Satellite precipitation data fusion method based on unsupervised adaptive fusion network
By dynamically selecting and fusing satellite precipitation data through an unsupervised adaptive fusion network, the error problem of satellite precipitation products in areas without rain gauges is solved, generating high-quality precipitation products that support hydrological research and water resource management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2024-03-18
- Publication Date
- 2026-06-12
Smart Images
Figure CN118211177B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of precipitation estimation technology, and in particular to a satellite precipitation data fusion method based on an unsupervised adaptive fusion network. Background Technology
[0002] Precipitation data plays a crucial role in hydrological and atmospheric research, contributing to a comprehensive understanding of water balance and climate change. Traditional precipitation estimation relies on rain gauge observations, which provide the most reliable data source. However, the uneven and limited coverage and distribution of rain gauges result in significant data gaps in some areas. To overcome this limitation, satellite remote sensing has emerged as a promising alternative for estimating precipitation in areas without rain gauges, offering broad coverage and near real-time capabilities. Remote sensing technology utilizes sensors to measure parameters such as cloud properties, atmospheric humidity, and temperature, providing valuable information for precipitation estimation and assessing the impacts of precipitation on water resources, flood forecasting, and climate change. Utilizing remote sensing technology allows for a more comprehensive understanding of precipitation patterns, contributing to efficient water resource management and accurate hydrological modeling.
[0003] Recently, several precipitation products based on remote sensing satellites have been released, with two widely used categories: IMERG and SM2RAIN. The former is a satellite precipitation product obtained using a "top-down" approach, estimating precipitation by analyzing atmospheric signals emitted and scattered by the hydrological medium. SM2RAIN, on the other hand, is a satellite precipitation product obtained using a "bottom-up" approach. It estimates terrestrial precipitation using satellite soil moisture observation data, fully utilizing the relationship between precipitation events and soil moisture. However, due to limitations in satellite sensors, retrieval algorithms, and spatial sampling frequency, these satellite-based precipitation products suffer from systematic biases and random errors. These factors cause discrepancies between satellite precipitation forecasts and rain gauge observations, posing a challenge to accurate precipitation event prediction. Therefore, it is necessary to overcome the limitations and errors of satellite precipitation products to ensure their reliable application in hydrological research. Recent studies have explored various methods for combining satellite precipitation and station precipitation data, including error analysis and machine learning. However, the lack of nonlinear capabilities limits the accuracy of these methods. The complex relationship between rain gauge observation data and satellite remote sensing data poses a challenge to accurately capturing precipitation patterns. Therefore, more robust precipitation estimation techniques are needed. Deep learning has emerged as a promising approach for precipitation estimation. Convolutional Neural Networks (CNNs) based on deep learning have demonstrated exceptional ability to capture the inherent patterns and relationships in precipitation data. Specifically, CNNs can effectively capture the spatial dependencies and correlations between rain gauge observation data and remote sensing data, while improving precipitation estimation by integrating information from multiple sources. Furthermore, more complex network architectures, such as multi-layer 3D-CNNs and ConvLSTMs, have been proposed to leverage spatiotemporal patterns of precipitation and enhance precipitation estimation in sparsely monitored areas. To further improve precipitation estimation in areas without rain gauge observations, a deep learning fusion algorithm based on an attention mechanism has been proposed. These networks can automatically learn spatial features and precipitation distribution from the input data, thereby integrating precipitation information from different sources.
[0004] However, it is worth noting that deep learning-based methods typically require a large amount of labeled data (i.e., rain gauge observations) for training. But rain gauge observations are difficult to obtain in some regions. Therefore, exploring unsupervised precipitation fusion methods is crucial, as these methods can fuse multi-source precipitation data even without rain gauge observations, thus addressing this challenge. Furthermore, developing adaptive fusion methods that can effectively fuse precipitation data from different sources is also significant for improving the accuracy of precipitation estimation. Summary of the Invention
[0005] This invention provides a satellite precipitation data fusion method based on an unsupervised adaptive fusion network to solve the technical problems mentioned in the background art.
[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:
[0007] This invention provides a satellite precipitation data fusion method based on an unsupervised adaptive fusion network, which specifically includes the following steps:
[0008] S1. Collect satellite precipitation data IMERG-Early and SM2RAIN, and construct an adaptive fusion network. The adaptive fusion network includes two first enhanced feature extraction modules, an adaptive fusion module, and a second enhanced feature extraction module.
[0009] S2. The satellite precipitation data IMERG-Early and SM2RAIN are processed through two first-level enhanced feature extraction modules to extract initial features, resulting in initial feature maps. and ;
[0010] S3, Initial feature map and The data is input into the adaptive fusion module, which dynamically selects the feature with the most information from different precipitation data to obtain the second fused feature. Then the second fusion feature The data is input into the second enhanced feature extraction module to obtain the fused precipitation product. ;
[0011] S4, based on integrated precipitation products The original satellite precipitation data was used as a reference, and an adaptive fusion network was trained using an unsupervised optimization method. The parameters of the adaptive fusion network were then optimized and updated to obtain the trained adaptive fusion network.
[0012] S5. Apply the trained adaptive fusion network to precipitation estimation.
[0013] Furthermore, the adaptive fusion network in S1 also includes two residuals;
[0014] The two first enhanced feature extraction modules are fused with the two residuals respectively and connected to the input of the adaptive fusion module. The output of the adaptive fusion module is connected to the second enhanced feature extraction module.
[0015] The first enhanced feature extraction module consists of three convolutional layers and three ReLU activation functions.
[0016] Furthermore, S2 is expressed by a formula, as follows:
[0017]
[0018] in, and These represent the input precipitation products, namely satellite precipitation data IMERG-Early and SM2RAIN; and These represent the ReLU activation function and the 3 × 3 convolutional layer, respectively.
[0019] Furthermore, step S3 specifically includes the following steps:
[0020] S31. Initial feature map and The data is input into the adaptive fusion module, which dynamically fuses different precipitation features from two parallel branches: features from satellite precipitation data IMERG-Early and SM2RAIN. It then uses element-wise summation to fuse these different precipitation features, resulting in the first fused feature. ;
[0021] S32, in the first fusion feature Channel statistics are generated using the global average pooling operator within the spatial dimension. ;
[0022] S33, Channel Statistics A new vector is obtained after the convolution operation. z Applying the sigmoid function to the new vector z yields the first attention activation value. Then subtract the first attention activation value from 1. Obtain the second attention activation value ;
[0023] S34. Utilizing the first attention activation value Second attention activation value For the initial feature map and Recalibration and aggregation yield the second fusion feature. ;
[0024] S35, Incorporating the second fusion feature The data is input into the second enhanced feature extraction module to obtain the fused precipitation product. .
[0025] Furthermore, the first fusion feature in S31 The formula is as follows:
[0026] .
[0027] Furthermore, the channel statistics in S32 The specific calculation formula is as follows:
[0028]
[0029] in, Represents the space pooling function; H , W These are the length and width of the feature map, respectively; This represents the c-th channel of the first fusion feature; i , j They represent H , W One of the pixels.
[0030] Furthermore, the first attention activation value in S33 Second attention activation value The formulas are as follows:
[0031]
[0032] in, This represents the sigmoid function, used to obtain the channel attention map.
[0033] Furthermore, the second fusion feature in S34 The formulas are as follows:
[0034] .
[0035] Furthermore, step S4 specifically includes the following steps:
[0036] S41. Constructing integrated precipitation products The total loss function includes both input precipitation products and the merged precipitation product. Structural similarity measurement loss and Functional loss;
[0037] S42. Repeat S2 to S4, adjusting the weights of the adaptive fusion network until the total loss function converges or falls below the set threshold; thus obtaining the trained adaptive fusion network.
[0038] Furthermore, the total loss in S41 is expressed by a formula, as follows:
[0039]
[0040] in, and These represent the total loss of satellite precipitation data IMERG-Early and SM2RAIN, respectively. and These represent the structural similarity measurement losses of satellite precipitation data IMERG-Early and SM2RAIN, respectively. and These represent the satellite precipitation data IMERG-Early and SM2RAIN, respectively. Functional loss;
[0041] and The calculation formulas are as follows:
[0042] .
[0043] The beneficial effects of this invention are:
[0044] 1. This invention designs an adaptive fusion module for selecting information features;
[0045] 2. This invention performs unsupervised optimization without observation data, which is of great significance for areas without ground stations;
[0046] 3. This invention also generates a set of high-quality precipitation products, which are of great significance for hydrological research. Attached Figure Description
[0047] Figure 1 This forms the overall framework of the present invention;
[0048] Figure 2 This is the overall framework for the adaptive fusion module. Detailed Implementation
[0049] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many other different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.
[0050] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0052] Reference Figure 1 This application provides a satellite precipitation data fusion method based on an unsupervised adaptive fusion network, which specifically includes the following steps:
[0053] S1. Collect satellite precipitation data IMERG-Early and SM2RAIN, and construct an adaptive fusion network. The adaptive fusion network includes two first enhanced feature extraction modules, an adaptive fusion module, and a second enhanced feature extraction module.
[0054] S2. The satellite precipitation data IMERG-Early and SM2RAIN are processed through two first-level enhanced feature extraction modules to extract initial features, resulting in initial feature maps. and ;
[0055] S3, Initial feature map and The data is input into the adaptive fusion module, which dynamically selects the feature with the most information from different precipitation data to obtain the second fused feature. Then the second fusion feature The data is input into the second enhanced feature extraction module to obtain the fused precipitation product. ;
[0056] S4, based on integrated precipitation products The original satellite precipitation data was used as a reference, and an adaptive fusion network was trained using an unsupervised optimization method. The parameters of the adaptive fusion network were then optimized and updated to obtain the trained adaptive fusion network.
[0057] S5. Apply the trained adaptive fusion network to precipitation estimation.
[0058] In some embodiments, the adaptive fusion network in S1 further includes two residuals;
[0059] The two first enhanced feature extraction modules are fused with the two residuals respectively and connected to the input of the adaptive fusion module. The output of the adaptive fusion module is connected to the second enhanced feature extraction module.
[0060] The first enhanced feature extraction module consists of a convolutional neural network (CNN) and three ReLU activation functions. The CNN comprises three stacked convolutional layers, each consisting of a learnable filter. These filters are convolved with the input data to extract local features and generate a feature map. By incorporating a CNN into this invention, the spatial dependencies of precipitation products can be extracted. In this case, the CNN can automatically learn the spatial dependency between the target precipitation value and its surrounding precipitation area through multiple convolutional operations.
[0061] In some embodiments, S2 is expressed by a formula, as follows:
[0062]
[0063] in, and These represent the input precipitation products, namely satellite precipitation data IMERG-Early and SM2RAIN; and These represent the ReLU activation function and the 3 × 3 convolutional layer, respectively.
[0064] In the initial feature extraction, the local feature extraction capability of convolution plays a crucial role in capturing local precipitation changes. This invention learns the spatial representation of precipitation data through a first enhanced feature extraction module, thereby providing initial features for subsequent fusion.
[0065] To effectively integrate the unique characteristics of different precipitation products, this invention designs an adaptive fusion module. This module is as follows: Figure 2 As shown, the algorithm consists of two operators: a fusion operation and a selection operation. The fusion operation combines information from different precipitation products to generate a global feature descriptor. Then, the selection operation uses these descriptors to recalibrate the feature maps from different branches and summarizes them. The purpose of these two operations is to adaptively combine and recalibrate the feature maps from different branches.
[0066] The following is a detailed description of the steps in S3 (i.e., the fusion operation and the selection operation). Specifically, S3 includes the following steps, where S31 to S32 are the fusion operation and S33 to S35 are the selection operation.
[0067] S31. Initial feature map and The data is input into the adaptive fusion module, which dynamically fuses different precipitation features from two parallel branches: features from satellite precipitation data IMERG-Early and SM2RAIN. It then uses element-wise summation to fuse these different precipitation features, resulting in the first fused feature. , ;
[0068] S32, in the first fusion feature Channel statistics are generated using the Global Average Pooling (GAP) operator within the spatial dimension. This channel statistics aggregates global spatial information;
[0069] S33, Channel Statistics A new vector is obtained after the convolution operation. z Applying the sigmoid function to the new vector z yields the first attention activation value. Then subtract the first attention activation value from 1. Obtain the second attention activation value ;
[0070] S34. Utilizing the first attention activation value Second attention activation value For the initial feature map and Recalibration and aggregation yield the second fusion feature. ;
[0071] S35, Incorporating the second fusion feature The data is input into the second enhanced feature extraction module to obtain the fused precipitation product. By employing this adaptive fusion module, effective fusion of precipitation products can be achieved. Furthermore, it retains the most critical information from the original products to the greatest extent possible.
[0072] In some embodiments, the first fusion feature in S31 The formula is as follows:
[0073] .
[0074] In some embodiments, channel statistics in S32 The specific calculation formula is as follows:
[0075]
[0076] in, Represents the space pooling function; H , W These are the length and width of the feature map, respectively; This represents the c-th channel of the first fusion feature; i , j They represent H , W One of the pixels.
[0077] In some embodiments, the first attention activation value in S33 Second attention activation value The formulas are as follows:
[0078]
[0079] in, This represents the sigmoid function, used to obtain the channel attention map.
[0080] In some embodiments, the second fusion feature in S34 The formulas are as follows:
[0081] .
[0082] This invention proposes an unsupervised optimization method to train an adaptive fusion network, i.e., it does not rely on rain gauge observation data. It measures the differences between the fused product and the original product separately, enabling the adaptive fusion network to optimize based on the input data. Using this unsupervised optimization method, the adaptive fusion network can effectively capture the inherent characteristics and precipitation patterns present in the original product, thereby improving precipitation estimation. The unsupervised optimization method is specifically shown in S4, which includes the following steps:
[0083] S41. Constructing integrated precipitation products The total loss function includes both input precipitation products and the merged precipitation product. Structural similarity measurement loss and Functional loss; these loss functions are designed to capture merged precipitation products The structural and content differences between the original and fused precipitation products are considered. By using these loss functions, the adaptive fusion network can better learn the complex characteristics of precipitation products and generate accurate fused precipitation outputs.
[0084] S42. Repeat S2 to S4, adjusting the weights of the adaptive fusion network until the total loss function converges or falls below the set threshold; thus obtaining the trained adaptive fusion network.
[0085] In some embodiments, the total loss in S41 is expressed by a formula, as follows:
[0086]
[0087] in, and These represent the total loss of satellite precipitation data IMERG-Early and SM2RAIN, respectively. and These represent the structural similarity measurement losses of satellite precipitation data IMERG-Early and SM2RAIN, respectively. and These represent the satellite precipitation data IMERG-Early and SM2RAIN, respectively. Functional loss;
[0088] and The calculation formulas are as follows:
[0089] .
[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A satellite precipitation data fusion method based on an unsupervised adaptive fusion network, characterized in that, Specifically, the steps include the following: S1. Collect satellite precipitation data IMERG-Early and SM2RAIN, and construct an adaptive fusion network. The adaptive fusion network includes two first enhanced feature extraction modules, an adaptive fusion module, and a second enhanced feature extraction module. S2. The satellite precipitation data IMERG-Early and SM2RAIN are processed through two first-level enhanced feature extraction modules to extract initial features, resulting in initial feature maps. and ; S3, Initial feature map and The data is input into the adaptive fusion module, which dynamically selects the feature with the most information from different precipitation data to obtain the second fused feature. Then the second fusion feature The data is input into the second enhanced feature extraction module to obtain the fused precipitation product. ; S4, based on integrated precipitation products The original satellite precipitation data was used as a reference, and an adaptive fusion network was trained using an unsupervised optimization method. The parameters of the adaptive fusion network were then optimized and updated to obtain the trained adaptive fusion network. S5. Apply the trained adaptive fusion network to precipitation estimation. S3 specifically includes the following steps: S31. Initial feature map and The data is input into the adaptive fusion module, which dynamically fuses different precipitation features from two parallel branches: features from satellite precipitation data IMERG-Early and SM2RAIN. It then uses element-wise summation to fuse these different precipitation features, resulting in the first fused feature. ; S32, in the first fusion feature Channel statistics are generated using the global average pooling operator within the spatial dimension. ; S33, Channel Statistics A new vector is obtained after the convolution operation. z Applying the sigmoid function to the new vector z yields the first attention activation value. Then subtract the first attention activation value from 1. Obtain the second attention activation value ; S34. Utilizing the first attention activation value Second attention activation value For the initial feature map and Recalibration and aggregation yield the second fusion feature. ; S35, Incorporating the second fusion feature The data is input into the second enhanced feature extraction module to obtain the fused precipitation product. ; Channel statistics in S32 The specific calculation formula is as follows: in, Represents the space pooling function; H , W These are the length and width of the feature map, respectively; This represents the c-th channel of the first fusion feature; i , j They represent H , W One of the pixels.
2. The satellite precipitation data fusion method according to claim 1, characterized in that, The adaptive fusion network in S1 also includes two residuals; The two first enhanced feature extraction modules are fused with the two residuals respectively and connected to the input of the adaptive fusion module. The output of the adaptive fusion module is connected to the second enhanced feature extraction module. The first enhanced feature extraction module consists of three convolutional layers and three ReLU activation functions.
3. The satellite precipitation data fusion method according to claim 1, characterized in that, S2 is expressed by a formula, as follows: in, and These represent the input precipitation products, namely satellite precipitation data IMERG-Early and SM2RAIN; and These represent the ReLU activation function and the 3 × 3 convolutional layer, respectively.
4. The satellite precipitation data fusion method according to claim 3, characterized in that, The first fusion feature in S31 The formula is as follows: 。 5. The satellite precipitation data fusion method according to claim 4, characterized in that, The first attention activation value in S33 Second attention activation value The formulas are as follows: in, This represents the sigmoid function, used to obtain the channel attention map.
6. The satellite precipitation data fusion method according to claim 4, characterized in that, The second fusion feature in S34 The formulas are as follows: 。 7. The satellite precipitation data fusion method according to claim 6, characterized in that, S4 specifically includes the following steps: S41. Constructing integrated precipitation products The total loss function includes both input precipitation products and the merged precipitation product. Structural similarity measurement loss and Functional loss; S42. Repeat S2 to S4, adjusting the weights of the adaptive fusion network until the total loss function converges or falls below the set threshold; thus obtaining the trained adaptive fusion network.
8. The satellite precipitation data fusion method according to claim 7, characterized in that, The total loss in S41 is expressed by a formula, as follows: in, and These represent the total loss of satellite precipitation data IMERG-Early and SM2RAIN, respectively. and These represent the structural similarity measurement losses of satellite precipitation data IMERG-Early and SM2RAIN, respectively. and These represent the satellite precipitation data IMERG-Early and SM2RAIN, respectively. Functional loss; and The calculation formulas are as follows: 。