Method for establishing precipitation estimation model by integrating multi-source data and precipitation estimation method

Abstract

The present invention provides a method for establishing a precipitation estimation model that integrates multi-source data and a precipitation estimation method, and relates to the technical field of remote sensing image processing. The method for establishing a precipitation estimation model that integrates multi-source data described in the present invention includes: acquiring observation data; using a gated recurrent neural network and a neural ordinary differential equation network to determine the hidden features that are continuously distributed in the time dimension according to the high-dimensional spatiotemporal features; using a self-attention mechanism and a multi-layer perception mechanism to determine the estimated precipitation according to the hidden features; using the observation data as input and the estimated precipitation as output to form a training set for training, and establishing a precipitation estimation model that integrates multi-source data when the parameter requirements are met. The technical solution described in the present invention achieves high-precision, fine-grained minute-level quantitative precipitation estimation by integrating multiple observation data using a deep learning method.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a precipitation estimation model establishment method and a precipitation estimation method that fuses multi-source data. Background Art

[0002] Precipitation is a common occurrence in daily life. Heavy rainfall, a key factor in triggering extreme natural disasters, causes significant personal and economic losses. Precipitation is affected by various meteorological factors and is unevenly distributed, creating significant inconvenience for meteorological researchers. Manual methods are inefficient and prone to errors and omissions. Existing algorithms, due to variations in data period, can only capture precipitation at larger scales, resulting in issues with accuracy and effectiveness. Summary of the Invention

[0003] The problem solved by the present invention is how to improve the accuracy and effectiveness of precipitation estimation results.

[0004] To solve the above problems, the present invention provides a method for establishing a precipitation estimation model by fusing multi-source data, comprising: acquiring observation data, wherein the observation data includes radar echo image data, satellite cloud image data, and rainfall data and numerical data monitored by ground automatic monitoring stations; using an attention mechanism to determine corresponding high-dimensional spatiotemporal features based on the observation data; using a gated recurrent neural network and a neural ordinary differential equation network to determine latent features continuously distributed in the time dimension based on the high-dimensional spatiotemporal features; using a self-attention mechanism and a multi-layer perception mechanism to determine estimated precipitation based on the latent features; using the observation data as input and the estimated precipitation as output to form a training set for training, and establishing a precipitation estimation model that fuses multi-source data when parameter requirements are met.

[0005] The method for establishing a precipitation estimation model by fusing multi-source data described in the present invention uses a deep learning method to fuse multiple observation data. On the one hand, it combines multiple observation data such as radar and satellite, which can give full play to the advantages of different data. On the other hand, due to the use of convolutional neural networks and gated recurrent units, it combines information in the time dimension and the space dimension, and makes high-dimensional spatiotemporal features continuous, ultimately achieving high-precision, fine-grained minute-level quantitative precipitation estimation.

[0006] Optionally, the use of an attention mechanism to determine the corresponding high-dimensional spatiotemporal features based on the observation data includes: using a recurrent neural network combined with a temporal attention mechanism to determine the high-dimensional temporal features based on the temporal attribute information of the observation data; using a deep residual network combined with a spatial attention mechanism to determine the high-dimensional spatial features based on the spatial attribute information of the observation data; determining the high-dimensional spatiotemporal features based on the high-dimensional temporal features and the high-dimensional spatial features.

[0007] The method for establishing a precipitation estimation model that integrates multi-source data described in the present invention adopts a temporal attention mechanism and a spatial attention mechanism to capture the temporal attribute information and spatial attribute information of the observation data respectively, and then determines high-dimensional spatiotemporal characteristics, which is conducive to better capturing the spatiotemporal characteristics of the input data and ultimately improving the accuracy of precipitation estimation.

[0008] Optionally, the specific steps of using the attention mechanism include: encoding the observation data and mapping it to a feature space to obtain features; determining a query vector, a key vector, and a value vector based on the features, and determining attention in combination with a normalized exponential function.

[0009] The method for establishing a precipitation estimation model by fusing multi-source data described in the present invention allocates different attention to different features through an attention mechanism, so that the network pays more attention to key information, and ultimately improves the accuracy of precipitation estimation.

[0010] Optionally, the method for establishing a precipitation estimation model by fusing multi-source data further includes: using a fully connected neural network to transform the features and fusing them through a dot product method.

[0011] The method for establishing a precipitation estimation model by fusing multi-source data described in the present invention uses a fully connected neural network to transform the features and fuse them through a dot product method, thereby obtaining continuous multi-source fusion high-dimensional features and realizing feature fusion.

[0012] Optionally, the use of a gated recurrent neural network and a neural ordinary differential equation network to determine the hidden features continuously distributed in the time dimension based on the high-dimensional spatiotemporal features includes: determining a reset gate and an update gate based on the hidden state at the previous moment and the input at the current moment; determining a hidden state difference based on the reset gate and the update gate, determining a corresponding differential equation based on the hidden state difference, and using the neural ordinary differential equation network to integrate the differential equation to determine the hidden features at any moment.

[0013] The method for establishing a precipitation estimation model that integrates multi-source data described in the present invention uses a gated recurrent neural network and a neural ordinary differential equation network to determine latent features that are continuously distributed in the time dimension. This can align the latent features of different data in time, thereby obtaining the latent features at any moment, outputting minute-level rainfall, and achieving more fine-grained minute-level quantitative precipitation estimation.

[0014] Optionally, the use of the self-attention mechanism and the multi-layer perception mechanism to determine the estimated precipitation based on the latent features includes: using the self-attention mechanism to assign different weights to the latent features, aggregating the latent features through an aggregation network, and performing regression through the multi-layer perception mechanism to determine the estimated precipitation.

[0015] The method for establishing a precipitation estimation model by fusing multi-source data described in the present invention adopts a self-attention mechanism and a multi-layer perception mechanism to determine the estimated precipitation amount, which is conducive to improving the accuracy of precipitation estimation.

[0016] Optionally, the method for establishing a precipitation estimation model by fusing multi-source data further includes: setting filtering conditions, and using a binary search algorithm to retrieve observation data that meets the filtering conditions from a database.

[0017] The method for establishing a precipitation estimation model by fusing multi-source data described in the present invention can improve the typicality of the precipitation estimation model by setting filtering conditions and retrieving observation data that meets the filtering conditions from a database using a binary search algorithm.

[0018] Optionally, the method for establishing a precipitation estimation model by fusing multi-source data further includes: balancing the data distribution of the training set by using SMOTE oversampling and EasyEnsemble undersampling.

[0019] The method for establishing a precipitation estimation model by fusing multi-source data described in the present invention balances the data distribution of the training set through SMOTE oversampling and EasyEnsemble undersampling, thereby improving the quality of the training set and further facilitating the improvement of the accuracy of precipitation estimation.

[0020] The present invention also provides a precipitation estimation method using multi-source data fusion, comprising: obtaining observational data of precipitation to be estimated; inputting the observational data of precipitation to be estimated into a precipitation estimation model based on multi-source data fusion established by the aforementioned method for establishing a precipitation estimation model based on multi-source data fusion, to obtain estimated precipitation. The aforementioned method for estimating precipitation using multi-source data fusion offers the same advantages as the aforementioned method for establishing a precipitation estimation model based on multi-source data fusion over existing technologies, and therefore will not be further elaborated upon here.

[0021] The present invention also provides a computer-readable storage medium storing a computer program. When read and executed by a processor, the computer program implements the above-described method for establishing a precipitation estimation model by fusing multi-source data or the method for estimating precipitation by fusing multi-source data. The advantages of the computer-readable storage medium and the above-described method for establishing a precipitation estimation model by fusing multi-source data over the prior art are the same and will not be further elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart of a method for establishing a precipitation estimation model by fusing multi-source data according to an embodiment of the present invention;

[0023] Figure 2 A modular schematic diagram of a method for establishing a precipitation estimation model by fusing multi-source data according to an embodiment of the present invention;

[0024] Figure 3 Schematic diagram of a spatiotemporal feature extraction module according to an embodiment of the present invention;

[0025] Figure 4 Schematic diagram of a fusion and continuous module according to an embodiment of the present invention;

[0026] Figure 5 Schematic diagram of a feature regression module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] Doppler weather radar, as a remote sensing method, emits electromagnetic waves that reflect and scatter off target objects, producing various echo patterns. This allows detection of meteorological echoes over a wide area. The direct cause of these echoes is the backscattering of electromagnetic waves by clouds and various hydrometeors in precipitation. Its high data resolution and wide scanning range enable all-day, all-weather observations, making it one of the most powerful small-scale precipitation monitoring methods. Ground-based automatic stations, as a direct means of measuring precipitation, can provide relatively accurate precipitation values ​​over a small area. However, due to their high maintenance costs, these stations are sparsely distributed and lack comprehensive coverage. Both Doppler radar and ground-based automatic station rainfall monitoring have their own advantages. Combining these strengths is an effective way to achieve more accurate precipitation estimates.

[0028] Currently, satellite data, as well as various other data sources such as temperature and humidity, are widely used in the precipitation field. However, existing technologies mostly focus on single-source inputs such as radar and satellite data. Precipitation variability is complex. As atmospheric motion changes, radar echoes can shift, change, and split over time. Furthermore, radar echo images are subject to a certain amount of ground clutter and erroneous information noise, and data may be missing during certain periods due to hardware issues. Therefore, technologies that rely solely on radar echo information cannot effectively improve the accuracy of model outputs. Satellite data has more channels and contains richer information, but the irregularity of data acquisition limits the progress of precipitation estimation at a finer temporal granularity. Different data contain different information and contribute to different aspects of the final precipitation results. However, different observation methods vary in spatial distribution and temporal scales. How to fully utilize the advantages of different observational data is a hot topic of research.

[0029] Traditional techniques currently used to address quantitative precipitation estimation include the optical flow method, which uses a continuous image sequence to detect subtle changes in object motion. This method estimates cloud motion from the current radar echo image and generates a radar echo map through semi-Lagrangian linear extrapolation. This method is more suitable for tracking and predicting echo targets with larger scale and longer durations. However, when echoes split or merge, the accuracy of the precipitation output drops rapidly. Another common technique is the cross-correlation method, one of the earlier researched tracking algorithms. This method treats the echoes across the entire radar image as a whole, tracking the movement of the entire echo region to obtain the final precipitation result. However, as the extrapolation time increases, the echoes experience varying degrees of radiometric distortion, reducing the accuracy of the final result. Numerical weather prediction technology, on the other hand, is based on meteorological physics methods, describing the evolution of weather based on the atmospheric environment and ultimately predicting future atmospheric motion and weather phenomena based on numerical results. However, this method is limited by startup time and cannot meet the requirements of short-term, minute-level precipitation prediction. The above-mentioned method uses only a single observation data set for modeling, which fails to capture atmospheric patterns and has poor robustness. Therefore, the present invention integrates multiple observation data sets and uses a deep learning neural network to train a regression model to achieve highly effective and accurate minute-level quantitative precipitation estimation.

[0030] In addition to traditional methods, there are also technologies that incorporate machine learning and deep learning. Existing machine learning technologies, due to limitations in machine learning's expressive power, cannot effectively model nonlinear relationships, further limiting model accuracy. Furthermore, due to the high feature requirements of the models, significant effort is required in feature engineering. Furthermore, regarding overfitting, traditional machine learning methods generally lack the ability to prevent it and place high demands on the amount and purity of training data, the dimensionality of features, and model complexity. Existing deep learning technologies also suffer from two issues. First, existing technologies mostly employ convolutional neural networks (CNNs) and feedforward neural networks (FNNs), which limit accuracy due to the translation invariance of CNNs. For example, in radar echo maps, areas with high echoes have a greater impact on the model's output of precipitation. Therefore, this patent utilizes an attention mechanism, assigning different weights to different regions through model calculations, allowing the model to focus more on areas with a greater impact on precipitation. Second, due to the limited operating cycle of radar echo data, existing technologies for quantitative precipitation estimation tasks are limited to hourly time granularity, which limits the effectiveness required for precipitation estimation tasks. Therefore, the present invention explores precipitation tasks with finer time granularity, realizes minute-level quantitative precipitation estimation, improves the effectiveness of the technology, and solves the shortcomings of existing methods that rely heavily on manual and feature processing, thereby improving the accuracy of the results.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, an embodiment of the present invention provides a method for establishing a precipitation estimation model that integrates multi-source data, including: obtaining observation data, wherein the observation data includes radar echo image data, satellite cloud image data, and rainfall data and numerical data monitored by ground automatic monitoring stations; using an attention mechanism to determine corresponding high-dimensional spatiotemporal features based on the observation data; using a gated recurrent neural network and a neural ordinary differential equation network to determine latent features continuously distributed in the time dimension based on the high-dimensional spatiotemporal features; using a self-attention mechanism and a multi-layer perception mechanism to determine estimated precipitation based on the latent features; using the observation data as input and the estimated precipitation as output to form a training set for training, and establishing a precipitation estimation model that integrates multi-source data when parameter requirements are met.

[0033] Specifically, in this embodiment, combined with Figure 2 As shown in the figure, observation data is collected from relevant meteorological data provided by various local meteorological bureaus and the National Meteorological Administration. This includes radar echo image data, satellite cloud image data, rainfall data from ground-based automatic monitoring stations, and numerical data such as temperature and humidity. These different observation data are input into the system. First, the spatiotemporal feature extraction module generates a high-dimensional spatiotemporal feature representation of the different observation data. Then, the fusion and continuation module generates a continuous fusion representation, combining the features of the different data. Finally, the attention regression module generates the final precipitation estimation result.

[0034] Combine Figure 3 As shown in Figure 1, the spatiotemporal feature extraction module consists of two key components. The first is the temporal feature extraction network, which focuses on the temporal attributes of the input observation data, capturing them to generate a high-dimensional temporal feature representation. The second is the spatial feature extraction network, which focuses on the spatial attributes of the input observation data, such as changes in radar echoes, and represents them as high-dimensional spatial features. The combination of these two components enables the spatiotemporal feature extraction module to better capture the spatiotemporal characteristics of the input data, such as the motion trends of radar echoes and the changing patterns of satellite cloud images.

[0035] Combine Figure 4 As shown in the figure, the fusion and continuation module accepts the high-dimensional spatiotemporal features of multi-source data, obtains a unified representation of different features through the feature conversion network, and fuses them. For the fused features, the continuation operation is used to obtain the hidden features that are continuously distributed in the time dimension.

[0036] Combine Figure 5As shown in the figure, the feature regression module receives the fused and continuous latent features as input, first obtains different weights through the self-attention network, then processes the continuous latent features through the aggregation network, and finally obtains the final minute-level precipitation output through the regression network.

[0037] In terms of data collection and feature engineering, the present invention pre-processes the multi-source precipitation observation data provided by the Shenzhen Meteorological Bureau and the National Meteorological Administration, removes missing values, and automatically aligns the data according to the regional longitude and latitude, and associates the multi-source data with the rainfall values ​​monitored by the ground automatic monitoring station, which is mainly used to solve the precipitation task in Guangdong Province. In order to improve the efficiency of data reading, the present invention processes the data into a standardized binary format to improve the training speed of the model. Guangdong Province belongs to the East Asian monsoon region, and has a mid-subtropical, south-subtropical and tropical climate from north to south. The precipitation is very abundant. Guangdong is frequently affected by typhoons and has many heavy rainfall weathers. Therefore, the precipitation data in Guangdong Province has good research value. The present invention uses relevant meteorological knowledge to realize the automatic screening of typical processes by manually formulating precipitation conditions, and screens out typical precipitation examples in recent years, totaling 58,061 typical examples. A typical example includes corresponding radar echo maps, satellite cloud maps, numerical data such as temperature and humidity, and data such as precipitation from ground automatic monitoring stations. All typical examples are divided into training set, validation set and test set. For the training set, we use data augmentation algorithms such as SMOTE oversampling and EasyEnsemble undersampling to balance the data distribution and improve its quality. The validation and test sets are left unprocessed to simulate real-world application scenarios and achieve more accurate results. After completing the above feature engineering, we use the data to train the model.

[0038] The input data of the model is not a static picture, but has spatiotemporal attributes, such as the changing trend of radar echoes over time, the movement trend of satellite cloud images over time, etc. In the process of feature extraction of input data, the spatiotemporal characteristics of the data must be fully considered. Therefore, the spatiotemporal feature extraction module independently developed by the present invention focuses on the temporal attribute characteristics and spatial attribute characteristics of the data respectively, and combines the two to obtain spatiotemporal feature representation. The temporal attribute feature special zone part uses RNN (Recurrent Neural Network) combined with temporal attention, and the spatial attribute feature extraction part uses ResNet (Deep Residual Network) with spatial attention mechanism to implement. Since different parts of the data have different contributions to feature representation, the present invention considers introducing an attention mechanism to assign different weights to them.

[0039] The calculation process of the attention mechanism is as follows:

[0040] 1. Encode the input data x and map it to the feature space. The formula is z i =Wx i +b, where x i Represents different multi-source data i, and obtains the encoded feature z i .

[0041] 2. For the encoded features obtained, the query, key, and value vectors are calculated respectively to calculate the weights. The formulas for calculating the query, key, and value vectors (query, key, value) are as follows:

[0042] Q i =W Q LayerNormlize(z i )

[0043] K i =W K LayerNormlize(z i )

[0044] V i =W V LayerNormlize(z i )

[0045] LayerNormlize represents operations such as translation and scaling of features in the feature space, normalizing the distribution of features so that the network can better learn the feature representation of the data. Q 、W K 、W V Represent different weight change matrices respectively.

[0046] 3. Calculate the formula using the dot product The final calculation results are obtained, where It is a scaling factor that plays a normalization role. SoftMax() is used to map the output results to the (0,1) interval and convert them into probabilities.

[0047] The attention mechanism allows the network to assign different levels of attention to different features, focusing on key information, such as areas with high echoes in radar data. This information has a greater impact on the final precipitation output. By combining temporal and spatial attention, a feedforward neural network can be used to obtain high-dimensional spatiotemporal features of the data.

[0048] Due to the different working cycles of multi-source data and ground automatic stations, the model is limited from exploring tasks with finer time granularity. Therefore, the present invention proposes a latent feature continuity module to realize the continuity of spatiotemporal features. For the feature continuity part, the present invention introduces GRU (Gate Recurrent Neural Network) and NeuralODE (Neural Ordinary Differential Equations) to realize the continuity of latent features (a key factor in achieving minute-level precipitation). There is a special gating mechanism in the gated recurrent neural network that can capture long-term time series information. The reset gate setting helps to capture short-term dependencies in the time series, while the update gate helps to capture long-term dependencies in the time series. Reset gate R t and update gate Z t The calculation formula is as follows:

[0049] R t =σ(X t W xr +H t-1 W hr +b r )

[0050] Z t =σ(X t W xz +H t-1 W hz +b z )

[0051] Among them, σ is the activation function sigmoid function, reset gate R t and update gate Z t The value range of each element in H is [0,1]. t-1 is the hidden state of the previous moment, X t is the input at the current moment. Next, through the formula:

[0052]

[0053] The candidate hidden state can be calculated, and then the new hidden state is obtained through the update memory stage of the gated recurrent neural network. xr 、W hr 、W xz 、W hz 、W xh 、W hh Represent different weight parameter matrices. b r 、b z 、b hThey represent different bias matrices respectively. tanh() is the tanh activation function, which can improve the representation ability of the model.

[0054]

[0055] From this we can get the hidden state difference as follows:

[0056]

[0057] Then the differential equation can be obtained through the following formula:

[0058]

[0059] in Represents the derivative of the hidden layer with respect to time t, which is parameterized by a deep neural network and further integrated by a neural ordinary differential equation network to obtain the hidden state at any time t0. This allows for the continuity of the hidden state. Therefore, for multi-source data with different periodic scales, the temporal dimension can be aligned through continuity, leading to finer-grained minute-level rainfall.

[0060] The multi-source high-dimensional spatiotemporal representations obtained by the spatiotemporal feature extraction module are used as input and serialized through a deep neural network to obtain continuous latent features. The features are transformed through an FCN (Fully Connected Neural Network) and then fused through dot products, resulting in continuous multi-source fused high-dimensional features.

[0061] For continuous latent features, aggregation regression is performed through the feature regression module, and a self-attention mechanism encoder is introduced. Different weights are assigned through the self-attention process to aggregate the continuous latent features, and regression is performed through a multi-layer perceptron to obtain the final precipitation.

[0062] In terms of model training, the processed training data is input into the model for training. The trained model accurately completes the precipitation estimation task, improves the effectiveness of precipitation, and achieves finer-grained minute-level quantitative precipitation estimation, which is not possible with traditional methods. Traditional methods such as optical flow, numerical estimation, and cross-correlation can only achieve quantitative precipitation estimation at hourly granularity. At the same time, this method uses minute-level precipitation results to accumulate hourly precipitation as an evaluation, which is higher than traditional methods in indicators such as RMSE. It can be seen that the use of deep learning methods to fuse multi-source data improves the accuracy of quantitative precipitation estimation tasks. For example, in terms of time granularity, existing methods such as optical flow, Cotrec, XGBoost (Extreme Gradient Boosting, extreme gradient boosting), and Random Forest can only achieve time granularity of 30 minutes and 1 hour, but the time granularity of the present invention can reach 5 minutes / 6 minutes.

[0063] Due to the time period difference between radar echo data and ground automatic station data, the radar echo map is updated every 6 minutes, while the precipitation data of the ground automatic station is updated every 5 minutes (each time the cumulative precipitation data of 5 minutes is updated). Therefore, in order to eliminate the time period difference, the existing method can only take the common multiple of the update period of the two. By accumulating 5 radar echo maps to 30 minutes (or accumulating 10 radar echo maps to 1 hour, or accumulating 20 radar echo maps to 2 hours), quantitative precipitation estimation with half-hour or hourly granularity is achieved. However, this approach limits the effectiveness of the task and data utilization. The present invention uses a gated recurrent neural network combined with a neural ordinary differential equation to continuousize the latent features, which can achieve quantitative precipitation estimation with a time granularity of 5 minutes or 6 minutes, and is updated synchronously with the ground automatic monitoring station and radar echo map, greatly improving the utilization of data (in fact, the present invention can achieve precipitation output with any minute granularity, but there is no practical application demand for other minute granularities. With the development of equipment, if necessary, the output of other minute granularities can be achieved by simply replacing the data set of the training model).

[0064] In this embodiment, a deep learning method is used to fuse multiple observation data. On the one hand, it combines multiple observation data such as radar and satellite data to give full play to the advantages of different data. On the other hand, due to the use of convolutional neural networks and gated recurrent units, the high-dimensional spatiotemporal features are continuous by combining information in the time and space dimensions, ultimately achieving high-precision, fine-grained minute-level quantitative precipitation estimation.

[0065] Optionally, the use of an attention mechanism to determine the corresponding high-dimensional spatiotemporal features based on the observation data includes: using a recurrent neural network combined with a temporal attention mechanism to determine the high-dimensional temporal features based on the temporal attribute information of the observation data; using a deep residual network combined with a spatial attention mechanism to determine the high-dimensional spatial features based on the spatial attribute information of the observation data; determining the high-dimensional spatiotemporal features based on the high-dimensional temporal features and the high-dimensional spatial features.

[0066] Specifically, in this embodiment, combined with Figure 3 As shown in the figure, the temporal attention mechanism and the spatial attention mechanism are used to capture the temporal attribute information and spatial attribute information of the observation data respectively, which are expressed as high-dimensional temporal features and high-dimensional spatial features. The combination of the two enables the spatiotemporal feature extraction module to better capture the spatiotemporal characteristics of the input data, such as the motion trend of radar echoes and the changing patterns of satellite cloud images.

[0067] In this embodiment, the temporal attention mechanism and the spatial attention mechanism are used to capture the temporal attribute information and spatial attribute information of the observation data respectively, and then determine the high-dimensional spatiotemporal features, which is conducive to better capturing the spatiotemporal characteristics of the input data and ultimately improving the accuracy of precipitation estimation.

[0068] Optionally, the specific steps of using the attention mechanism include: encoding the observation data and mapping it to a feature space to obtain features; determining a query vector, a key vector, and a value vector based on the features, and determining attention in combination with a normalized exponential function.

[0069] Specifically, in this embodiment, query is a query vector representing information related to the learning task, key is a key vector representing the attributes of the pixel itself, and value is a value vector representing the feature representation of the pixel. Because different parts of the data contribute differently to the feature representation, an attention mechanism is introduced to assign different weights to them. This allows different features to be given different attention, allowing the network to focus more on key information, such as areas with high echo in radar data, which have a greater impact on the final precipitation output.

[0070] In this embodiment, different attention is assigned to different features through the attention mechanism, so that the network pays more attention to key information, and ultimately improves the accuracy of precipitation estimation.

[0071] Optionally, the method for establishing a precipitation estimation model by fusing multi-source data further includes: using a fully connected neural network to transform the features and fusing them through a dot product method.

[0072] Specifically, in this embodiment, the features are converted through FCN (Fully Connected Neural Network) and then fused through dot product, thereby obtaining continuous multi-source fused high-dimensional features.

[0073] In this embodiment, a fully connected neural network is used to transform the features and fuse them through a dot product method, thereby obtaining continuous multi-source fused high-dimensional features and realizing feature fusion.

[0074] Optionally, the use of a gated recurrent neural network and a neural ordinary differential equation network to determine the hidden features continuously distributed in the time dimension based on the high-dimensional spatiotemporal features includes: determining a reset gate and an update gate based on the hidden state at the previous moment and the input at the current moment; determining a hidden state difference based on the reset gate and the update gate, determining a corresponding differential equation based on the hidden state difference, and using the neural ordinary differential equation network to integrate the differential equation to determine the hidden features at any moment.

[0075] Specifically, in this embodiment, the continuity of the hidden state can be achieved through the neural ordinary differential equation plus the gated recurrent network, which can then align the latent features of different data in time. By deducing the neural ordinary differential equation, the latent features at any time can be obtained, and the rainfall at the minute level can be output.

[0076] Other existing methods cannot achieve minute-level precipitation mainly due to the differences in the periodicity of multi-source data. For example, the radar echo map is updated every 6 minutes, and the automatic station rainfall is updated every 5 minutes. There is a difference in the period between the two. Therefore, in order to utilize the data, the common multiple of the two, 30 minutes or 1 hour, is taken as the time granularity for quantitative precipitation estimation.

[0077] In this embodiment, a gated recurrent neural network and a neural ordinary differential equation network are used to determine the latent features that are continuously distributed in the time dimension, so that the latent features of different data can be aligned in time, and then the latent features at any time can be obtained, and the rainfall at the minute level can be output, thereby achieving a more fine-grained minute-level quantitative precipitation estimation.

[0078] Optionally, the use of the self-attention mechanism and the multi-layer perception mechanism to determine the estimated precipitation based on the latent features includes: using the self-attention mechanism to assign different weights to the latent features, aggregating the latent features through an aggregation network, and performing regression through the multi-layer perception mechanism to determine the estimated precipitation.

[0079] Specifically, in this embodiment, for the continuous latent features, aggregation regression is performed through the feature regression module, and a self-attention mechanism encoder is introduced. Different weights are assigned through the self-attention process to aggregate the continuous latent features, and regression is performed through the multi-layer perceptron to obtain the final precipitation.

[0080] In this embodiment, the self-attention mechanism and the multi-layer perception mechanism are used to determine the estimated precipitation, which is conducive to improving the accuracy of precipitation estimation.

[0081] Optionally, the method for establishing a precipitation estimation model by fusing multi-source data further includes: setting filtering conditions, and using a binary search algorithm to retrieve observation data that meets the filtering conditions from a database.

[0082] Specifically, in this embodiment, relevant meteorological knowledge is used to automatically screen typical processes by manually formulating precipitation conditions, and typical precipitation samples in recent years are screened out as observation data samples, which can improve the typicality of the precipitation estimation model.

[0083] In this embodiment, by setting filtering conditions and using a binary search algorithm to retrieve observation data that meets the filtering conditions from the database, the typicality of the precipitation estimation model can be improved.

[0084] Optionally, the method for establishing a precipitation estimation model by fusing multi-source data further includes: balancing the data distribution of the training set by using SMOTE oversampling and EasyEnsemble undersampling.

[0085] Specifically, in this embodiment, a typical sample includes corresponding radar echo images, satellite cloud images, numerical data such as temperature and humidity, and precipitation data from automatic ground monitoring stations. All typical samples are divided into training, validation, and test sets. For the training set, data augmentation algorithms such as SMOTE oversampling and EasyEnsemble undersampling are used to balance the data distribution and improve the quality of the training set. The validation and test sets are left unprocessed to simulate real-world application scenarios and obtain more accurate results and indicators.

[0086] In this embodiment, the data distribution of the training set is balanced through SMOTE oversampling and EasyEnsemble undersampling, thereby improving the quality of the training set and further improving the accuracy of precipitation estimation.

[0087] Another embodiment of the present invention also provides a precipitation estimation method by fusing multi-source data, comprising: obtaining observation data of the precipitation to be estimated, inputting the observation data of the precipitation to be estimated into the precipitation estimation model by fusing multi-source data established by the above-mentioned precipitation estimation model establishment method by fusing multi-source data, and obtaining estimated precipitation.

[0088] Another embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is read and executed by a processor, it implements the above-mentioned method for establishing a precipitation estimation model by fusing multi-source data or the method for precipitation estimation by fusing multi-source data.

[0089] Although the present invention is disclosed as above, the scope of protection disclosed by the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for establishing a precipitation estimation model by integrating multi-source data, characterized in that: include: Acquiring observation data, wherein the observation data includes radar echo image data, satellite cloud image data, and rainfall data and numerical data monitored by ground automatic monitoring stations; Using an attention mechanism to determine corresponding high-dimensional spatiotemporal features based on the observed data; Using a gated recurrent neural network and a neural ordinary differential equation network to determine the hidden features continuously distributed in the time dimension according to the high-dimensional spatiotemporal features; Using a self-attention mechanism and a multi-layer perception mechanism to determine the estimated precipitation according to the latent features; The observed data is used as input, and the estimated precipitation is used as output to form a training set for training, and when the parameter requirements are met, a precipitation estimation model integrating multi-source data is established; Wherein, the use of the attention mechanism to determine the corresponding high-dimensional spatiotemporal features according to the observed data includes: Using a recurrent neural network combined with a temporal attention mechanism, high-dimensional temporal features are determined according to the temporal attribute information of the observed data; Using a deep residual network combined with a spatial attention mechanism to determine high-dimensional spatial features based on the spatial attribute information of the observed data; Determine the high-dimensional spatiotemporal feature according to the high-dimensional time series feature and the high-dimensional spatial feature; Among them, the specific steps of adopting the attention mechanism include: The observed data is encoded and mapped to a feature space to obtain features, wherein the features are expressed as: z i =Wx i +b; Among them, z i represents the feature, x i represents different observed data, W and b are corresponding coefficients; Determine a query vector, a key vector, and a value vector according to the features, and determine attention in combination with a normalized exponential function; Wherein, the query vector is expressed as: Q i =W Q LayerNormlize(z i ); Wherein, the key vector is expressed as: K i =W K LayerNormlize(z i ); Wherein, the value vector is expressed as: V i =In V LayerNormlize(z i ); Among them, Q i , K i and V i denote the query vector, the key vector and the value vector respectively, W Q , W K and W V They represent different weight change matrices respectively, and LayerNormlize represents the translation and scaling changes of the features in the feature space; Among them, the specific steps of using the attention mechanism also include using a fully connected neural network to transform the features and fuse them through a dot product method, wherein the dot product calculation formula is expressed as: Among them, attention(Q i ,K i ,V i ) represents the attention, SoftMax() maps the output result to the interval (0,1) and converts it into probability. represents the scaling factor; Wherein, the use of gated recurrent neural network and neural ordinary differential equation network to determine the hidden features continuously distributed in the time dimension according to the high-dimensional spatiotemporal features includes: Determine the reset gate and update gate based on the hidden state at the previous moment and the input at the current moment; Wherein, the reset gate is expressed as: R t =σ(X t W xr +H t-1 W hr +b r ); Wherein, the update gate is expressed as: Z t =σ(X t W xz +H t-1 W hz +b z ); Among them, σ represents the activation function sigmoid function, H t-1 represents the hidden state at the previous moment, X t Represents the input at the current moment; Determine a hidden state difference according to the reset gate and the update gate, determine a corresponding differential equation according to the hidden state difference, and use a neural network of ordinary differential equations to integrate the differential equation to determine a hidden feature at any time; Among them, the candidate hidden state is expressed as: Among them, W xr , W hr , W xz , W hz , W xh and W hh Represent different weight parameter matrices, b r , b z , b h They represent different bias matrices respectively, and tanh() represents the tanh activation function; Among them, the new hidden state is expressed as: Among them, the hidden state difference is expressed as: Wherein, the differential equation is expressed as: in, represents the derivative of the hidden layer with respect to time t; The method of using the self-attention mechanism and the multi-layer perception mechanism to determine the estimated precipitation according to the latent features includes: The self-attention mechanism is used to assign different weights to the latent features, the latent features are aggregated through an aggregation network, and the estimated precipitation is determined by regression through the multi-layer perception mechanism.

2. The method for establishing a precipitation estimation model by integrating multi-source data according to claim 1, characterized in that: Also includes: Set filtering conditions and use a binary search algorithm to retrieve observation data that meets the filtering conditions from the database.

3. The method for establishing a precipitation estimation model by integrating multi-source data according to claim 1, characterized in that: Also includes: The data distribution of the training set is balanced by SMOTE oversampling and EasyEnsemble undersampling.

4. A precipitation estimation method integrating multi-source data, characterized in that: include: Obtain observation data of precipitation to be estimated, input the observation data of precipitation to be estimated into a precipitation estimation model for fusing multi-source data established by the method for establishing a precipitation estimation model for fusing multi-source data according to any one of claims 1 to 3, and obtain estimated precipitation.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is read and executed by the processor, the method for establishing a precipitation estimation model by fusing multi-source data according to any one of claims 1 to 3 or the method for estimating precipitation by fusing multi-source data according to claim 4 is implemented.

Images