Deep learning-based downscaling and fusion collaborative calculation method and system for daily precipitation products

Abstract

The present application provides a collaborative calculation method and system for downscaling and fusing daily precipitation products based on deep learning, belonging to the technical field of data processing. Based on remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources, daily-scale datasets and monthly-scale datasets are generated; a data spatio-temporal fusion and downscaling model is constructed, and the model is used to train and predict the daily-scale dataset and the monthly-scale dataset respectively to obtain daily-scale precipitation estimates and monthly-scale precipitation estimates; on the monthly scale, the daily-scale precipitation estimates are corrected using the monthly-scale precipitation estimates. This solution performs downscaling on the monthly scale and the daily scale respectively. Vegetation information data is added to the monthly-scale downscaling to calculate accurate monthly precipitation, and the daily-scale downscaling is used to estimate the proportion of daily precipitation in the total monthly precipitation. Multiplying the two obtains the final daily precipitation, improving the accuracy of daily-scale precipitation estimation.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly relates to a collaborative calculation method, system and electronic device for downscaling and fusing daily precipitation products based on deep learning. Background Technique

[0002] High-resolution daily precipitation data is crucial for understanding regional and global precipitation patterns, simulating future climate scenarios, and assessing the ecological impacts of climate change. However, due to the strong spatio-temporal heterogeneity of daily precipitation, obtaining high-resolution daily precipitation data still faces significant challenges. Traditional production of daily precipitation data relies on ground monitoring data from meteorological observation stations. Although these station data can provide accurate precipitation estimates, their spatial information is limited. Specifically, meteorological observation stations are very unevenly distributed in space, with sparse station densities in rural areas, mountainous areas, forests, etc., greatly limiting their reconstruction of spatio-temporal changes in precipitation patterns. The progress of remote sensing technology has provided an important information source for obtaining high-resolution and spatially continuous daily precipitation data.

[0003] In recent years, remote sensing precipitation products based on multiple satellite sensors have been successively released, such as Global Satellite Mapping of Precipitation (GSMaP), Integrated Multi-satellite Retrievals for GPM (IMERG), Climate Hazards center InfraRed Precipitation with Station data (CHIRPS), and CPC MORPHing technique (CMORPH). According to previous evaluations, GSMaP, IMERG, and CMORPH have shown good daily precipitation accuracy in China. However, the resolution and accuracy of existing precipitation products are still insufficient to meet more refined regional scales, especially for meteorological and hydrological research in small regions.

[0004] Statistical downscaling methods combined with high-resolution environmental variables have been widely used in daily precipitation data production. Since precipitation is considered to be closely related to various environmental factors such as topography and vegetation, many methods for downscaling monthly-scale and annual-scale precipitation products using high-resolution environmental data have been developed. For example, geographically weighted regression (GWR) was introduced into the downscaling of monthly-scale TRMM data, taking into account the spatial heterogeneity of the relationship between precipitation and the normalized difference vegetation index (NDVI) as well as the digital elevation model (DEM). Studies have shown that the downscaling accuracy of support vector machines (SVMs) is not only better than methods including random forests (RFs), GWR, multiple linear regression (MLR), and exponential regression (ER), but also has good consistency with monthly ground observation data compared to the original TRMM data. Currently, these related studies mainly focus on precipitation data at the annual and monthly scales because there is a significant correlation between precipitation and local environmental factors at these time scales. However, at finer time scales (such as daily), this relationship usually changes significantly or even does not exist.

[0005] To address this issue, some new methods have been introduced. The common strategy is to first accumulate precipitation estimates to the monthly or annual scale and establish their relationship with local environmental variables. After spatial downscaling at the coarse scale, temporal downscaling of daily precipitation estimates is then performed. This strategy assumes that the proportion of daily precipitation in monthly precipitation can be accurately detected at the original spatial resolution. However, due to the limited ability of precipitation products in daily precipitation, especially for extreme precipitation, this assumption is usually difficult to achieve. Missed or false alarms of extreme precipitation in precipitation products may lead to significant biases in daily precipitation estimates for the entire month.

[0006] Therefore, an improved technical solution is needed to address the deficiencies of the above-mentioned existing technologies. Summary of the Invention

[0007] The purpose of this application is to provide a deep learning-based collaborative calculation method, system, and electronic device for downscaling and fusing daily precipitation products, to solve or alleviate the problem of insufficient accuracy in existing daily-scale precipitation estimates, and to produce daily precipitation products with higher resolution and accuracy.

[0008] To achieve the above purpose, this application provides the following technical solutions:

[0009] In the first aspect, this application provides a deep learning-based collaborative calculation method for downscaling and fusing daily precipitation products, including:

[0010] Generating a daily-scale dataset and a monthly-scale dataset based on remote sensing precipitation products, meteorological observation station data, vegetation, and terrain data from multiple sources; wherein, the monthly-scale dataset contains vegetation information, and the daily-scale dataset does not contain vegetation information;

[0011] Build a data spatio-temporal fusion and downscaling model based on deep learning, and use the daily-scale dataset and the monthly-scale dataset to train the data spatio-temporal fusion and downscaling model respectively to obtain a trained model, and use the trained model to make predictions respectively to obtain daily-scale precipitation estimation and monthly-scale precipitation estimation;

[0012] Among them, the spatial resolution of the daily-scale precipitation estimation and the monthly-scale precipitation estimation is the target resolution, and the target resolution is higher than the original spatial resolution of the remote sensing precipitation product;

[0013] The data spatio-temporal fusion and downscaling model includes: a Residual Network (ResNet) module, a Bidirectional Long Short-Term Memory (BiLSTM) module, and an attention mechanism module; the ResNet module is used to learn the spatial neighborhood features of the input data at each time step and input them into the BiLSTM module; the BiLSTM module is used to capture the temporal context relationship of the spatial neighborhood features, and the attention mechanism module is used to weight the output of the BiLSTM module to generate the final model output;

[0014] On the monthly scale, use the monthly-scale precipitation estimation to correct the daily-scale precipitation estimation to obtain high-resolution daily precipitation data.

[0015] In some possible implementation manners, based on the remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources, generate a daily-scale dataset and a monthly-scale dataset, including:

[0016] Perform spatial interpolation processing on the remote sensing precipitation products from multiple sources to obtain spatially interpolated precipitation data;

[0017] Perform resampling processing on the vegetation and terrain data to obtain resampled vegetation and terrain data;

[0018] Among them, the spatial resolution of the spatially interpolated precipitation data and the resampled vegetation and terrain data is the target resolution;

[0019] Perform time-scale accumulation processing on the interpolated precipitation data to obtain daily-scale precipitation data and monthly-scale precipitation data;

[0020] Based on the daily-scale precipitation data and the monthly-scale precipitation data, with each observation station in the meteorological observation station data as the center, extract the daily-scale precipitation and monthly-scale precipitation of N×N pixels around it, and extract the vegetation information and terrain information of N×N pixels around it from the resampled vegetation and terrain data;

[0021] Combine the daily-scale precipitation of N×N pixels with the terrain information to generate the daily-scale dataset; combine the monthly-scale precipitation of N×N pixels with the vegetation information and the terrain information to generate the monthly-scale dataset;

[0022] Where N is a positive integer greater than 1.

[0023] In some possible implementation manners, the terrain data is DEM data. Accordingly, with each observation site in the meteorological observation station data as the center, extract the terrain information of N×N pixels around it, specifically:

[0024] With each observation site in the meteorological observation station data as the center, use the DEM data to extract the elevation, slope, and aspect of N×N pixels around the center point.

[0025] In some possible implementation manners, the vegetation information includes: the Normalized Difference Vegetation Index (NDVI).

[0026] In some possible implementation manners, the ResNet module includes multiple residual blocks, and the multiple residual blocks are used to continuously compress the input data in the spatial dimension for each time step and learn its spatial neighborhood features;

[0027] The BiLSTM module includes multiple BiLSTM layers connected in sequence, and is used to receive the spatial neighborhood features learned by the ResNet module and process them in sequence to obtain the output, hidden layer state, and cell state of the corresponding layer;

[0028] The attention mechanism module is used to multiply the output of the final layer of the BiLSTM module by the final hidden layer state and normalize it to obtain attention weights, and apply the attention weights to the output of the final layer to generate the final model output.

[0029] In some possible implementation manners, the data spatio-temporal fusion and downscaling model further includes: a fully connected layer and a Dropout layer connected in sequence with the attention mechanism module.

[0030] In some possible implementation manners, after generating the daily-scale dataset and the monthly-scale dataset, it further includes: dividing the daily-scale dataset and the monthly-scale dataset to obtain corresponding training data, validation data, and test data;

[0031] Accordingly, training the data spatio-temporal fusion and downscaling model includes:

[0032] Input the training data into the data spatio-temporal fusion and downscaling model, and repeatedly adjust the hyperparameters according to the accuracy evaluation index value of the data spatio-temporal fusion and downscaling model on the validation data until the accuracy index value reaches the preset accuracy condition, and output the trained data spatio-temporal fusion and downscaling model;

[0033] Use the trained data spatio-temporal fusion and downscaling model to predict the test data, and measure the unbiased accuracy of the trained data spatio-temporal fusion and downscaling model according to the accuracy evaluation index of the prediction result.

[0034] In some possible implementation manners, on the monthly scale, use the estimated monthly-scale precipitation to correct the estimated daily-scale precipitation to obtain high-resolution daily precipitation data. Specifically:

[0035] For any specified month, perform normalization processing on the time series data composed of the estimated daily-scale precipitation values of the specified month to obtain the normalization processing result;

[0036] Multiply the normalization processing result by the estimated monthly-scale precipitation value corresponding to the specified month to obtain high-resolution daily precipitation data.

[0037] In a second aspect, the present embodiment provides a deep learning-based collaborative computing system for downscaling and fusing daily precipitation products, including:

[0038] A data acquisition unit configured to generate a daily-scale data set and a monthly-scale data set based on remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources; wherein, the monthly-scale data set contains vegetation information, and the daily-scale data set does not contain vegetation information;

[0039] A model construction and prediction unit configured to use a pre-constructed data spatio-temporal fusion and downscaling model to process the daily-scale data set and the monthly-scale data set respectively, and correspondingly obtain an estimated daily-scale precipitation and an estimated monthly-scale precipitation;

[0040] Wherein, the spatial resolution of the estimated daily-scale precipitation and the estimated monthly-scale precipitation is the target resolution, and the target resolution is higher than the original spatial resolution of the remote sensing precipitation product;

[0041] The data spatio-temporal fusion and downscaling model includes: a Residual Network (ResNet) module, a Bidirectional Long Short-Term Memory (BiLSTM) module, and an attention mechanism module. The ResNet module is used to learn the spatial neighborhood features of the input data at each time step and input them into the BiLSTM module. The BiLSTM module is used to capture the temporal context relationship of the spatial neighborhood features, and the attention mechanism module is used to weight the output of the BiLSTM module to generate the daily-scale precipitation estimate and the monthly-scale precipitation estimate.

[0042] A correction unit, configured to correct the daily-scale precipitation estimate on a monthly basis using the monthly-scale precipitation estimate to obtain high-resolution daily precipitation data.

[0043] In a third aspect, this embodiment provides an electronic device, including: a memory for storing instructions executed by one or more processors of the electronic device; a processor, when the processor executes the instructions in the memory, enables the electronic device to implement the steps of the method described in any of the above embodiments.

[0044] The method for collaborative calculation of downscaling and fusion of daily precipitation products based on deep learning provided in the embodiments of this application first fuses and downscales the daily-scale and monthly-scale precipitation, that is, directly models the daily-scale and monthly-scale precipitation data to obtain their respective precipitation estimates, laying a foundation for subsequent bias correction. A data spatio-temporal fusion and downscaling model combining a Residual Network (ResNet), a Bidirectional Long Short-Term Memory (BiLSTM), and an attention mechanism (abbreviated as the RNBLA model) is used. This model considers the relationship between precipitation and vegetation and terrain, and integrates data from multiple remote sensing precipitation products and meteorological observation stations to improve the accuracy of precipitation estimation. Since the correlation between local environmental factors and daily precipitation is not significant at the daily scale, the daily-scale precipitation estimate directly generated by the model is biased. However, there is a relatively stable relationship between monthly-scale precipitation and environmental factors, so the generated monthly-scale precipitation estimate has higher accuracy. Therefore, the generated monthly-scale precipitation estimate is used to correct the bias of the daily-scale precipitation on a monthly basis, avoiding errors that may be caused by missed or false alarms of extreme precipitation under traditional downscaling methods, thereby obtaining a daily-scale precipitation data product with higher resolution and higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic flowchart of the method for collaborative calculation of downscaling and fusion of daily precipitation products based on deep learning.

[0046] Figure 2Logic diagram of a deep learning-based daily precipitation product downscaling and fusion collaborative calculation method provided according to some embodiments of the present application.

[0047] Figure 3 Schematic diagram of the single model input area range for the data spatio-temporal fusion and downscaling model.

[0048] Figure 4 Schematic diagram of the overall structure of the data spatio-temporal fusion and downscaling model.

[0049] Figure 5 Schematic diagram of the specific structure of the data spatio-temporal fusion and downscaling model.

[0050] Figure 6 Schematic diagram of the structure of an electronic device provided for some embodiments. Detailed implementation manners

[0051] For the convenience of understanding the technical solutions of the present application, the following provides an exemplary description of relevant terms.

[0052] (1) Remote sensing precipitation product:

[0053] Remote sensing precipitation products refer to a set of data on the precipitation distribution and intensity on the Earth's surface collected, processed, and retrieved through remote sensing sensors carried on satellites. Such products use multi-band observations of the atmosphere and the Earth's surface by satellites (such as visible light, infrared, microwave, etc.), combined with numerical models and algorithms, to generate spatio-temporally continuous precipitation information.

[0054] As described in the background art, currently, remote sensing precipitation products covering the global range include: GSMaP, IMERG, CHIRPS, and CMORPH. Among them, the Chinese name of GSMaP is Satellite Precipitation Mapping, its spatial resolution is 0.1° (about 11 km), and the time resolution is half an hour / hour; the Chinese name of IMERG is Global Precipitation Measurement Multi-Satellite Integrated Retrieval, and its spatial resolution and time resolution are the same as those of GSMaP, which are 0.1° and half an hour / hour respectively; the Chinese name of CHIRPS is Climate Hazards Center Infrared Precipitation with Station Data Fusion Product, its spatial resolution is 0.05° (about 5.5 km), and the time resolution is daily / monthly; the Chinese name of CMORPH is CPC Morphing Technique Precipitation Product, its spatial resolution is 8 km, and the time resolution is half an hour / hour. It can be seen that compared with traditional ground monitoring data relying on meteorological observation stations, each remote sensing precipitation product can provide precipitation data with wide coverage and spatial continuity. However, the highest spatial resolution of these remote sensing precipitation products is only 5.5 km, while local-scale research usually requires a higher spatial resolution (such as a spatial resolution of 0.01° (about 1.1 km) or higher). Therefore, the existing remote sensing precipitation products are difficult to meet the needs of local-scale research.

[0055] (2) Original spatial resolution, target spatial resolution:

[0056] The original spatial resolution is the spatial resolution of the remote sensing precipitation product, and the target spatial resolution is the spatial resolution of the precipitation estimate obtained after downscaling. During the downscaling process, the original spatial resolution usually refers to a low spatial resolution (such as 0.1°), and the target spatial resolution is a high spatial resolution (such as 0.01°).

[0057] (3) Statistical downscaling method:

[0058] The statistical downscaling method is a technique that uses mathematical and statistical models to convert low-resolution precipitation data (such as the 0.1° spatial resolution of remote sensing precipitation products) into high-resolution (such as 0.01° or higher) precipitation data. Its core idea is to utilize the statistical relationship between environmental variables (such as terrain, vegetation index, etc.) and precipitation to spatially refine the low-resolution precipitation data, thereby generating a more refined and higher-resolution precipitation distribution.

[0059] However, the prerequisite for implementing the statistical downscaling method is that there is a significant correlation between precipitation and local environmental factors. Such a prerequisite faces limitations in the time scale. Specifically, on the annual and monthly scales, there is a significant correlation between precipitation and local environmental factors, while on the daily scale, the correlation between precipitation and local environmental factors decreases or even disappears.

[0060] For this reason, the existing technology obtains high-precision daily-scale precipitation data through a data accumulation strategy, that is, first accumulates precipitation estimates to the monthly or annual scale, establishes its relationship with local environmental variables, and adopts a downscaling method that first deals with space and then time to obtain daily-scale precipitation data. However, during extreme precipitation events, such a data accumulation strategy often leads to large biases in daily precipitation estimates.

[0061] In view of this, the method provided in this application proposes a novel daily precipitation downscaling strategy. By using a more advanced deep learning module (a combination of a residual network (ResNet), a bidirectional long short-term memory network BiLSTM, and an attention mechanism (Attention) to form a deep neural network, which constitutes a data spatio-temporal fusion and downscaling model (abbreviation: model)), first uses the model to estimate precipitation on the monthly and daily scales respectively, and then uses the generated monthly-scale precipitation estimate to correct the bias of the daily-scale precipitation, so as to reduce the errors caused by extreme precipitation underreporting or false reporting in remote sensing precipitation products, thereby producing a daily precipitation product with higher resolution and better accuracy.

[0062] The following describes the embodiments of this application with reference to the accompanying drawings.

[0063] The embodiments of this application can be applied to Figure 6In the electronic device shown, the electronic device may be, but is not limited to, mobile terminals such as mobile phones, tablet computers, handheld computers, personal digital assistants (PDAs), etc., smart home devices such as smart TVs, smart cameras, wearable devices such as smart bracelets, smart watches, smart glasses, or other computer devices such as desktop computers, laptop computers, ultra-mobile personal computers (UMPCs), netbooks, and smart screens.

[0064] As Figure 6 As shown, the electronic device 200 may include one or more of the following components: a processor 201, a memory 203, a communication interface 202, and a communication bus 204. Among them, the memory 203 may be connected to the processor 201 through the bus 204. The bus may transfer data between the processor 201 and the memory 203. The bus may be divided into an address bus, a data bus, a control bus, etc.

[0065] The processor 201 may include one or more processing cores. The processor 201 can connect various parts within the entire electronic device 200 using various interfaces and circuits. By running or executing instructions, programs, code sets, or instruction sets stored in the memory 203, and by invoking the data stored in the memory 203, it can perform various functions of the electronic device 200 and process data. Exemplarily, the processor 201 may include an application processor (AP), a modem processor, a CPU, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic array (PLA), and / or a neural-network processing unit (NPU), etc. Among them, the CPU mainly processes the operating system, user interface, application programs, etc.; the GPU is responsible for rendering and drawing the content to be displayed; the NPU is used to implement artificial intelligence (AI) functions; the modem is used to handle wireless communications. Different processing units can be independent devices or integrated in one or more processors. For example, the multiple processing units shown above are all integrated in one SoC, or the AP is a separate semiconductor chip and the other processing units are integrated in one SoC. This application does not make any limitations in this regard.

[0066] The memory 203 may include a random access memory (RAM), may also include a read-only memory (ROM), and may further include a non-transitory computer-readable storage medium. The memory 203 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 203 may include a program storage area and a data storage area. Among them, the program storage area may store instructions for implementing the operating system, instructions for at least one function, such as the collaborative calculation method for downscaling and fusing daily precipitation products based on deep learning, etc.; the data storage area may store data collected or generated according to the use of the electronic device 200, such as input data of remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources, or generated daily-scale data sets, monthly-scale data sets, daily-scale precipitation estimates, monthly-scale precipitation estimates, etc.

[0067] In addition, those skilled in the art can understand that the structure of the electronic device 200 shown in the above figures does not limit the electronic device 200. The electronic device may include more or fewer components than shown in the figures, or combine some components, or have different component arrangements. For example, the electronic device 200 also includes components such as a microphone, a speaker, a radio frequency circuit, a sensor, an audio circuit, a power supply, and a Bluetooth module, which will not be elaborated here.

[0068] This embodiment provides a collaborative calculation method for downscaling and fusing daily precipitation products based on deep learning, such as Figure 1 shown, the method includes the following steps:

[0069] Step S101, generate a daily-scale data set and a monthly-scale data set based on remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources.

[0070] In this embodiment, for data set preparation, remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources need to be collected separately.

[0071] Among them, the remote sensing precipitation products include the mainstream GSMaP, IMERG and CMORPH. At the same time, the terrain data, vegetation data and meteorological observation station data of the study area are obtained. The above data can be downloaded from their respective official websites or other data providers. Among them, the GSMaP, IMERG and CMORPH data are specifically: precipitation data with a half-hour / hour time resolution from July 2000 to June 2022. The spatial resolutions of GSMaP, IMERG and CMORPH refer to the previous text and will not be elaborated here.

[0072] Terrain data includes: DEM data, and vegetation information includes: Normalized Difference Vegetation Index (NDVI).

[0073] In some embodiments, in step S101, based on remote sensing precipitation products, meteorological observation station data, vegetation, and terrain data from multiple sources, daily-scale datasets and monthly-scale datasets are generated, including:

[0074] In step S111, spatial interpolation processing is performed on precipitation products from multiple sources to obtain spatially interpolated precipitation data.

[0075] In step S121, resampling processing is performed on vegetation and terrain data to obtain resampled vegetation and terrain data.

[0076] Among them, the spatial resolutions of the spatially interpolated precipitation data, resampled vegetation, and terrain data are all the target spatial resolution.

[0077] The purposes of the above steps S111 and S121 are to preprocess the collected data. Exemplarily:

[0078] For remote sensing precipitation products: The bilinear interpolation method is used to interpolate the spatial resolution of all remote sensing precipitation products to the target spatial resolution, such as 0.01°.

[0079] In terms of environmental variables, NDVI and DEM are resampled to the target spatial resolution, such as 0.01°, and the longitude and latitude of the center of each grid are calculated.

[0080] In step S131, time-scale accumulation processing is performed on the interpolated precipitation data to obtain daily-scale precipitation data and monthly-scale precipitation data.

[0081] That is to say, after spatially interpolating the remote sensing precipitation products, in terms of the time scale, the estimated precipitation values from 12:00 noon of the previous day to 12:00 noon today (UTC) in the remote sensing precipitation products are accumulated to obtain the accumulated daily precipitation estimated value (i.e., the daily-scale precipitation data). The daily precipitation estimated value is accumulated on a monthly scale to obtain the monthly-scale precipitation data. Such processing is beneficial for the precipitation data to match the daily precipitation observation values of Chinese meteorological observation stations.

[0082] In step S141, based on the daily-scale precipitation data and monthly-scale precipitation data, with each observation station in the meteorological observation station data as the center, the daily-scale precipitation and monthly-scale precipitation of N×N pixels around it are extracted, and the vegetation information and terrain information of N×N pixels around it are extracted from the resampled vegetation and terrain data. Among them, N is a positive integer greater than 1.

[0083] Here, the purpose of step S141 is to incorporate the information of the adjacent areas of the meteorological observation stations into the model prediction. Specifically: taking each observation site in the meteorological observation station data as the center, the daily-scale precipitation and monthly-scale precipitation of the N×N pixels around it are extracted as the single model input area.

[0084] The size of the sub-grid (i.e., the input area) that meets the model input requirements is a trade-off between the computational cost and the accuracy. After testing and verification, 49×49 pixels can not only enable the model to achieve the expected accuracy but also maintain the computational efficiency. Therefore, as a preferred embodiment, the value of N is 49, that is, 49×49 (0.49°×0.49°) pixels around each observation site are extracted as the single model input area.

[0085] Figure 3 The schematic diagram of the single model input area range of the data spatio-temporal fusion and downscaling model is shown. The left figure shows the result of superimposing the spatially interpolated and time-scale accumulated remote sensing precipitation products, the resampled vegetation and terrain data with the meteorological observation stations (black dots in the left figure). Taking each meteorological observation station (such as meteorological observation station i) as the center, 49×49 pixels (black grids in the left figure) around it are extracted to obtain the neighborhood range shown in the right figure, which is used as the single input area of the model.

[0086] In some embodiments, the terrain information specifically includes elevation, slope, and aspect. Therefore, in step S141, taking each observation site in the meteorological observation station data as the center, the terrain information of the N×N pixels around it is extracted. Specifically: taking each observation site in the meteorological observation station data as the center, the elevation, slope, and aspect of the N×N pixels around the center point are extracted using DEM data.

[0087] Step S151: Combine the daily-scale precipitation of the N×N pixels with the terrain information to generate a daily-scale data set; combine the monthly-scale precipitation of the N×N pixels with the vegetation information and terrain information to generate a monthly-scale data set.

[0088] It should be noted that since the time scale of the vegetation information (such as NDVI data) is usually monthly scale, it can only be directly applied to the fusion and downscaling processing of monthly-scale precipitation. Therefore, in step S151, the monthly-scale data set contains the combination of the vegetation information, monthly-scale precipitation, and terrain information of the N×N pixels, while the generated daily-scale data set only includes the combination of the daily-scale precipitation and terrain information of the N×N pixels.

[0089] Further, after generating the daily-scale dataset and the monthly-scale dataset, it further includes: dividing the daily-scale dataset and the monthly-scale dataset to obtain corresponding training data, validation data, and test data, laying a data foundation for subsequent model training and prediction.

[0090] Step S102: Construct a data spatio-temporal fusion and downscaling model based on deep learning, and separately train the data spatio-temporal fusion and downscaling model using the daily-scale dataset and the monthly-scale dataset to obtain a trained model, and use the trained model to perform separate predictions to correspondingly obtain daily-scale precipitation estimation and monthly-scale precipitation estimation.

[0091] Among them, the spatial resolutions of the daily-scale precipitation estimation and the monthly-scale precipitation estimation are the target spatial resolutions, and the target spatial resolution is higher than the original spatial resolution of the remote sensing precipitation product.

[0092] The data spatio-temporal fusion and downscaling model includes: a Residual Network (ResNet) module, a Bidirectional Long Short-Term Memory (BiLSTM) module, and an attention mechanism module; the ResNet module is used to learn the spatial neighborhood features of the input data at each time step and input them into the BiLSTM module (also known as the BiLSTM cell); the BiLSTM module is used to capture the temporal context relationship of the spatial neighborhood features, and the attention mechanism module is used to weight the output of the BiLSTM module to generate the final model output.

[0093] Specifically, the ResNet module includes multiple residual blocks, and the multiple residual blocks are used to continuously compress the input data at each time step in the spatial dimension and learn its spatial neighborhood features; the BiLSTM module includes multiple BiLSTM layers connected in sequence, which are used to receive the spatial neighborhood features learned by the ResNet module and process them in sequence to obtain the output, hidden layer state, and cell state of the corresponding layer; the attention mechanism module is used to multiply the output of the final layer of the BiLSTM module by the final hidden layer state and perform normalization to obtain the attention weight, and apply the attention weight to the output of the final layer to generate the final model output.

[0094] Further, the data spatio-temporal fusion and downscaling model further includes: a fully connected layer and a Dropout layer connected in sequence to the attention mechanism module.

[0095] In this embodiment, in terms of model design, in order to capture the spatial and temporal correlations between meteorological observation stations and satellite estimations, the RNBLA model uses the ResNet module to learn spatial neighborhood features and the BiLSTM module to capture temporal context relationships. In addition, in order to selectively emphasize the most relevant and informative parts of the input sequence to achieve better prediction results, an attention mechanism module is introduced after the BiLSTM module to achieve weighting.

[0096] Figure 4 It is a schematic diagram of the overall structure of the data spatio-temporal fusion and downscaling model. As Figure 4 shown, first, for the target time step T, data for a total of n time steps before and after it are collected. At each time step from T1 to Tn, the remote sensing precipitation product and environmental variables (terrain and vegetation) within the model input area are concatenated into a three-dimensional array. Secondly, these three-dimensional arrays are each passed through a convolution and a max pooling operation, and then continuously compressed in the spatial dimension through five residual blocks while being enhanced at the feature level. Next, the n resulting feature vectors are sequentially input into a BiLSTM module containing three BiLSTM layers, generating three outputs: the BiLSTM output, the hidden state, and the cell state. Then, by multiplying the BiLSTM output by the final hidden state and normalizing, the attention weights (Attention weights) are obtained and applied to the BiLSTM output to generate the final output (that is, the Attention weights are normalized and weighted summed to obtain the Attention output). Finally, several fully connected layers are used to capture the relationships and patterns between the outputs of the attention module and remap them to the desired output length. That is to say, in order to prevent the model from overfitting, fully connected layers and Dropout layers are used after the attention mechanism module.

[0097] Figure 5 It is a schematic diagram of the specific structure of the data spatio-temporal fusion and downscaling model. Refer to Figure 5 , the model combines a convolutional neural network (CNN), a residual block (Residual block), a bidirectional LSTM (BiLSTM), an attention mechanism (Attention), and a fully connected layer (Linear), and is a hybrid model. The complete process and related hyperparameters executed at each time step Ti from T1 to Tn are as follows: There are four different data sources in the input part of the model, namely the remote sensing precipitation product and environmental variables (NDVI, DEM), and location information. Each has a data dimension of 49×49 but different numbers of channels. The shape of the remote sensing precipitation product is (49, 49, 3), NDVI is single-channel, DEM has three channels, and the location information has two channels; then these data are concatenated in the channel dimension through a Concatenate operation to obtain the concatenated tensor; next is the Convolution2D layer, using a 5×5 convolutional kernel, outputting 16 channels, followed by ReLU activation and MaxPooling2D(2×2); subsequently, it is input into multiple residual blocks for processing. The output channels of the multiple residual blocks are 32, 64, 128, 256, and 512 in sequence. The structure of each residual block includes two 3×3 convolutional layers, and the output channels of each convolutional layer are h, corresponding to different numbers of channels, each convolutional layer is followed by BatchNorm and ReLU, and then the input and output are added together (i.e., residual connection); after multiple residual blocks is the Flatten layer, which flattens the feature map output by the last residual block through the Flatten operation, and at this time, the flattened feature map of each time step is obtained. The feature time series composed of the feature maps of all time steps enter the BiLSTM layer in sequence. There are three BiLSTM layers with 64, 256, and 512 hidden units respectively; next is the Attention layer, which performs an attention mechanism on the output of the BiLSTM to refine the weighting of the BiLSTM output. Specifically, it multiplies the BiLSTM output by the final hidden layer state and normalizes it to obtain the attention weights; then it is followed by multiple fully connected layers and Dropout layers. The fully connected layers include 256, 128, 64, 32, and 1 neurons. There is a ReLU activation between the fully connected layer and the Dropout layer, and the Dropout layer uses a ratio of 0.1; finally, a value is output, using the ReLU activation function, for the regression task, that is, predicting precipitation.

[0098] After the RNBLA model is constructed, step S102 includes:

[0099] Step S112: Input the training data into the data spatio-temporal fusion and downscaling model, and repeatedly adjust the hyperparameters according to the accuracy evaluation index value of the data spatio-temporal fusion and downscaling model on the validation data until the accuracy index value reaches the preset accuracy condition, and output the trained data spatio-temporal fusion and downscaling model.

[0100] Step S122: Use the trained data spatio-temporal fusion and downscaling model to predict the test data, and measure the unbiased accuracy of the trained data spatio-temporal fusion and downscaling model according to the accuracy evaluation index of the prediction result.

[0101] Based on the foregoing description, the time scale of the NDVI data is the monthly scale and can only be directly applied to the fusion and downscaling processing of monthly-scale precipitation. Therefore, the RNBLA model is used to perform data fusion and downscaling on daily-scale and monthly-scale precipitation respectively. When processing the daily-scale dataset, the NDVI data is not input, and the daily-scale precipitation estimate and monthly-scale precipitation estimate with a resolution of the target resolution (such as 0.01°) are obtained respectively.

[0102] Step S103: On the monthly scale, use the monthly-scale precipitation estimate to correct the daily-scale precipitation estimate to obtain high-resolution daily precipitation data.

[0103] In some embodiments, step S103: On the monthly scale, use the monthly-scale precipitation estimate to correct the daily-scale precipitation estimate to obtain high-resolution daily precipitation data, which specifically includes the following steps:

[0104] Step S113: For any specified month, perform normalization processing on the time series data composed of the daily-scale precipitation estimation values of the specified month to obtain the normalization result.

[0105] Step S123: Multiply the normalization result by the monthly-scale precipitation estimation value corresponding to the specified month to obtain high-resolution daily precipitation data.

[0106] Specifically, in Step S103, bias correction is performed on the daily-scale precipitation estimation at the monthly scale, and the correction process includes: normalizing the daily precipitation estimation sequence of a given month and multiplying it by the monthly precipitation estimation value of that month.

[0107] It should be noted that traditional normalization processing is usually used to eliminate the dimension or data differences, while the normalization processing in this application is to convert the daily precipitation estimation value into the proportion of the daily precipitation in the total monthly precipitation, which is used to reflect the relative contribution of the daily precipitation in the specified whole month. Then, multiply the normalization result by the monthly-scale precipitation estimation value corresponding to the specified month, use the high-precision monthly precipitation estimation value as a benchmark to correct the daily precipitation, utilize the stability of the monthly precipitation, and at the same time combine the time dynamic characteristics of the daily precipitation to avoid the problem of precipitation estimation deviation caused by extreme precipitation errors, and improve the consistency of the monthly precipitation and the daily precipitation.

[0108] Through the above steps, the RNBLA model successfully integrates three mainstream remote sensing precipitation products (GSMaP, IMERG, and CMORPH) and in-situ observation data, and generates high-resolution (spatial resolution of 0.01°) or higher-resolution daily precipitation distribution data in the study area (such as China). Compared with the original spatial resolution of the remote sensing precipitation products (such as the original spatial resolution of GSMaP is 0.1°, the original spatial resolution of IMERG is 0.1°, and the original spatial resolution of CMORPH is 8 km), the daily precipitation estimation calculated by the RNBLA model obtains richer spatial precipitation details than the original products, which is more suitable for regional fine research.

[0109] Measure the unbiased accuracy of the trained data spatio-temporal fusion and downscaling model according to the accuracy evaluation indexes of the prediction results. In this example, CC, MAE, RMSE, Bias, POD, FAR, and CSI are selected as the accuracy evaluation indexes. Through the verification and comparison with the ground observation stations, compared with the GSMaP, IMERG, and CMORPH products, the CC of the daily precipitation data obtained by the method provided in this embodiment is increased by 9.21%, the MAE is reduced by 16.02%, the RMSE is reduced by 18.14%, and the absolute value of Bias is reduced by 69.57%. The detailed results are shown in Table 1, and Table 1 is as follows:

[0110] Table 1 Comparison of the overall accuracy between the RNBLA model and different precipitation products

[0111]

[0112] The results show that: the method proposed in this embodiment successfully improves the estimation accuracy of daily precipitation data and enhances the simulation accuracy of the daily precipitation field. In addition, the comprehensive comparison of the three indicators of POD, FAR, and CSI shows that: this method can improve the prediction accuracy of precipitation by the data. The experiment also plotted the spatial distribution maps of each evaluation index. It can be analyzed that, spatially, the precipitation estimation of the RNBLA model generally obtains lower errors. The MAE at the station location is reduced by 16% - 32% compared with the original precipitation product, and the RMSE is reduced by 18% - 35%. The RNBLA model has a Bias closer to 0 in 78.33% of the stations, alleviating the overestimation or underestimation bias of the remote sensing precipitation product to a certain extent.

[0113] To further understand the above process, the following combines Figure 2 to reorganize the overall logic of this embodiment again. Figure 2 It is a logic diagram of the downscaling and fusion collaborative calculation method for daily precipitation products based on deep learning provided according to some embodiments of the present application. As Figure 2 shown, the method provided in this embodiment can be executed according to the following steps:

[0114] (a) Dataset preparation, including remote sensing precipitation products, DEM, NDVI, location, and meteorological observation station data. Through multi-source satellite precipitation data from different sources, precipitation information with different accuracies is provided; spatial coordinate features are provided through location information (latitude and longitude), vegetation information is provided through NDVI, and the digital elevation model DEM provides topographic features such as elevation, slope, and aspect. In short, by integrating multi-source remote sensing data and ground observation data, the spatio-temporal heterogeneity features are fused to improve the model's modeling ability for complex environments (such as terrain and vegetation).

[0115] (b) Data preprocessing, including interpolating the remote sensing precipitation products, accumulating the precipitation products from the half-hour / hour time scale to the daily and monthly scales, resampling the NDVI and DEM data, and extracting sub-grids that meet the model input requirements. Through data preprocessing, data missing in space or time is filled, the spatial resolution and time resolution of different data sources are unified, and the model's ability to capture spatial context information is enhanced through local range sub-grid extraction.

[0116] (c) Model training and hyperparameter tuning, including repeatedly adjusting hyperparameters according to the model's performance on the validation data until the optimal accuracy is achieved, resulting in a trained model. The RNBLA model adopts a hybrid architecture of ResNet + BiLSTM + Attention, combining the local feature extraction ability of CNN and the temporal modeling ability of BiLSTM to adapt to the spatio-temporal dynamics of precipitation data, and enhancing the model's sensitivity to key features through the attention mechanism.

[0117] (d) Precipitation evaluation and prediction, using the proposed RNBLA model to fuse and downscale remote sensing precipitation products, and using evaluation metrics on the test data to measure the unbiased accuracy of the model.

[0118] (e) Daily precipitation correction, using the monthly precipitation estimate to correct the daily precipitation estimate at the monthly scale.

[0119] In summary, although a large number of studies have confirmed a significant correlation between precipitation and vegetation, and using known vegetation information for precipitation downscaling can improve accuracy. However, the inventors believe that the correlation is only significant at the annual and monthly scales, and greatly weakens or even disappears at the daily scale. Currently, the vegetation NDVI index is only reliable at the monthly scale. These two reasons make it difficult for traditional daily precipitation methods to directly utilize vegetation information. The method provided in this embodiment performs downscaling simultaneously at the monthly and daily scales. In the monthly scale, vegetation information data is added to calculate accurate monthly precipitation. The daily scale downscaling does not consider vegetation information and is used to estimate the proportion of daily precipitation in the total monthly precipitation. Finally, the monthly precipitation is multiplied by the daily precipitation proportion to obtain the final daily precipitation. The advantages of this downscaling framework are as follows: 1) It cleverly uses the vegetation information at the monthly scale and incorporates it into the final daily scale precipitation estimate during the correction stage; 2) The original remote sensing precipitation product has obvious deficiencies in extreme precipitation, often resulting in large errors or even missed reports, which will completely change the proportion of daily precipitation in the total monthly precipitation. In this embodiment, a model estimate is first performed on the daily precipitation alone to improve the accuracy of the proportion of daily precipitation in the total monthly precipitation; 3) The monthly scale data is relatively stable and less affected by accidental factors, while the daily precipitation data is greatly affected by random factors and local effects. By correcting the data at the monthly scale, these random fluctuations can be smoothed and the random error of the data can be reduced. Therefore, the daily precipitation downscaling result calculated through this framework is superior to both the direct daily scale precipitation result and the result of time downscaling using the original value of the precipitation product after monthly scale precipitation.

[0120] Based on the same inventive concept, this embodiment provides a deep learning-based collaborative calculation system for daily precipitation product downscaling and fusion, which includes:

[0121] A data acquisition unit, configured to generate a daily-scale dataset and a monthly-scale dataset based on precipitation products, meteorological observation station data, vegetation, and terrain data from multiple sources;

[0122] A model construction and prediction unit, configured to use a pre-constructed data spatio-temporal fusion and downscaling model to process the daily-scale dataset and the monthly-scale dataset respectively, and correspondingly obtain a daily-scale precipitation estimate and a monthly-scale precipitation estimate;

[0123] Among them, the spatial resolutions of the daily-scale precipitation estimate and the monthly-scale precipitation estimate are the target spatial resolutions, and the target spatial resolution is higher than the original spatial resolution of the precipitation product;

[0124] The data spatio-temporal fusion and downscaling model includes: a Residual Network (ResNet) module, a Bidirectional Long Short-Term Memory (BiLSTM) module, and an attention mechanism module; the ResNet module is used to learn the spatial neighborhood features of the input data at each time step and input them into the BiLSTM module; the BiLSTM module is used to capture the temporal context relationship of the spatial neighborhood features, and the attention mechanism module is used to weight the output of the BiLSTM module to generate a daily-scale precipitation estimate and a monthly-scale precipitation estimate;

[0125] A correction unit, configured to correct the daily-scale precipitation estimate monthly using the monthly-scale precipitation estimate to obtain high-resolution daily precipitation data.

[0126] The deep learning-based daily precipitation product downscaling and fusion collaborative computing system provided in this embodiment can implement the processes and steps of the deep learning-based daily precipitation product downscaling and fusion collaborative computing method provided in any of the above embodiments and achieve the same technical effects, which will not be elaborated here one by one.

[0127] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A method for downscaling and fusion collaborative calculation of daily precipitation products based on deep learning, characterized in that: include: Generate a daily-scale dataset and a monthly-scale dataset based on remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources; wherein the monthly-scale dataset contains vegetation information, and the daily-scale dataset does not contain vegetation information; Based on deep learning, a data spatiotemporal fusion and downscaling model is constructed, and the data spatiotemporal fusion and downscaling model is trained separately using the daily scale data set and the monthly scale data set to obtain a trained model, and the trained model is used to perform predictions respectively to obtain a daily scale precipitation estimate and a monthly scale precipitation estimate; The spatial resolution of the daily-scale precipitation estimation and the monthly-scale precipitation estimation is the target resolution, and the target resolution is higher than the original spatial resolution of the remote sensing precipitation product. The data spatiotemporal fusion and downscaling model includes: a residual network ResNet module, a bidirectional long short-term memory network BiLSTM module and an attention mechanism module; the ResNet module is used to learn the spatial neighborhood features of the input data of each time step and input it to the BiLSTM module; the BiLSTM module is used to capture the temporal contextual relationship of the spatial neighborhood features, and the attention mechanism module is used to weight the output of the BiLSTM module to generate a final model output; On the monthly scale, the daily precipitation estimate is corrected using the monthly precipitation estimate to obtain high-resolution daily precipitation data, specifically including: for any specified month, normalizing the time series data consisting of the daily precipitation estimate of the specified month to obtain a normalized result; and multiplying the normalized result by the monthly precipitation estimate corresponding to the specified month to obtain high-resolution daily precipitation data.

2. The method according to claim 1, characterized in that Based on the remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from various sources, daily and monthly scale datasets are generated, including: Performing spatial interpolation processing on the remote sensing precipitation products from the multiple sources to obtain spatially interpolated precipitation data; Resampling the vegetation and terrain data to obtain resampled vegetation and terrain data; The spatial resolutions of the precipitation data after spatial interpolation and the vegetation and terrain data after resampling are both the target resolution; Perform time scale accumulation processing on the interpolated precipitation data to obtain daily and monthly precipitation data; Based on the daily precipitation data and the monthly precipitation data, taking each observation station in the meteorological observation station data as the center, extracting the daily precipitation and the monthly precipitation of the surrounding N×N pixels, and extracting the vegetation information and the terrain information of the surrounding N×N pixels from the resampled vegetation and terrain data; Combining the daily-scale precipitation of N×N pixels with the terrain information to generate the daily-scale dataset; combining the monthly-scale precipitation of N×N pixels with the vegetation information and the terrain information to generate the monthly-scale dataset; Wherein, N is a positive integer greater than 1.

3. The method according to claim 2, characterized in that The terrain data is DEM data. Accordingly, taking each observation station in the meteorological observation station data as the center, the terrain information of the surrounding N×N pixels is extracted, specifically: Taking each observation station in the meteorological observation station data as the center, the DEM data is used to extract the elevation, slope and slope direction of N×N pixels around the center point.

4. The method according to claim 2, characterized in that: The vegetation information includes: Normalized Difference Vegetation Index NDVI.

5. The method according to claim 1, characterized in that The ResNet module includes a plurality of residual blocks, and the plurality of residual blocks are used to continuously compress the input data of each time step in the spatial dimension and learn its spatial neighborhood features; The BiLSTM module includes a plurality of BiLSTM layers connected in sequence, which are used to receive the spatial neighborhood features learned by the ResNet module and process them in sequence to obtain the output, hidden layer state and cell state of the corresponding layer; The attention mechanism module is used to multiply the output of the final layer of the BiLSTM module by the final hidden layer state and normalize it to obtain an attention weight, and apply the attention weight to the output of the final layer to generate the final model output.

6. The method according to claim 5, characterized in that The data spatiotemporal fusion and downscaling model also includes: a fully connected layer and a Dropout layer connected to the attention mechanism module in sequence.

7. The method according to claim 1, characterized in that After generating the daily and monthly scale datasets, it also includes: Dividing the daily scale data set and the monthly scale data set to obtain corresponding training data, verification data, and test data; Accordingly, the data spatiotemporal fusion and downscaling model is trained, including: Input the training data into the data spatiotemporal fusion and downscaling model, and repeatedly adjust the hyperparameters according to the accuracy evaluation index value of the data spatiotemporal fusion and downscaling model on the verification data until the accuracy index value reaches the preset accuracy condition, and output the trained data spatiotemporal fusion and downscaling model; The trained data spatiotemporal fusion and downscaling model is used to predict the test data, and the unbiased accuracy of the trained data spatiotemporal fusion and downscaling model is measured according to the accuracy evaluation index of the prediction result.

8. A daily precipitation product downscaling and fusion collaborative computing system based on deep learning, characterized in that: include: A data acquisition unit is configured to generate a daily scale data set and a monthly scale data set based on remote sensing precipitation products, meteorological observation station data, vegetation and terrain data from multiple sources; wherein the monthly scale data set contains vegetation information, and the daily scale data set does not contain vegetation information; A model building and prediction unit is configured to process the daily scale data set and the monthly scale data set respectively using a pre-built data spatiotemporal fusion and downscaling model to obtain a daily scale precipitation estimate and a monthly scale precipitation estimate respectively; The spatial resolution of the daily-scale precipitation estimation and the monthly-scale precipitation estimation is the target resolution, and the target resolution is higher than the original spatial resolution of the remote sensing precipitation product. The data spatiotemporal fusion and downscaling model includes: a residual network ResNet module, a bidirectional long short-term memory network BiLSTM module and an attention mechanism module; the ResNet module is used to learn the spatial neighborhood features of the input data of each time step and input it into the BiLSTM module; the BiLSTM module is used to capture the temporal contextual relationship of the spatial neighborhood features, and the attention mechanism module is used to weight the output of the BiLSTM module to generate the daily scale precipitation estimate and the monthly scale precipitation estimate; a correction unit configured to correct the daily-scale precipitation estimate at a monthly level using the monthly-scale precipitation estimate to obtain high-resolution daily precipitation data; The correction unit is further configured to: for any specified month, normalize the time series data consisting of the daily precipitation estimate of the specified month to obtain a normalized processing result; multiply the normalized processing result by the monthly precipitation estimate corresponding to the specified month to obtain high-resolution daily precipitation data.

9. An electronic device, characterized in that: include: a memory for storing instructions executed by one or more processors of the electronic device; The processor, when the processor executes the instructions in the memory, can enable the electronic device to implement the steps of any one of the methods described in claims 1 to 7.

Images