitation data downscaling model, it also includes: Obtain the pre-processed multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to different regions at each historical time interval; Constructing data samples, wherein each of the data samples includes pre-processed multi-satellite fusion remote sensing precipitation observation data, terrain elevation data, and longitude and latitude data corresponding to a region in a historical time interval; Using each of the data samples to perform at least one round of iterative training on the generative adversarial network, and in each round of training, inputting the downscaled image data output by the generator into the discriminator and optimizing the generator according to the determination result output by the discriminator; The generator in the generative adversarial network that has completed iterative training is determined as a precipitation data downscaling model.

4. The precipitation data processing method based on multi-source data fusion according to claim 3 is characterized in that: The generator includes a multimodal input processing module, an attention enhancement module, a residual module and an upsampling reconstruction module connected in sequence; the attention enhancement module includes a channel attention enhancement unit and a spatial attention enhancement unit connected in series; The multimodal input processing module is used to perform feature fusion processing on the multi-satellite fused remote sensing precipitation observation data, terrain elevation data and latitude and longitude data corresponding to the target area at the target time interval to obtain a corresponding fusion feature map; The channel attention enhancement unit is used to generate a channel attention enhancement feature map corresponding to the fusion feature map based on the channel attention mechanism; The spatial attention enhancement unit is used to generate the spatial and channel attention enhancement feature maps corresponding to the channel attention enhancement feature map based on the spatial attention mechanism; The residual module is used to perform gradient stabilization processing on the spatial and channel attention enhancement feature maps; The upsampling reconstruction module is used to upsample the spatial and channel attention enhancement feature maps after gradient stabilization processing to obtain downscaled image data with a resolution higher than that of the multi-satellite fusion remote sensing precipitation observation data.

5. The precipitation data processing method based on multi-source data fusion according to claim 4 is characterized in that: The multimodal input processing module includes: a precipitation image feature extraction unit, a terrain feature extraction unit and a feature splicing unit; The precipitation image feature extraction unit is used to perform channel dimension-upgrading feature extraction on the multi-satellite fusion remote sensing precipitation observation data corresponding to the target area at the target time interval based on the first convolution layer and the PReLU activation function in sequence, so as to obtain precipitation observation image feature data corresponding to the multi-satellite fusion remote sensing precipitation observation data; The terrain feature extraction unit includes three downsampling layers connected in sequence, and the downsampling layers are used to perform channel dimension up-conversion feature extraction on the terrain elevation data corresponding to the target area at the target time interval and the longitude data and latitude data in the longitude and latitude data in sequence based on the second convolution layer and the PReLU activation function, so that the terrain feature extraction unit outputs the terrain elevation feature data corresponding to the terrain elevation data, the longitude feature data corresponding to the longitude data, and the latitude feature data corresponding to the latitude data respectively; The feature stitching layer is used to perform feature fusion on the precipitation observation image feature data, the terrain elevation feature data, the longitude feature data and the latitude feature data in the channel dimension to obtain a corresponding fusion feature map.

6. The precipitation data processing method based on multi-source data fusion according to claim 4 is characterized in that: The channel attention enhancement unit includes: a global average pooling layer, a global maximum pooling layer, a shared fully connected layer, a first Sigmoid activation layer and a first channel weight addition unit; The global average pooling layer is used to extract the global average feature map of the fused feature map; The global maximum pooling layer is used to extract the global maximum feature map of the fused feature map; The shared fully connected layer is used to perform nonlinear transformation on the global average feature map and the global maximum feature map based on the sequentially connected ReLU activation function and the channel compression unit to obtain a corresponding fully connected feature map; The first Sigmoid activation layer is used to generate a channel weight map corresponding to the fully connected feature map within a preset interval based on the Sigmoid function, wherein the preset interval is [0, 1]; The first channel weight addition unit is used to perform channel weight addition on the channel weight map and the fusion feature map to obtain a corresponding channel attention enhancement feature map.

7. The precipitation data processing method based on multi-source data fusion according to claim 4 is characterized in that: The spatial attention enhancement unit includes: an average pooling layer, a maximum pooling layer, a channel splicing layer, a third convolutional layer, a second Sigmoid activation layer and a second channel weight addition unit; The average pooling layer is used to extract the average response map of the channel attention enhanced feature map; The maximum pooling layer is used to extract the maximum response map of the channel attention enhancement feature map; The channel splicing layer is used to splice the average response map and the maximum response map along the channel dimension based on the ReLU activation function and the channel compression unit connected in sequence to obtain a corresponding feature splicing map; The third convolutional layer is used to learn the spatial weight distribution of the feature splicing graph to obtain the corresponding spatial feature graph; The second Sigmoid activation layer is used to generate a spatial weight map corresponding to the spatial feature map based on a Sigmoid function; The second channel weight addition unit is used to perform channel weight addition on the spatial weight map and the channel attention enhancement feature map to obtain the corresponding spatial and channel attention enhancement feature map.

8. A precipitation data processing device based on multi-source data fusion, characterized in that: include: A precipitation data downscaling module is used to input the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data corresponding to the target area at the target time interval into a preset precipitation data downscaling model, so that the precipitation data downscaling model performs feature fusion, channel attention enhancement and space attention enhancement processing on the multi-satellite fusion remote sensing precipitation observation data, terrain elevation data and longitude and latitude data, and outputs downscaled image data with a higher resolution than the multi-satellite fusion remote sensing precipitation observation data; The downscaled image correction module is used to correct the downscaled image data according to the station precipitation observation data corresponding to each of the multiple observation stations in the target area at the target time interval, so as to obtain the target precipitation observation data of the target area at the target time interval.

9. An electronic 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 the computer program, the precipitation data processing method based on multi-source data fusion as described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the precipitation data processing method based on multi-source data fusion as described in any one of claims 1 to 7 is implemented.

Images