Multi-source rainfall fusion method and system based on improved convolutional long short-term memory network

Abstract

The invention relates to the technical field of meteorological monitoring, in particular to a multi-source rainfall fusion method and system based on an improved convolutional long short-term memory network, and the method comprises the steps: constructing a regional planar rainfall field based on commercial microwave link rainfall observation data and rainfall observation data of a ground rainfall station, and constructing a rainfall data fusion training sample; setting a rainfall label of a rainfall data fusion training sample by using a high-precision ground rainfall station planar rainfall field, wherein the ground rainfall stations are dispersedly arranged at partial positions in the area; and constructing a rainfall time-space fusion model based on the convolutional neural network and the long-short-term memory network, training the rainfall time-space fusion model by using the rainfall data fusion training sample, and predicting a regional rainfall field in the to-be-detected region by using the trained model. According to the invention, the precision and temporal-spatial resolution of rainfall products can be improved, and reliable data support is provided for refined early warning and forecasting of flood disasters.

Description

Technical Field

[0001] This invention relates to the field of meteorological monitoring technology, and in particular to a multi-source rainfall fusion method and system based on an improved convolutional long short-term memory network. The method uses commercial microwave links supported by the improved convolutional long short-term memory network to fuse multi-source rainfall and obtain high-precision, high-spatiotemporal-resolution rainfall products. Background Technology

[0002] Rainfall, as a major contributing factor to floods, directly determines the flood formation process and the risk level of secondary geological disasters through its spatiotemporal distribution characteristics. High-precision rainfall monitoring data is a crucial scientific basis for constructing disaster early warning systems and implementing disaster prevention and mitigation measures. Current mainstream rainfall monitoring technologies mainly include three types of observation methods: ground-based station observations, weather radar estimation, and satellite remote sensing retrieval. However, existing monitoring systems have significant technical limitations: the spatial distribution of ground-based observation stations is highly heterogeneous, the construction and maintenance costs of meteorological radar systems are high, and the temporal resolution of satellite remote sensing technology is significantly insufficient. Therefore, the development of multi-source rainfall data fusion technology has become a crucial breakthrough for improving the spatiotemporal accuracy of hydrological and meteorological monitoring.

[0003] Multi-source rainfall data fusion methods are mainly based on mathematical principles such as weighted averaging and regression analysis, aiming to correct systematic errors in the rainfall observation process. Typical methods include probability matching, objective statistical analysis, co-kriging interpolation, variational assimilation, and Kalman filtering algorithms. Although these traditional methods can achieve spatiotemporal modeling of rainfall, their effectiveness is usually limited by idealized assumptions. Therefore, there is an urgent need for a spatiotemporal modeling technique for rainfall to improve the accuracy and resolution of rainfall products to meet the needs of early warning and forecasting in complex meteorological environments. Summary of the Invention

[0004] To address the issue of insufficient accuracy and resolution in existing rainfall product modeling, this invention provides a multi-source rainfall fusion method and system based on an improved convolutional long short-term memory network. By fusing commercial microwave link observation data and ground rain gauge data, it obtains high-precision, high spatiotemporal resolution rainfall products, providing reliable data support for refined early warning and forecasting of flood disasters.

[0005] According to the design scheme provided by this invention, on the one hand, a multi-source rainfall fusion method based on an improved convolutional long short-term memory network is provided, comprising:

[0006] The process involves acquiring rainfall observation data from commercial microwave links and rainfall observation data from ground rain gauges. A first regional isometric rainfall field is constructed based on the commercial microwave link rainfall observation data. A second regional isometric rainfall field is constructed using rainfall observation data from ground rain gauges within the region. Rainfall data fusion training samples are then constructed based on the first and second regional isometric rainfall fields and topographic parameters. Rainfall labels for the rainfall data fusion training samples are set using the high-precision second regional isometric rainfall field. The ground rain gauges are distributed in some locations within the region.

[0007] A spatiotemporal fusion model for rainfall is constructed based on convolutional neural networks and long short-term memory networks. The spatiotemporal fusion model for rainfall is trained using training samples fused with rainfall data to obtain a target model for spatiotemporal fusion of rainfall. The spatiotemporal fusion model for rainfall extracts regional spatial features of precipitation based on convolutional neural networks, extracts regional temporal features of precipitation based on long short-term memory networks, and uses a channel attention mechanism to learn and quantify the importance of regional precipitation features.

[0008] The topographic parameters, microwave link precipitation observation data, and ground rain gauge precipitation observation data of the area to be measured are obtained. Based on the microwave link precipitation observation data and the ground rain gauge precipitation observation data, the first isometric precipitation field and the second isometric precipitation field of the area to be measured are constructed respectively. The topographic parameters, the first isometric precipitation field and the second isometric precipitation field of the area to be measured are used as inputs to the spatiotemporal fusion target model of precipitation. The spatiotemporal fusion target model of precipitation is used to predict the precipitation field at each location in the area to be measured.

[0009] As a multi-source rainfall fusion method based on an improved convolutional long short-term memory network, this invention further constructs a regional first areal rainfall field based on rainfall observation data from commercial microwave links, comprising:

[0010] An anomaly detection algorithm is used to remove outliers based on signal propagation characteristics. Combined with terrain data, continuous and visible microwave links are obtained, and microwave link observation data is acquired.

[0011] Time alignment of microwave link observation data is performed based on ground rain gauge observation data, and the wet and dry periods of rainfall are determined based on the microwave link signal attenuation amplitude and the moving standard deviation.

[0012] By using a sliding window to traverse the time series of microwave link signals, the minimum signal attenuation amplitude within each sliding window is used as the basic attenuation benchmark for the corresponding time period.

[0013] Based on the empirical model of wireless microwave rain attenuation characteristics proposed by the International Telecommunication Union (ITU-R model), wireless microwave link data is inverted to obtain microwave link observation data. By discretizing the microwave link observation data, a discrete point set of microwave link data is obtained. Based on this discrete point set, a first isometric rainfall field in the region is constructed using a trend surface interpolation algorithm.

[0014] As a multi-source rainfall fusion method based on an improved convolutional long short-term memory network, this invention further utilizes rainfall observation data from surface rain gauges within the region to construct a second isometric rainfall field, comprising:

[0015] The raw rainfall data from ground rain gauges in the region are accumulated over time, and data sequences at specified time intervals are obtained through linear interpolation.

[0016] A second isometric precipitation field for the region is constructed based on the data sequence and using a trend surface interpolation algorithm.

[0017] As a multi-source rainfall fusion method based on an improved convolutional long short-term memory network, this invention further constructs rainfall data fusion training samples based on the first isometric rainfall field and the second isometric rainfall field of the region, as well as topographic parameters, including:

[0018] For the first and second areal precipitation fields in the region, extract the topographic parameters, commercial microwave link precipitation observation data and ground rain gauge precipitation observation data corresponding to a fixed time step.

[0019] Based on the extracted terrain parameters and rainfall observation data, a training dataset was established to correspond in time and space between commercial microwave link rainfall and ground rain gauge rainfall.

[0020] As a multi-source rainfall fusion method based on an improved convolutional long short-term memory network, the present invention further trains a rainfall spatiotemporal fusion model using rainfall data fusion training samples, including:

[0021] The model training loss function is set based on the mean square error;

[0022] The rainfall data fusion training samples are divided into a training set for training the model and a test set for validating the trained model according to a preset ratio.

[0023] The training dataset uses rain gauge observation data, commercial microwave link observation data, and terrain parameter data as inputs to the model training, rainfall labels as ground truth, and the model is trained based on the model training loss function.

[0024] The trained model was validated using a test set, and the optimal model was selected as the target model for rainfall fusion.

[0025] As a multi-source rainfall fusion method based on an improved convolutional long short-term memory network in this invention, the model training loss function is further expressed as: Where n is the number of samples in the set, y i It is true. These are predicted values.

[0026] Furthermore, this invention also provides a multi-source rainfall fusion system based on an improved convolutional long short-term memory network, comprising: a sample construction module, a model training module, and a fusion output module, wherein...

[0027] The sample construction module is used to acquire commercial microwave link rainfall observation data and ground rain gauge rainfall observation data. Based on the commercial microwave link rainfall observation data, a first regional isometric rainfall field is constructed. Using the rainfall observation data from ground rain gauges within the region, a second regional isometric rainfall field is constructed. Based on the first and second regional isometric rainfall fields and topographic parameters, rainfall data fusion training samples are constructed. Rainfall labels are set on the rainfall data fusion training samples using the high-precision second regional isometric rainfall field. The ground rain gauges are scattered in some locations within the region. The isometric rainfall field describes the regional precipitation phenomenon based on meteorological observation data and topographic parameters.

[0028] The model training module is used to construct a rainfall spatiotemporal fusion model based on convolutional neural networks and long short-term memory networks. The rainfall spatiotemporal fusion model is trained using rainfall data fusion training samples to obtain the rainfall spatiotemporal fusion target model. The rainfall spatiotemporal fusion model extracts regional precipitation spatial features based on convolutional neural networks, extracts regional precipitation temporal features based on long short-term memory networks, and uses a channel attention mechanism to learn and quantify the importance of regional precipitation features.

[0029] The fusion output module is used to acquire topographic parameters of the area to be measured, microwave link precipitation observation data, and ground rain gauge precipitation observation data. Based on the microwave link precipitation observation data and the ground rain gauge precipitation observation data, it constructs the first isometric precipitation field and the second isometric precipitation field of the area to be measured, respectively. The topographic parameters of the area to be measured, the first isometric precipitation field, and the second isometric precipitation field are used as inputs to the spatiotemporal fusion target model of precipitation. The spatiotemporal fusion target model of precipitation is used to predict the precipitation field at each location in the area to be measured.

[0030] The beneficial effects of this invention are:

[0031] To address the technical bottlenecks in the current development of high-precision, high-spatiotemporal-resolution rainfall products, this invention fuses multi-source rainfall data from commercial microwave links using an improved convolutional long short-term memory network. This achieves the generation of high-precision, high-spatiotemporal-resolution rainfall products. Considering the issue of inconsistent intervals in rain gauge observation data, a time accumulation method is used for data standardization. Furthermore, considering the non-standardization of microwave link observation data and the impact of other factors on microwave attenuation, an anomaly detection algorithm is used to filter the original microwave link observation data. A dry / wet period discrimination model is established, and a sliding window method is used to dynamically correct the baseline attenuation, which is then compared with rainfall data. The observation data from the stations maintain the same temporal resolution. Based on the empirical model of radio microwave rain attenuation characteristics proposed by the International Telecommunication Union (ITU-R model), radio microwave link data is inverted to obtain microwave link observation data. Considering that the microwave link observation data has a linear structure, trend surface interpolation algorithm is used to obtain isometric microwave link rainfall inversion data. For rain gauge data, isometric rainfall field is constructed through trend surface interpolation algorithm, and deep feature fusion of multi-source data is achieved based on an improved convolutional long short-term memory network to realize the generation of high-precision, high spatiotemporal resolution rainfall products, providing reliable data support for refined early warning and forecasting of flood disasters. Attached image description:

[0032] Figure 1 This is a schematic diagram of the multi-source rainfall fusion process based on an improved convolutional long short-term memory network in the embodiment.

[0033] Figure 2 This is a schematic diagram illustrating the principle of the commercial microwave link multi-source rainfall data fusion algorithm in the embodiment;

[0034] Figure 3 This is a schematic diagram of the structure of the improved convolutional long short-term memory network in the embodiment. Detailed implementation method:

[0035] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0036] Precipitation, as a crucial component of the water cycle, plays a vital role in connecting the biosphere, hydrosphere, lithosphere, and atmosphere. Its spatial distribution is extremely complex, profoundly impacting meteorology, hydrology, and related processes. Therefore, improving the accuracy of precipitation estimation is of great significance for hydrology, meteorology, climate, agriculture, ecology, and natural disaster prevention. To enhance the accuracy and effectiveness of precipitation estimation, this invention provides a multi-source precipitation fusion method based on an improved convolutional long short-term memory network, such as... Figure 1 As shown, it specifically includes the following content:

[0037] S101. Acquire commercial microwave link rainfall observation data and ground rain gauge rainfall observation data. Construct a first regional isometric rainfall field based on the commercial microwave link rainfall observation data. Construct a second regional isometric rainfall field using the rainfall observation data from ground rain gauges within the region. Construct rainfall data fusion training samples based on the first and second regional isometric rainfall fields and terrain parameters. Set rainfall labels for the rainfall data fusion training samples using the high-precision second regional isometric rainfall field (containing more rain gauge data). The ground rain gauges are dispersed in some locations within the region.

[0038] Specifically, the first areal precipitation field in the region, constructed based on precipitation observation data from commercial microwave links, can be designed to include:

[0039] An anomaly detection algorithm is used to remove outliers based on signal propagation characteristics. Combined with terrain data, continuous and visible microwave links are obtained, and microwave link observation data is acquired.

[0040] Time alignment of microwave link observation data is performed based on ground rain gauge observation data, and the wet and dry periods of rainfall are determined based on the microwave link signal attenuation amplitude and the moving standard deviation.

[0041] By using a sliding window to traverse the time series of microwave link signals, the minimum signal attenuation amplitude within each sliding window is used as the basic attenuation benchmark for the corresponding time period.

[0042] Based on the empirical model of wireless microwave rain attenuation characteristics proposed by the International Telecommunication Union (ITU-R model), wireless microwave link data is inverted to obtain microwave link observation data. By discretizing the microwave link observation data, a discrete point set of microwave link data is obtained. Based on this discrete point set, a first isometric rainfall field in the region is constructed using a trend surface interpolation algorithm.

[0043] To address the issue of inconsistent intervals in rain gauge observation data, a time-cumulative method can be used for data standardization. Specifically, the raw rainfall data is accumulated over time, and linear interpolation is used to standardize the data to a fixed time interval. Anomaly detection algorithms are employed to remove outliers based on signal propagation characteristics. Combined with topographic data, continuous and visible microwave links are obtained through topographic analysis, thus preserving the observation data generated by these links. Time alignment processing is then performed on the microwave link observation data to ensure consistency with the time reference and resolution of the rain gauge observation data.

[0044] Considering that the attenuation and standard deviation of microwave signals are relatively stable when there is no rainfall, but when rainfall occurs, the signal attenuation fluctuates significantly, and the standard deviation increases accordingly, in this embodiment, as... Figure 2In the algorithm shown, the moving standard deviation method can be used to determine the wet and dry periods of rainfall. For a wireless microwave attenuation sequence, the total attenuation rate formula for the microwave link is as follows:

[0045] A(t)=A B (t)+A R (t) (1)

[0046] In the formula, A(t) is the total attenuation rate of the microwave link at time t. B (t) represents the basic attenuation rate of the microwave link at time t, A R (t) represents the rain attenuation rate of the microwave link at time t, with units of decibels per kilometer (dB / km).

[0047] In the sliding window W t =[tw,t]

[0048]

[0049] In the formula, N W The number of data items in the window. This represents the average total microwave attenuation rate. This represents the standard deviation within the window.

[0050] The standard deviation threshold can be determined using 24-hour dry period decay data. If the rainfall exceeds the threshold, it is considered rainfall; if the rainfall is less than or equal to the threshold, it is considered no rainfall.

[0051] During microwave signal propagation, signal attenuation caused by factors such as changes in atmospheric water vapor content is defined as basic attenuation. This attenuation caused by non-precipitation factors constitutes a systematic error source, affecting the accuracy of microwave-based rainfall retrieval algorithms, especially during and around precipitation events. To reduce the interference of non-meteorological attenuation, a reliable microwave attenuation benchmark needs to be established before microwave rainfall retrieval calculations. In this embodiment, a sliding standard deviation algorithm can be used for quantitative analysis of basic attenuation. The specific method is as follows: A sliding window is set to traverse the microwave signal time series, and the minimum attenuation value within each window is identified as the benchmark for basic attenuation during that period. By continuously sliding the window, the basic attenuation value is dynamically corrected, thereby completing the calculation of basic attenuation. The formula for this method is shown below:

[0052]

[0053] In the formula, n represents the number of time points, {r1, r2, ... r... n} is an n-dimensional sequence of data representing the signal attenuation during wireless microwave link transmission, where Δ represents the base attenuation value.

[0054] Given the linear distribution of commercial microwave link observation data, it is necessary to transform it into a planar precipitation field using spatial interpolation methods. Before interpolation, the microwave link data can be discretized into a discrete point set. Based on the assumption of spatial uniformity of precipitation intensity along the microwave link path, five discretization schemes can be used: sampling only the link center point (S1), sampling only the two end nodes (S2), sampling at 1000m intervals (S3), sampling at 500m intervals (S4), and sampling at 200m intervals (S5). By systematically comparing the impact of different discretization schemes on the reconstruction accuracy of the two-dimensional precipitation field, the optimal sampling interval was finally determined to be 200m (S5 scheme). Based on the discrete point set generated by this optimal discretization scheme, a high-resolution planar microwave link precipitation field is constructed using a trend surface interpolation algorithm.

[0055] The raw rainfall data from ground rain gauges within the region are accumulated over time, and a data sequence for a specified time interval is obtained through linear interpolation. Based on the data sequence, a second isometric rainfall field for the region can be constructed using a trend surface interpolation algorithm.

[0056] Among them, the rainfall data fusion training samples constructed based on the first and second isometric rainfall fields of the region and topographic parameters can be designed to include:

[0057] For the first and second areal precipitation fields in the region, extract the topographic parameters, commercial microwave link precipitation observation data and ground rain gauge precipitation observation data corresponding to each time step;

[0058] Based on the extracted terrain parameters and rainfall observation data, a training dataset was established to correspond in time and space between commercial microwave link rainfall and ground rain gauge rainfall.

[0059] Rain gauges are typically deployed in a dispersed manner during actual operation. Since microwave link observation data are all predicted values, the dispersed ground rain gauge observation data is used as ground truth labels to construct a sample set based on the corresponding location and time points. In this solution, rain gauge data is integrated into the wireless microwave link observation data, and the rain gauges are divided into training and test sets, with ratios of 70% and 30%, respectively. The rain gauge observation data is naturally also divided into training and test sets. For example, if there are 100 rain gauge locations in a certain study area, the data from 70 of these locations can be selected as the ground truth labels for the training set (used to construct a high-precision areal rainfall field), the data from 35 of these locations can be used as the training set, and the observation data from the remaining 30 rain gauges can be used to construct a test set to test and evaluate the accuracy of the rainfall prediction model after training.

[0060] S102. Construct a spatiotemporal fusion model of rainfall based on convolutional neural network and long short-term memory network. Train the spatiotemporal fusion model of rainfall using rainfall data fusion training samples to obtain the spatiotemporal fusion target model of rainfall. The spatiotemporal fusion model of rainfall extracts regional precipitation spatial features based on convolutional neural network, extracts regional precipitation temporal features based on long short-term memory network, and uses channel attention mechanism to learn and quantify the importance of regional precipitation features.

[0061] Specifically, the rainfall fusion model is trained using training samples from the fusion of rainfall data, including:

[0062] The model training loss function is set based on the mean square error;

[0063] The rainfall data fusion training samples are divided into a training set for training the model and a test set for validating the trained model according to a preset ratio.

[0064] The training dataset uses rain gauge observation data, commercial microwave link observation data, and terrain parameter data as inputs to the model training, rainfall labels as ground truth, and the model is trained based on the model training loss function.

[0065] The trained model was validated using a test set, and the optimal model was selected as the target model for rainfall fusion.

[0066] Considering that existing multi-source rainfall data fusion methods often operate under assumptions, leading to significant errors in real-world scenarios, this embodiment employs an improved convolutional long short-term memory network model incorporating a channel attention mechanism to achieve effective fusion of multi-source rainfall data from commercial microwave links. This improved model integrates the spatial feature extraction capabilities of convolutional neural networks with the temporal modeling advantages of long short-term memory networks, and innovatively introduces a channel attention mechanism. This mechanism dynamically learns and quantifies the feature importance of each input data channel (including rain gauge observation data, wireless microwave rainfall inversion data, and terrain feature parameters such as elevation, slope, and aspect), thereby achieving adaptive weight allocation for key feature channels while suppressing interference from irrelevant or noisy channels.

[0067] Figure 3This is a schematic diagram of the structure of an improved convolutional long short-term memory (LSTM) network. The structure of the improved LSM network includes: a channel attention module (e.g., Equations 5 and 6); an input gate (e.g., Equation 7) that controls the input data such as rainfall, elevation, slope, and aspect into the memory unit; a forget gate (e.g., Equation 8) that controls the degree to which the memory unit forgets historical information; a memory unit (e.g., Equation 9) that generates new candidate memories based on the input and hidden states; an output gate (e.g., Equation 10) that controls the amount of information output from the memory unit; and a convolution operation that processes spatial sequence information such as rainfall, elevation, and slope. The formula for ConvLSTM is shown below.

[0068] M c (X)=σ(MLP(AvgPool(X))+MLP(MaxPool(X))) (5)

[0069] X′=M c (X)*X (6)

[0070] In the formula, X represents the original feature, X′ represents the feature after channel attention processing, AvgPool represents average pooling, MaxPool represents max pooling, MLP represents a fully connected layer, and σ represents the Sigmoid activation function.

[0071]

[0072] In the formula, For Hadamard operations, * represents convolution, σ is the Sigmoid activation function, and i t Preserve the probability of the state of the input gate, f t Preserve the probability of the forget gate state, c t Let o be the state at time t. t Let h be the output probability of the output gate at time t. t For the output of the hidden layer, W xi W hti W ci W xf W hf W cf W xc W hc W ho and W co b represents the weight term. f b i b c and b o This represents the deviation term. The calculations for σ and tanh are as follows:

[0073]

[0074] The improved convolutional long short-term memory network uses mean squared error as the loss function, and its formula is as follows:

[0075]

[0076] In the formula, n is the number of samples, y i For the true value, These are predicted values.

[0077] This improved convolutional long short-term memory network enables deep fusion of rain gauge observation data and wireless microwave link observation data, thereby obtaining high-precision, high spatiotemporal resolution rainfall products.

[0078] S103. Obtain the topographic parameters, microwave link precipitation observation data, and ground rain gauge precipitation observation data of the area to be measured. Based on the microwave link precipitation observation data and the ground rain gauge precipitation observation data, construct the first isometric precipitation field and the second isometric precipitation field of the area to be measured, respectively. Use the topographic parameters, the first isometric precipitation field, and the second isometric precipitation field of the area to be measured as inputs to the spatiotemporal fusion target model of precipitation. Use the spatiotemporal fusion target model of precipitation to predict the precipitation field at each location in the area to be measured.

[0079] The precipitation in the study area is fused based on the spatiotemporal fusion target model of precipitation. The microwave link precipitation observation data and the ground rain gauge precipitation observation data are correlated in space and time. The accuracy of microwave link precipitation prediction at various locations in the study area (including locations in the study area where no ground rain gauge is set) is improved by using the dispersed ground rain gauge precipitation observation data.

[0080] Furthermore, based on the above method, this embodiment of the invention also provides a multi-source rainfall fusion system based on an improved convolutional long short-term memory network, comprising: a sample construction module, a model training module, and a fusion output module, wherein,

[0081] The sample construction module is used to acquire commercial microwave link rainfall observation data and ground rain gauge rainfall observation data, construct a first regional isometric rainfall field based on the commercial microwave link rainfall observation data, construct a second regional isometric rainfall field using the rainfall observation data from ground rain gauges within the region, construct rainfall data fusion training samples based on the first and second regional isometric rainfall fields and topographic parameters, and set rainfall labels for the rainfall data fusion training samples using the high-precision second regional isometric rainfall field. The ground rain gauges are dispersed in some locations within the region.

[0082] The model training module is used to construct a rainfall spatiotemporal fusion model based on convolutional neural networks and long short-term memory networks. The rainfall spatiotemporal fusion model is trained using rainfall data fusion training samples to obtain the rainfall spatiotemporal fusion target model. The rainfall spatiotemporal fusion model extracts regional precipitation spatial features based on convolutional neural networks, extracts regional precipitation temporal features based on long short-term memory networks, and uses a channel attention mechanism to learn and quantify the importance of regional precipitation features.

[0083] The fusion output module is used to acquire the topographic parameters of the area to be measured. It takes the topographic parameters of the area to be measured, the microwave link precipitation observation data of the area to be measured, and the ground rain gauge precipitation observation data of the area to be measured as inputs to the spatiotemporal fusion target model of rainfall, and uses the spatiotemporal fusion target model of rainfall to predict the regional rainfall field in the area to be measured.

[0084] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0086] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0087] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0088] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-source rainfall fusion method based on an improved convolutional long short-term memory network, characterized in that, Include: Acquire rainfall observation data from commercial microwave links and rainfall observation data from ground rain gauges. Construct a first regional isometric rainfall field based on the commercial microwave link rainfall observation data. Construct a second regional isometric rainfall field using rainfall observation data from ground rain gauges within the region. Construct rainfall data fusion training samples based on the first and second regional isometric rainfall fields and topographic parameters. Set rainfall labels for the rainfall data fusion training samples using the second regional isometric rainfall field. The ground rain gauges are dispersed in some locations within the region. A spatiotemporal fusion model for rainfall is constructed based on convolutional neural networks and long short-term memory networks. The spatiotemporal fusion model for rainfall is trained using training samples fused with rainfall data to obtain a target model for spatiotemporal fusion of rainfall. The spatiotemporal fusion model for rainfall extracts regional spatial features of precipitation based on convolutional neural networks, extracts regional temporal features of precipitation based on long short-term memory networks, and uses a channel attention mechanism to learn and quantify the importance of regional precipitation features. The topographic parameters, microwave link precipitation observation data, and ground rain gauge precipitation observation data of the area to be measured are obtained. Based on the microwave link precipitation observation data and the ground rain gauge precipitation observation data, the first isometric precipitation field and the second isometric precipitation field of the area to be measured are constructed respectively. The topographic parameters, the first isometric precipitation field and the second isometric precipitation field of the area to be measured are used as inputs to the spatiotemporal fusion target model of precipitation. The spatiotemporal fusion target model of precipitation is used to predict the precipitation field at each location in the area to be measured.

2. The multi-source rainfall fusion method based on an improved convolutional long short-term memory network according to claim 1, characterized in that, The first isometric precipitation field in the region was constructed based on precipitation observation data from commercial microwave links, including: An anomaly detection algorithm is used to remove outliers based on signal propagation characteristics. Combined with terrain data, continuous and visible microwave links are obtained, and microwave link observation data is acquired. Time alignment of microwave link observation data is performed based on ground rain gauge observation data, and the wet and dry periods of rainfall are determined based on the microwave link signal attenuation amplitude and the moving standard deviation. By using a sliding window to traverse the time series of microwave link signals, the minimum signal attenuation amplitude within each sliding window is used as the basic attenuation benchmark for the corresponding time period. Based on the empirical model of wireless microwave rain attenuation characteristics proposed by the International Telecommunication Union, wireless microwave link data is inverted to obtain microwave link observation data; by discretizing the microwave link observation data, a discrete point set of microwave link data is obtained, and based on this discrete point set, a trend surface interpolation algorithm is used to construct the first isometric rainfall field in the region.

3. The multi-source rainfall fusion method based on an improved convolutional long short-term memory network according to claim 1, characterized in that, A second areal precipitation field for the region is constructed using rainfall observation data from surface rain gauges within the region, including: The raw rainfall data from ground rain gauges in the region are accumulated over time, and data sequences at specified time intervals are obtained through linear interpolation. A second isometric precipitation field for the region is constructed based on the data sequence and using a trend surface interpolation algorithm.

4. The multi-source rainfall fusion method based on an improved convolutional long short-term memory network according to claim 1, characterized in that, A training sample for rainfall data fusion is constructed based on the first and second areal rainfall fields of the region and topographic parameters, including: For the first and second areal precipitation fields in the region, extract the topographic parameters, commercial microwave link precipitation observation data and ground rain gauge precipitation observation data corresponding to a fixed time step. Based on the extracted terrain parameters and rainfall observation data, a training dataset was established to correspond in time and space between commercial microwave link rainfall and ground rain gauge rainfall.

5. The multi-source rainfall fusion method based on an improved convolutional long short-term memory network according to claim 1, characterized in that, The spatiotemporal fusion model of rainfall is trained using training samples fused from rainfall data, including: The model training loss function is set based on the mean square error; The rainfall data fusion training samples are divided into a training set for training the model and a test set for validating the trained model according to a preset ratio. The training dataset uses rain gauge observation data, commercial microwave link observation data, and terrain parameter data as inputs to the model training, rainfall labels as ground truth, and the model is trained based on the model training loss function. The trained model was validated using a test set, and the optimal model was selected as the target model for rainfall fusion.

6. The multi-source rainfall fusion method based on an improved convolutional long short-term memory network according to claim 5, characterized in that, The model training loss function is expressed as: Where n is the number of samples in the set, y i It is true. These are predicted values.

7. A multi-source rainfall fusion system based on an improved convolutional long short-term memory network, characterized in that, It includes: a sample construction module, a model training module, and a fusion output module, among which, The sample construction module is used to acquire commercial microwave link rainfall observation data and ground rain gauge rainfall observation data, construct a first regional isometric rainfall field based on the commercial microwave link rainfall observation data, construct a second regional isometric rainfall field using the rainfall observation data from ground rain gauges within the region, construct rainfall data fusion training samples based on the first and second regional isometric rainfall fields and topographic parameters, and set rainfall labels for the rainfall data fusion training samples using the second regional isometric rainfall field. The ground rain gauges are dispersed in some locations within the region. The model training module is used to construct a rainfall spatiotemporal fusion model based on convolutional neural networks and long short-term memory networks. The rainfall spatiotemporal fusion model is trained using rainfall data fusion training samples to obtain the rainfall spatiotemporal fusion target model. The rainfall spatiotemporal fusion model extracts regional precipitation spatial features based on convolutional neural networks, extracts regional precipitation temporal features based on long short-term memory networks, and uses a channel attention mechanism to learn and quantify the importance of regional precipitation features. The fusion output module is used to acquire topographic parameters of the area to be measured, microwave link precipitation observation data, and ground rain gauge precipitation observation data. Based on the microwave link precipitation observation data and the ground rain gauge precipitation observation data, it constructs the first isometric precipitation field and the second isometric precipitation field of the area to be measured, respectively. The topographic parameters of the area to be measured, the first isometric precipitation field, and the second isometric precipitation field are used as inputs to the spatiotemporal fusion target model of precipitation. The spatiotemporal fusion target model of precipitation is used to predict the precipitation field at each location in the area to be measured.

8. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 6.

Images