Multi-band multi-source collaborative atmospheric precipitable water vapor retrieval method based on deep neural network, medium and equipment

Abstract

The application provides a multi-band multi-source collaborative atmospheric precipitable water inversion method based on a deep neural network, a medium and equipment, and belongs to the technical field of atmospheric remote sensing, and comprises the following steps: collecting multi-source satellite observation features, performing spatiotemporal collaborative matching on the multi-source satellite observation features; constructing an original data set composed of the multi-source satellite observation features and measured atmospheric precipitable water after spatiotemporal alignment; constructing a U-Net deep neural network; preprocessing the multi-source satellite observation features and the atmospheric precipitable water in the original data set, and dividing the original data set into a training set, a verification set and a test set; training the U-Net deep neural network by using the training set, iteratively optimizing network parameters on the training set, monitoring performance indicators on the verification set, and comparing and displaying the final performance of the network on the test set; and predicting the atmospheric precipitable water by using the U-Net deep neural network obtained after training. The application realizes high-precision, high-spatial-resolution and reliable atmospheric precipitable water inversion under all-weather conditions.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric remote sensing technology, specifically relating to a method, medium, and equipment for inverting atmospheric precipitable water based on a deep neural network and multi-band, multi-source collaborative method. Background Technology

[0002] Precipitable water vapor (PWV) is a key parameter characterizing the total amount of water vapor per unit area column, playing a central role in global energy balance, water cycle, and climate evolution. In astronomy, water vapor is a major source of signal attenuation and noise in the microwave and infrared bands; therefore, high-precision PWV retrieval is equally crucial for high-precision astronomical observations such as millimeter-wave interferometric imaging and terahertz spectral line detection. Currently, methods for acquiring PWV data mainly encompass a variety of technologies, from ground-based detection to satellite remote sensing.

[0003] Ground-based detection methods mainly include radiosonde observation, global navigation satellite system (GNSS) remote sensing, and ground-based microwave radiometers. Radiosonde observation, by releasing weather balloons to directly measure parameters in various atmospheric layers, is the most authoritative method for obtaining vertical water vapor profiles; however, its observation costs are high and the distribution of stations is extremely uneven. In contrast, atmospheric precipitable water retrieved based on tropospheric delay information from GNSS signals has the advantages of high accuracy and high temporal resolution, serving as a benchmark for verifying the accuracy of other retrieval results. Ground-based microwave radiometers utilize the absorption characteristics of the atmosphere in specific microwave frequency bands, passively receiving downlink thermal radiation brightness temperature to achieve near-real-time retrieval of total water vapor column and lower atmospheric vertical profiles. However, a common limitation of ground-based detection methods is that the effective measurement range is limited to the area near the antenna or station, making it difficult to achieve large-area continuous spatial coverage.

[0004] Satellite remote sensing measures atmospheric water vapor absorption mechanisms in different bands, including near-infrared, thermal infrared, and passive microwave, and is the primary means of obtaining large-scale, continuously covered atmospheric precipitable water products. While near-infrared retrieval can provide high spatial resolution products at the kilometer or even meter level, it is heavily reliant on solar radiation reflected from the Earth's surface and is limited to operation under clear daytime conditions. Thermal infrared retrieval, although capable of day and night observation, still cannot penetrate clouds and is highly susceptible to interference from ground emissivity and atmospheric profiles. Microwave retrieval, while capable of penetrating clouds for all-weather retrieval, is limited by the diffraction limit of physical antennas, typically resulting in spatial resolution only in the tens of kilometers range, and significant errors can occur against complex terrain. Furthermore, traditional atmospheric precipitable water retrieval in different bands heavily relies on fixed physical or numerical retrieval models built based on prior knowledge. These models usually require iterative solutions using complex radiative transfer equations. Differences in surface type and real-time changes in atmospheric temperature profiles can lead to insufficient overall accuracy and stability of microwave physical retrieval algorithms.

[0005] In summary, existing single-band observation data inversion methods have significant limitations in four dimensions: spatial resolution, observation frequency, all-weather accuracy, and, most importantly, the accuracy of the inversion model. How to construct a framework for atmospheric precipitable water inversion that leverages the advantages of wide-area satellite coverage, achieves high-precision ground-based benchmarks, and provides reliable, high-spatial-resolution data under all weather conditions is a pressing issue in the field of atmospheric remote sensing. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by providing a multi-band, multi-source collaborative atmospheric precipitable water inversion method, medium, and equipment based on deep neural networks. This method eliminates the need for complex prior physical inversion models or pre-defined parameterization schemes found in traditional inversion methods. It directly combines the multi-dimensional feature extraction capabilities of deep neural networks with multi-band, multi-source raw satellite payload observation data. After inputting multi-band observation features into the deep neural network, high-precision ground-based global navigation satellite system (GNSS) measured values ​​are used as the objective function for constraint. Backpropagation optimizes the objective function calculated from the predicted results and the ground reference, enabling the network to directly learn the nonlinear mapping from the raw payload observation data space to the high-resolution atmospheric precipitable water data space. This avoids the accuracy loss associated with traditional inversion processes based on prior physical models. Ultimately, this invention achieves high-precision, high spatial resolution, and reliable atmospheric precipitable water inversion under all weather conditions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for retrieving atmospheric precipitable water based on a deep neural network with multiple bands and multiple sources, comprising the following steps: Step 1: Collect multi-source satellite observation features, and use the measured atmospheric precipitable water record as the spatiotemporal anchor point to perform spatiotemporal collaborative matching of multi-source satellite observation features; Step 2: Construct the original dataset consisting of spatiotemporally aligned multi-source satellite observation features and measured atmospheric precipitable water. Step 3: Construct the U-Net deep neural network and its target loss function. The network input is multi-source satellite observation features, and the output is atmospheric precipitable water. Step 4: Preprocess the multi-source satellite observation features and atmospheric precipitable water in the original dataset, and divide it into training set, validation set and test set based on the Global Navigation Satellite System sites; Step 5: Train the U-Net deep neural network using the training set, iteratively optimize the network parameters on the training set, monitor performance metrics on the validation set, and compare and demonstrate the final network performance on the test set. Step 6: Use the trained U-Net deep neural network to predict atmospheric precipitable water.

[0009] Optionally, in step 1, the multi-source satellite observation features include microwave brightness temperature, infrared and visible light radiation brightness temperature, and reflectivity; in the original dataset, each row of samples represents the mapping between multi-source satellite observation features and atmospheric precipitable water at a specific spatiotemporal location.

[0010] Optionally, in step 2, the spatiotemporal collaborative matching specifically includes: In the time dimension, the time difference between microwave brightness temperature and measured atmospheric precipitable water is limited to a ±5 minute window, and this is used to coordinate the matching of infrared and visible light radiation brightness temperature and reflectivity within a ±5 minute tolerance. In the spatial dimension, a nearest neighbor mapping strategy is adopted, with a distance threshold of 0.1°, to extract multi-source satellite observation features at the nearest pixel to the measured atmospheric precipitable water location.

[0011] Optionally, in step 3, the U-Net deep neural network uses a symmetrical encoder-decoder structure; the network input is... Tensors of dimension, filled by setting 0 to 1 Dimension; in the encoder section, The dimensional tensor is passed through a convolutional block and downsampled, and this operation is repeated 4 times to finally transform it into a bottleneck layer. In the decoder section, the bottleneck layer's tensor is upsampled and then concatenated with the tensor of the corresponding level's encoder convolutional block. This process continues to extract features through the convolutional block, repeating the process four times. Finally, the atmospheric precipitable water content is calculated in the output layer through a fully connected network.

[0012] Optionally, in step 3, the target loss function for:

[0013] in, and These are atmospheric precipitable water values ​​predicted by the network and measured by global navigation satellite system stations, respectively. This represents the total number of samples currently participating in the calculation.

[0014] Optionally, in step 4, the preprocessing includes: performing normalization on the multi-source satellite observation features in the original dataset; and processing the atmospheric precipitable water... conduct The numerical adjustment, This indicates the adjusted atmospheric precipitable water content, and then... Normalization is performed.

[0015] Optionally, in step 4, the training set, validation set, and test set are divided in a ratio of 8:1:1.

[0016] In a second aspect, the present invention provides a computer-readable storage medium storing a computer program that causes a computer to execute the multi-band, multi-source, collaborative atmospheric precipitable water inversion method based on deep neural networks as described in the first aspect.

[0017] Thirdly, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the multi-band multi-source collaborative atmospheric precipitable water inversion method based on deep neural networks as described in the first aspect.

[0018] The beneficial effects of this invention are: (1) This invention directly establishes a nonlinear mapping from satellite payload characteristics to atmospheric precipitable water through deep learning, avoiding strong dependence on physical radiative transfer models and real-time atmospheric profile data. This greatly reduces the complexity of the inversion model and significantly improves the inversion accuracy and robustness of the model in different geographical regions and complex meteorological environments.

[0019] (2) This invention utilizes a specially designed deep neural network structure to achieve direct alignment between the inversion result and the target high-resolution grid. Compared with traditional methods that require "low-resolution inversion and spatial interpolation", this architecture fundamentally avoids the accuracy loss caused by spatial resampling through efficient one-time forward computation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the training and application of the U-Net deep neural network.

[0021] Figure 2This is a schematic diagram showing the distribution of NGL GNSS stations in Australia.

[0022] Figure 3 This is the architecture diagram of the U-Net deep neural network.

[0023] Figure 4 It shows the curves of the objective function and RMSE as a function of the number of iterations.

[0024] Figure 5 It is a comparison between model predictions and actual PWV values.

[0025] Figure 6 It is a comparison of the manifold structure of the test set data. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0027] In one embodiment, this invention proposes a multi-band, multi-source collaborative atmospheric precipitable water inversion method based on deep neural networks. This method specifically includes: constructing a deep learning model with multi-band, multi-source satellite raw observation data as input variables and high-precision, high spatial resolution, and reliable atmospheric precipitable water under all weather conditions as output variables. An objective function is constructed by calculating the mean square error between the deep neural network prediction results and the true values ​​of measured atmospheric precipitable water products, directly driving the iterative optimization of the deep neural network without relying on complex prior physical models. When the validation set accuracy reaches a preset training step size or a set accuracy threshold, the model weights at this point are saved for subsequent end-to-end inversion of atmospheric precipitable water at any geographical location. This method significantly improves the precision of atmospheric water vapor detection and achieves all-weather, cross-scale inversion.

[0028] The basic principle of the multi-band, multi-source data collaborative atmospheric precipitable water inversion method based on deep neural networks is as follows: High-precision atmospheric precipitable water data is used as a physical constraint benchmark to drive a deep neural network to establish a nonlinear mapping from the multi-band, multi-source remote sensing data space to the atmospheric precipitable water data space. Specifically, this method constructs a dataset containing multi-band and corresponding location-specific true atmospheric precipitable water values ​​by spatiotemporally synchronizing microwave detection data with infrared and visible light observation data, which have high spatial resolution. Based on this, multi-band, multi-source observation parameters are fed as input feature vectors into a defined deep neural network. First, the predicted atmospheric precipitable water value is obtained through the forward propagation process of the deep neural network, and the objective function (such as mean square error) between this prediction result and the true atmospheric precipitable water value used as the monitoring target is calculated. Subsequently, based on this objective function, the network weights are iteratively optimized using backpropagation and gradient descent. By iteratively executing the above "forward-backward" process, the model weights are continuously updated until final convergence, ultimately establishing a mapping from multi-band, multi-source remote sensing data space to atmospheric precipitable water data space, such as... Figure 1 As shown.

[0029] Taking the total precipitable atmospheric water in Australia as an example, the inversion process is as follows: (1) This embodiment uses the measured atmospheric precipitable water record as the spatiotemporal anchor point to perform strict spatiotemporal collaborative matching of multi-source heterogeneous satellite data, achieving effective alignment of multi-source observation data such as infrared, microwave, and visible light. The specific matching method is as follows: In the time dimension, the time difference between the passive microwave brightness temperature and the measured atmospheric precipitable water is strictly limited to a window of ±5 minutes, and the infrared and visible light radiation brightness temperature and reflectivity within the ±5 minute tolerance are matched accordingly; In the spatial dimension, the nearest neighbor mapping strategy is adopted, with a distance threshold of 0.1°, to extract the above-mentioned multi-source satellite observation features at the pixel closest to the measured atmospheric precipitable water location. After removing invalid samples, the final dataset is constructed. This dataset accurately correlates satellite observation features with atmospheric precipitable water in spatiotemporal terms, laying a standardized data-driven foundation for neural network learning of the nonlinear mapping from multi-band, multi-source remote sensing feature space to atmospheric precipitable water space.

[0030] (2) Construct an original dataset consisting of satellite observation features (microwave brightness temperature, infrared and visible light radiation brightness temperature and reflectivity) and high-precision measured atmospheric precipitable water in a spatiotemporal alignment. In the dataset, each row of samples represents the mapping between multi-source, multi-platform satellite observation features and atmospheric precipitable water at a specific spatiotemporal location. The selection of input features focuses on the complementarity of multi-band radiation characteristics. The satellite observation features selected in this embodiment specifically include passive microwave brightness temperature (vertical and horizontal polarization brightness temperatures of 18, 23, 36, and 89 GHz from the AMSR2 payload) which has all-weather penetration capability and can invert water vapor information under cloud masks, and infrared and visible light radiation brightness temperature and reflectivity (bright temperature, reflectivity, and satellite elevation azimuth angle from the AHI payload, passband numbers 1 to 16) which provide high spatial resolution texture and capture local subtle water vapor changes. Detailed descriptions of each feature are shown in Table 1. The supervision labels uniformly adopt the measured atmospheric precipitable water at the Global Navigation Satellite System sites provided by the Nevada Geodesy Laboratory (e.g., Figure 2 (As shown).

[0031] Table 1 Satellite Data

[0032] (3) In order to extract complex nonlinear manifold mappings from high-dimensional features fused from multiple sources, this embodiment adopts an improved U-Net deep neural network architecture (e.g., Figure 3 (As shown). This network improves upon the traditional U-Net deep neural network for processing two-dimensional image information. In terms of overall topology, the network uses a symmetrical encoder-decoder structure. The basic unit of the network is a convolutional block containing two one-dimensional convolutions; the specific dimensional changes of the tensor in each convolutional layer are labeled... Figure 3 Network input is Tensors of dimension, filled by setting 0 to 1 Dimensions. In the encoder section, The dimensional tensor is passed through a convolutional block and downsampled, and this operation is repeated 4 times to finally transform it into a bottleneck layer. In the decoder section, the bottleneck layer's tensor is upsampled and then concatenated with the tensor of the corresponding level's encoder convolutional block. This process continues through the convolutional block to extract features, repeating four times. Finally, the output layer passes through a fully connected network to calculate the atmospheric precipitable water content.

[0033] It is worth noting that, unlike the traditional U-Net, this model not only adjusts the two-dimensional convolution to one-dimensional convolution to meet the actual data dimensionality requirements of this embodiment, but also adopts the same padding strategy in all convolutional blocks. This operation ensures that the tensor dimension remains unchanged before and after the convolution calculation, avoiding the problem in the traditional U-Net deep neural network where the tensor dimension is reduced due to the lack of padding, which requires pruning features during skip connections. This maximizes the integrity of multi-scale feature fusion.

[0034] This architecture fully leverages the multi-scale feature extraction capabilities of deep convolutional networks and the advantages of skip connections. The encoder path captures deep physical information of atmospheric features through layer-by-layer convolution; skip connections directly transfer the high-resolution spatial details preserved in the encoding layer to the decoding layer, effectively compensating for the spatial scale limitations of microwave observations. The output layer undergoes dimensionality transformation to ensure that the resolution of the inversion result is directly aligned with the input feature map or the preset target resolution, thereby achieving end-to-end direct physical mapping.

[0035] (4) Before entering network training, the original features are preprocessed to optimize the model convergence performance. First, the input feature vector is normalized, and the specific mapping rule is as follows:

[0036] in These are the initial values ​​for the corresponding features. , These are the maximum and minimum values ​​of the corresponding feature.

[0037] This operation aims to eliminate the impact of differences in the dimensions of different satellite payloads on gradient updates. Secondly, the target label (atmospheric precipitable water) is... The numerical adjustment and specific mapping rules are as follows:

[0038] This mapping optimizes the shape of the target label distribution (approximately adjusting the long-tailed distribution to a Gaussian distribution), suppresses the skewed distribution characteristics of atmospheric precipitable water, and reduces the variance of large water vapor distributions under extreme weather conditions, thus enabling the model to maintain stable inversion accuracy across different water vapor levels. Simultaneously, normalization is also performed on this adjusted target label, with the specific mapping rules as follows:

[0039] in This is the initial value for atmospheric precipitable water. These are the normalized label values. , These are the maximum and minimum values ​​of precipitable water after logarithmic transformation.

[0040] (5) For the continuous variable regression task of atmospheric precipitable water inversion, this embodiment selects the mean square error as the target loss function and directly calculates the difference between the network predicted value and the ground measured benchmark. Norm error. Its mathematical expression is as follows:

[0041] in , To correspond to the predicted atmospheric precipitable water volume from the horizontal network and the measured atmospheric precipitable water volume from the Global Navigation Satellite System (GNSS) stations, This represents the total number of samples currently participating in the calculation.

[0042] Using this loss function, the gradient is calculated through the backpropagation algorithm and the network weights are iteratively optimized to make the prediction results approximate the actual measured values ​​of the global navigation satellite system.

[0043] (6) To ensure the model has good spatial generalization ability, this embodiment adopts a site-based partitioning strategy. All global navigation satellite system sites are divided into training set, validation set and test set in a ratio of 8:1:1. The network parameters are iteratively optimized on the training set, the performance indicators are monitored on the validation set to prevent the model from overfitting to the features of specific geographical locations of the stations, and the final model performance is compared and displayed on the test set.

[0044] (7) During model training, when the validation set performance reaches the preset training step size or triggers the early stopping threshold, the iteration is stopped (e.g., Figure 4 (As shown) and save the current optimal network structure and weight parameters. Then, forward propagation is performed on the test set, and the model's prediction accuracy is checked by calculating evaluation metrics such as mean squared error and correlation coefficient. A scatter plot of the predicted and measured values ​​is then plotted (e.g., ...). Figure 5 (As shown). The validated model can be directly deployed in the real-time inversion process. By inputting raw satellite observation data and geographic location information, it can instantly output high-resolution, all-weather atmospheric precipitable water products.

[0045] (8) Analysis of the trained model on the test set data revealed that, compared to an untrained deep neural network and a neural network trained on random noise data, the deep neural network in this embodiment successfully extracted high-dimensional manifold information (such as...) from the atmospheric precipitable water data. Figure 6 (As shown).

[0046] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program that enables a computer to execute the multi-band, multi-source, collaborative atmospheric precipitable water inversion method based on deep neural networks described in the foregoing embodiments.

[0047] In another embodiment, the present invention proposes an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the multi-band, multi-source collaborative atmospheric precipitable water inversion method based on deep neural networks as described in the foregoing embodiment.

[0048] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0049] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. 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 implementation should not be considered beyond the scope of this application.

[0050] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for inverting atmospheric precipitable water based on a deep neural network using multi-band, multi-source collaborative methods, characterized in that... Includes the following steps: Step 1: Collect multi-source satellite observation features, and use the measured atmospheric precipitable water record as the spatiotemporal anchor point to perform spatiotemporal collaborative matching of multi-source satellite observation features; Step 2: Construct the original dataset consisting of spatiotemporally aligned multi-source satellite observation features and measured atmospheric precipitable water. Step 3: Construct the U-Net deep neural network and its target loss function. The network input is multi-source satellite observation features, and the output is atmospheric precipitable water. Step 4: Preprocess the multi-source satellite observation features and atmospheric precipitable water in the original dataset, and divide it into training set, validation set and test set based on the Global Navigation Satellite System sites; Step 5: Train the U-Net deep neural network using the training set, iteratively optimize the network parameters on the training set, monitor performance metrics on the validation set, and compare and demonstrate the final network performance on the test set. Step 6: Use the trained U-Net deep neural network to predict atmospheric precipitable water.

2. The method for inverting atmospheric precipitable water based on a deep neural network using a multi-band, multi-source collaborative approach as described in claim 1, characterized in that: In step 1, the multi-source satellite observation features include microwave brightness temperature, infrared and visible light radiation brightness temperature, and reflectivity; in the original dataset, each row of samples represents the mapping between multi-source satellite observation features and atmospheric precipitable water at a specific spatiotemporal location.

3. The method for inverting atmospheric precipitable water based on a deep neural network using a multi-band, multi-source collaborative approach, as described in claim 2, is characterized in that: In step 2, the spatiotemporal collaborative matching specifically involves: In the time dimension, the time difference between microwave brightness temperature and measured atmospheric precipitable water is limited to a ±5 minute window, and this is used to coordinate the matching of infrared and visible light radiation brightness temperature and reflectivity within a ±5 minute tolerance. In the spatial dimension, a nearest neighbor mapping strategy is adopted, with a distance threshold of 0.1°, to extract multi-source satellite observation features at the nearest pixel to the measured atmospheric precipitable water location.

4. The method for inverting atmospheric precipitable water based on a deep neural network using a multi-band, multi-source collaborative approach as described in claim 1, characterized in that: In step 3, the U-Net deep neural network uses a symmetrical encoder-decoder structure; the network input is... Tensors of dimension, filled by setting 0 to 1 Dimension; in the encoder section, The dimensional tensor is passed through a convolutional block and downsampled, and this operation is repeated four times to finally transform it into a bottleneck layer. In the decoder section, the bottleneck layer's tensor is upsampled and then concatenated with the tensor of the corresponding level's encoder convolutional block. This process continues to extract features through the convolutional block, repeating the process four times. Finally, the atmospheric precipitable water content is calculated in the output layer through a fully connected network.

5. The method for inverting atmospheric precipitable water based on a deep neural network using a multi-band, multi-source collaborative approach as described in claim 1, characterized in that: In step 3, the target loss function for: in, and These are atmospheric precipitable water values ​​predicted by the network and measured by global navigation satellite system stations, respectively. This represents the total number of samples currently participating in the calculation.

6. The method for inverting atmospheric precipitable water based on a deep neural network using a multi-band, multi-source collaborative approach as described in claim 1, characterized in that: In step 4, the preprocessing includes: performing normalization on the multi-source satellite observation features in the original dataset; and processing the atmospheric precipitable water... conduct The numerical adjustment, This indicates the adjusted atmospheric precipitable water content, and then... Normalization is performed.

7. The method for inverting atmospheric precipitable water based on a deep neural network using a multi-band, multi-source collaborative approach as described in claim 1, characterized in that: In step 4, the training set, validation set, and test set are divided in a ratio of 8:1:

1.

8. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the multi-band, multi-source, collaborative atmospheric precipitable water inversion method based on deep neural networks as described in any one of claims 1-7.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method for inverting atmospheric precipitable water based on a deep neural network as described in any one of claims 1-7.

Images