Atmospheric precipitable water quantity inversion method and system based on multi-source satellite remote sensing data

Abstract

The invention provides an atmospheric precipitable water quantity inversion method and system based on multi-source satellite remote sensing data, and relates to the technical field of remote sensing data. After clear sky pixels are screened, radiation brightness temperature data, digital elevation model data and spatio-temporal feature coding data are processed through three feature extraction branches, obtained feature representations are spliced and input into a feature fusion network, a spatial attention module is introduced in the feature fusion stage, finally, a composite loss function is adopted for training, and the feature fusion network is constructed. And high-precision atmospheric precipitable water quantity inversion is realized. According to the method, the multi-source data characteristics are fully utilized, and the inversion precision and the space continuity are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data technology, and in particular to a method and system for retrieving atmospheric precipitable water based on multi-source satellite remote sensing data. Background Technology

[0002] Depending on the spectral band, remote sensing inversion methods can be divided into near-infrared inversion, thermal infrared inversion, and microwave remote sensing inversion. Near-infrared inversion mainly utilizes the 0.94μm water vapor absorption band, with the MODIS near-infrared algorithm as a representative example; however, this method is only suitable for cloudless daytime conditions. Domestic scholars such as Hu Xiuqing et al. developed a weighted average algorithm based on FY-3A MERSI near-infrared data, finding that the 940nm channel performs better in dry atmospheres, while 905nm is more sensitive under humid conditions; however, the lookup table needs optimization to reduce system bias. Thermal infrared inversion is based on the difference in water vapor absorption between split-window channels (11μm and 12μm), such as the covariance-variance ratio algorithm proposed by Sobrino et al., but changes in surface emissivity limit accuracy, and this method is sensitive to the upper atmosphere; lower-level water vapor distribution needs to be combined with other data. Furthermore, thermal infrared remote sensing methods can achieve all-day observation of atmospheric precipitable water in land and ocean areas, but due to the inability of thermal infrared bands to penetrate clouds, it is only suitable for inversion under clear-sky conditions. In contrast, microwave remote sensing possesses all-weather detection capabilities, and its principle is to invert atmospheric parameters using microwave radiation signals. Because the emissivity of the sea surface is significantly lower than that of the land surface in the microwave band, it is easier to distinguish atmospheric and surface radiation information. Therefore, microwave remote sensing is mainly used for inverting precipitable water over the ocean. This study targets the land region of my country; therefore, the thermal infrared inversion method was chosen to invert PWV under clear-sky conditions.

[0003] In recent years, methods for inverting thermal infrared PWV (Power Width Volume) from geostationary meteorological satellites have mainly included regression analysis, physical iterative regression, and machine learning algorithms. Tan et al. used infrared channel brightness temperature data from FY-4A AGRI as input, combined with surface parameters provided by numerical weather prediction as the background field, to develop a physical inversion method. They found that FY-4A PWV products have advantages in spatiotemporal resolution, with an overall accuracy better than 4 mm, making them suitable for climate research, weather forecasting, and GNSS meteorology. However, under high water vapor conditions, FY-4A PWV is underestimated and requires calibration before use. Domestic scholars such as Li Wanbiao developed a physical iterative algorithm for FY-4A AGRI data, but it has significant errors in low water vapor regions such as the Qinghai-Tibet Plateau. In terms of deep learning applications, Liu et al. used CNN to invert PWV from GOES-16 ABI data, improving accuracy by about 20% compared to traditional methods. However, the CNN architecture is directly transferred from image recognition and needs optimization for the physical mechanism of water vapor inversion. Lee et al. compared the performance of random forests, XGBoost, and DNNs in PWV inversion, finding that DNNs performed best in complex nonlinear relationships. Zhou et al. constructed a high-precision and spatially resolution machine learning PWV inversion model using multiple fusion models, finding that CNNs, with their spatial feature extraction capabilities, were the optimal model, significantly improving the accuracy and detail of water vapor products; deep learning models outperformed traditional machine learning and linear models. However, research on FY-4B AGRI data is limited, especially regarding all-weather inversion in the Chinese region, which lacks systematic research. Summary of the Invention

[0004] This invention provides a method and system for retrieving atmospheric precipitable water based on multi-source satellite remote sensing data, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for retrieving atmospheric precipitable water based on multi-source satellite remote sensing data, comprising:

[0006] Meteorological data from a geostationary satellite multispectral imager is acquired and reprojected onto a unified spatial grid. Clear-sky pixels are then selected based on the cloud masking data of the meteorological data, and cloud-covered areas are removed to obtain valid input data.

[0007] The multi-band radiation brightness temperature data in the effective input data is input to the first feature extraction branch, and multi-scale feature extraction is performed through the encoder-decoder structure. The encoder downsamples step by step to extract features at different scales, and the decoder restores the spatial resolution by upsampling and fuses the features of each level of the encoder with the corresponding level of the decoder through skip connections to obtain the first feature representation.

[0008] The digital elevation model data of the meteorological data is input into the second feature extraction branch, and the modulation features of topography on water vapor distribution are extracted through a convolutional neural network to obtain the second feature representation;

[0009] The spatiotemporal feature encoding data of the meteorological data is input into the third feature extraction branch, and the dynamic modulation features of the time series on the water vapor distribution are extracted through a one-dimensional convolutional neural network to obtain the third feature representation;

[0010] The first feature representation, the second feature representation, and the third feature representation are concatenated and input into the feature fusion network. A spatial attention mechanism is introduced to map the concatenated features into a single-channel inversion result through dimensionality reduction using convolutional layers. The training process uses a composite loss function, which includes mean squared error loss, mean absolute error loss, and gradient consistency loss. The gradient consistency constraint can adjust for overfitting. The single-channel inversion result is output as an atmospheric precipitable water product.

[0011] The meteorological data includes at least one of the following: multi-band radiation brightness temperature data, cloud mask data, digital elevation model data, and atmospheric reanalysis gridded data.

[0012] Reprojecting the meteorological data onto a unified spatial grid includes:

[0013] The geographic coordinate information corresponding to each pixel in the radiation brightness temperature data is obtained by parsing the geographic location data; a geographic coordinate system of the target spatial grid is established, wherein the target spatial grid covers a predetermined geographic area and has a uniform spatial resolution;

[0014] Based on the geographic coordinate information, the original pixels of the radiation brightness temperature data are mapped to the corresponding grid points of the target spatial grid through coordinate transformation, thereby completing the spatial registration of the radiation brightness temperature data;

[0015] Spatial interpolation is performed on the atmospheric reanalysis grid data to resample the original spatial resolution grid data to a spatial resolution consistent with the target spatial grid, resulting in downscaled atmospheric reanalysis grid data;

[0016] A resampling operation is performed on the digital elevation model data to adjust it to a spatial resolution and geographic range consistent with the target spatial grid.

[0017] The radiation brightness temperature data, the downscaled atmospheric reanalysis grid data, and the resampled digital elevation model data are spatially aligned pixel-by-pixel on the target spatial grid to form a spatially registered multi-source dataset.

[0018] After obtaining the first feature representation, the method further includes:

[0019] A spatial attention mechanism is applied to the first feature representation, and a spatial weight mask is generated through convolution operation and weighted to enhance the feature response of the active convection region and the region with significant water vapor gradient, thus obtaining a weighted first feature representation.

[0020] Multi-scale feature extraction is performed using an encoder-decoder structure. The encoder downsamples at different scales to extract features, while the decoder upsamples to restore spatial resolution and fuses features from each encoder level with corresponding features from the decoder via skip connections.

[0021] Multiple downsampling levels are set in the encoder. Each downsampling level includes convolution and pooling operations. The convolution operation extracts the feature representation at the current scale, and the pooling operation reduces the spatial resolution of the feature map and increases the receptive field. After downsampling at each level, a multi-scale feature pyramid from coarse to fine is obtained.

[0022] In the decoder, upsampling levels are set up to correspond to the number of encoder levels. Each upsampling level restores the spatial resolution of the feature map through transpose convolution or interpolation operations.

[0023] At each upsampling level, the feature map recovered by the decoder at the current level is skipped and connected to the feature map at the corresponding spatial scale level in the encoder. The skip connection fuses the spatial detail information of the encoder feature map into the decoder feature map through feature concatenation operation, compensating for the high-frequency spatial information lost during the upsampling process.

[0024] The skip connections enable the decoder to retain the multi-scale spatial structure features extracted by the encoder while restoring spatial resolution, and output a first feature representation that includes coarse-grained semantic information and fine-grained spatial details.

[0025] The training process also includes:

[0026] An adaptive optimizer is used to iteratively update the parameters of the feature fusion network. The adaptive optimizer dynamically adjusts the learning step size by maintaining the first-order moment estimate and second-order moment estimate of each parameter, and introduces a weight decay mechanism to impose regularization constraints on the parameters to suppress overfitting.

[0027] A learning rate scheduling strategy is set up to warm up the model by gradually increasing the learning rate in the initial training phase, and then gradually decaying the learning rate to its peak value according to a predetermined pattern in the main training phase. This allows the model parameters to converge quickly in the early stage of training and be fine-tuned in the later stage.

[0028] During training, the combined loss function value and predefined performance evaluation index on the validation set are monitored in real time. Training is terminated when the validation set performance index does not improve within a preset number of consecutive rounds, and the network parameters with the best performance on the validation set are saved as the final model parameters.

[0029] A second aspect of the present invention provides an atmospheric precipitable water retrieval system based on multi-source satellite remote sensing data, comprising:

[0030] The first unit is used to acquire meteorological data from a geostationary satellite multispectral imager, reproject the meteorological data onto a unified spatial grid, and filter clear-sky pixels based on the cloud mask data of the meteorological data, removing cloud-covered areas to obtain valid input data.

[0031] The second unit is used to input the multi-band radiation brightness temperature data from the effective input data into the first feature extraction branch, and perform multi-scale feature extraction through an encoder-decoder structure. The encoder downsamples step by step to extract features at different scales, and the decoder restores the spatial resolution by upsampling and fuses the features of each level of the encoder with the corresponding level of the decoder through skip connections to obtain the first feature representation.

[0032] The third unit is used to input the digital elevation model data of the meteorological data into the second feature extraction branch, and extract the modulation features of topography on water vapor distribution through a convolutional neural network to obtain the second feature representation;

[0033] The fourth unit is used to input the spatiotemporal feature encoded data of the meteorological data into the third feature extraction branch, and extract the dynamic modulation features of spatiotemporal water vapor distribution through a one-dimensional convolutional neural network to obtain the third feature representation;

[0034] The fifth unit is used to concatenate the first feature representation, the second feature representation, and the third feature representation and input them into the feature fusion network. The feature fusion network maps the concatenated features into a single-channel inversion result through layer-by-layer dimensionality reduction. The training process uses a combined loss function composed of a mean square error term, a mean absolute error term, and a gradient smoothing term. The gradient smoothing term constrains the spatial continuity by calculating the gradient difference between the inversion result and the true value in the horizontal and vertical directions. The single-channel inversion result is output as a gridded product of atmospheric precipitable water.

[0035] A third aspect of the embodiments of the present invention,

[0036] An electronic device is provided, comprising:

[0037] processor;

[0038] Memory used to store processor-executable instructions;

[0039] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0040] Fourth aspect of the present invention,

[0041] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0042] The beneficial effects of this application are as follows:

[0043] Compared with traditional single-source data methods, this invention utilizes geostationary satellite multispectral imager data combined with digital elevation model data to establish a multi-source data fusion inversion framework, effectively capturing the spatiotemporal variation characteristics of atmospheric water vapor and the influence of topography on water vapor distribution, thereby improving inversion accuracy and reliability.

[0044] By employing an encoder-decoder structure for multi-scale feature extraction and combining it with a skip connection mechanism to fuse information from different scales, the model can simultaneously preserve the global structure and local details of water vapor distribution, thereby enhancing its ability to express water vapor features in different regions.

[0045] The innovative introduction of gradient smoothing constraint terms to construct a combined loss function effectively constrains the spatial continuity of the inversion results, solves the problem of abrupt changes in inversion results at regional boundaries in traditional methods, ensures the physical rationality of the product, and improves its application adaptability under complex terrain and meteorological conditions. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the atmospheric precipitable water inversion method based on multi-source satellite remote sensing data according to an embodiment of the present invention.

[0047] Figure 2 A graph showing the historical changes in model training loss, RMSE, and MAE;

[0048] Figure 3 A scatter plot of PWV inversion accuracy;

[0049] Figure 4 The histogram of PWV inversion error distribution;

[0050] Figure 5 This is a scatter plot of PWV inversion accuracy. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0053] Existing research largely focuses on FY-4A or foreign satellite data, and the application value of FY-4B's newly added channels in water vapor retrieval has not yet been explored. FY-4B has optimized channel configurations compared to FY-4A, especially the newly added 7.42μm channels, which have unique advantages in water vapor retrieval. However, existing technologies have failed to fully utilize the information from these new channels, limiting further improvements in retrieval accuracy. This patent, through a multi-branch fusion architecture, fully utilizes the seven channels of FY-4B AGRI data, achieving a high-precision retrieval with an RMSE of 1.10mm and an R² of 0.988, improving accuracy by approximately 70% compared to traditional methods, thus fully exploring the application potential of FY-4B data.

[0054] Traditional methods mostly rely on "single-point" or "pixel-by-pixel" inversion, failing to fully utilize the valuable spatial context information contained within image data. Physical models and statistical traditional machine learning models typically treat each pixel as an independent processing unit, utilizing only its own multi-channel brightness and temperature information for calculation or prediction. However, the distribution of atmospheric water vapor is spatially continuous and highly correlated. The water vapor conditions, cloud morphology, texture, and spatial structure surrounding a pixel all contain crucial indicative information about the pixel's true water vapor content. Pixel-by-pixel processing completely ignores this neighborhood information. For example, a pixel located in a clear-sky area, if adjacent to a deep convection cloud, is likely to have a much higher water vapor content than a pixel located in the center of a large, uniform clear-sky area. Traditional methods cannot capture these differences in spatial patterns, thus limiting the accuracy and physical realism of the inversion. This patent adopts a U-Net encoder-decoder architecture, which fully utilizes spatial context information through multi-scale feature extraction and skip connections to achieve smooth transitions between adjacent pixels at a high resolution of 4km, with deviations controlled within ±2mm, effectively solving the problem of insufficient utilization of spatial features.

[0055] Existing CNN architectures are mostly directly transferred from the field of image recognition, lacking specific designs for the physical mechanisms of water vapor retrieval. Most existing technologies improve PWV retrieval accuracy by fusing multi-source information from existing PWV data, but lack network architecture designs specifically tailored to the characteristics of FY-4B AGRI data. Existing deep learning models fail to fully consider the unique characteristics of atmospheric water vapor retrieval, such as the physical meaning, spatial continuity, and temporal variation patterns of multi-channel brightness temperature data, resulting in limited model expressive power and difficulty in handling complex multi-source data fusion tasks. This patent designs a specialized multi-branch hybrid architecture, including a brightness temperature data branch, a DEM terrain branch, and a spatial attention fusion module, optimized for the physical mechanisms of water vapor retrieval, achieving deep fusion of multimodal information and significantly improving the model's expressive power and retrieval accuracy.

[0056] Atmospheric radiative transfer is an extremely complex nonlinear process, and existing technologies have limitations in fitting this complexity. While physical models have clear mechanisms, they rely on numerous simplifications and assumptions about atmospheric and surface conditions. These assumptions are often difficult to fully satisfy under real, variable conditions, leading to model errors. Furthermore, solving the inverse problem of the radiative transfer equation is itself an ill-posed problem; even small observational errors can cause significant deviations in the inversion results. In statistical and traditional machine learning methods, linear regression models clearly cannot characterize highly nonlinear mapping relationships. Although models like random forests have stronger nonlinear fitting capabilities, they essentially still segment the feature space and are relatively weak at learning smooth, continuous features with complex hierarchical structures in images. Moreover, these models are highly sensitive to the coverage and representativeness of training samples. When encountering extreme weather or underlying surface conditions not present in the training set, their generalization ability may decrease, producing unreliable predictions. This patent utilizes the powerful nonlinear fitting capabilities of deep learning models to maintain stable performance under three different atmospheric conditions: drought, normal, and high humidity. In particular, under high humidity conditions, the R² reaches 0.9818. Even under extreme weather conditions, it can still accurately capture the PWV distribution characteristics with the deviation controlled within ±2 mm, fully demonstrating the model's strong generalization ability.

[0057] China's complex topography, particularly in regions like the Qinghai-Tibet Plateau, necessitates improved accuracy in water vapor retrieval. Existing technologies lack systematic research on water vapor retrieval accuracy under varying topographic conditions, resulting in significant differences in accuracy between plains and mountainous areas. Current methods do not adequately consider the impact of topography on atmospheric parameter distribution and lack effective mechanisms for fusing topographic information, leading to a significant decrease in retrieval accuracy in complex terrain regions. This patent addresses this issue by introducing a DEM branch to specifically process topographic data and learn the static modulation patterns of water vapor distribution influenced by topography. This significantly improves retrieval accuracy in complex terrain regions, effectively resolving the accuracy degradation problem caused by topographic influence.

[0058] Different technical approaches involve trade-offs between computational cost and efficiency, making it difficult to achieve a balance. A complete physical radiative transfer model is computationally intensive, and iterative optimization pixel-by-pixel is extremely time-consuming, posing a significant challenge to operational applications requiring high timeliness. While methods such as lookup tables can be used for acceleration, this comes at the cost of some accuracy. Traditional machine learning models are fast in the inference phase, but they still face engineering challenges in feature extraction and data loading when processing large-scale, high-dimensional satellite imagery data, and they cannot benefit from the high parallelization acceleration of hardware such as GPUs like deep learning models. This patented model, on GPU hardware, takes less than 5 seconds to infer a complete FY-4B AGRI image (422×932 grid), fully meeting the timeliness requirements of near-real-time meteorological products, achieving a perfect combination of high accuracy and high efficiency, and possessing the capability to support near-real-time weather forecasting and warning operations.

[0059] refer to Figures 1 to 5 In view of this, this application proposes a method for inverting atmospheric precipitable water based on multi-source satellite remote sensing data, including:

[0060] Meteorological data from a geostationary satellite multispectral imager is acquired and reprojected onto a unified spatial grid. Clear-sky pixels are then selected based on the cloud masking data of the meteorological data, and cloud-covered areas are removed to obtain valid input data.

[0061] The multi-band radiation brightness temperature data in the effective input data is input to the first feature extraction branch, and multi-scale feature extraction is performed through the encoder-decoder structure. The encoder downsamples step by step to extract features at different scales, and the decoder restores the spatial resolution by upsampling and fuses the features of each level of the encoder with the corresponding level of the decoder through skip connections to obtain the first feature representation.

[0062] The digital elevation model data of the meteorological data is input into the second feature extraction branch, and the modulation features of topography on water vapor distribution are extracted through a convolutional neural network to obtain the second feature representation;

[0063] The spatiotemporal feature encoding data of the meteorological data is input into the third feature extraction branch, and the dynamic modulation features of the time series on the water vapor distribution are extracted through a one-dimensional convolutional neural network to obtain the third feature representation;

[0064] The first feature representation, the second feature representation, and the third feature representation are concatenated and input into the feature fusion network. A spatial attention mechanism is introduced to map the concatenated features into a single-channel inversion result through dimensionality reduction using convolutional layers. The training process uses a composite loss function, which includes mean squared error loss, mean absolute error loss, and gradient consistency loss. The gradient consistency constraint can adjust for overfitting. The single-channel inversion result is output as an atmospheric precipitable water product.

[0065] The meteorological data includes at least one of the following: multi-band radiation brightness temperature data, cloud mask data, digital elevation model data, and atmospheric reanalysis gridded data.

[0066] Reprojecting the meteorological data onto a unified spatial grid includes:

[0067] The geographic coordinate information corresponding to each pixel in the radiation brightness temperature data is obtained by parsing the geographic location data; a geographic coordinate system of the target spatial grid is established, wherein the target spatial grid covers a predetermined geographic area and has a uniform spatial resolution;

[0068] Based on the geographic coordinate information, the original pixels of the radiation brightness temperature data are mapped to the corresponding grid points of the target spatial grid through coordinate transformation, thereby completing the spatial registration of the radiation brightness temperature data;

[0069] Spatial interpolation is performed on the atmospheric reanalysis grid data to resample the original spatial resolution grid data to a spatial resolution consistent with the target spatial grid, resulting in downscaled atmospheric reanalysis grid data;

[0070] A resampling operation is performed on the digital elevation model data to adjust it to a spatial resolution and geographic range consistent with the target spatial grid.

[0071] The radiation brightness temperature data, the downscaled atmospheric reanalysis grid data, and the resampled digital elevation model data are spatially aligned pixel-by-pixel on the target spatial grid to form a spatially registered multi-source dataset.

[0072] After obtaining the first feature representation, the method further includes:

[0073] A spatial attention mechanism is applied to the first feature representation, and a spatial weight mask is generated through convolution operation and weighted to enhance the feature response of the active convection region and the region with significant water vapor gradient, thus obtaining a weighted first feature representation.

[0074] Multi-scale feature extraction is performed using an encoder-decoder structure. The encoder downsamples at different scales to extract features, while the decoder upsamples to restore spatial resolution and fuses features from each encoder level with corresponding features from the decoder via skip connections.

[0075] Multiple downsampling levels are set in the encoder. Each downsampling level includes convolution and pooling operations. The convolution operation extracts the feature representation at the current scale, and the pooling operation reduces the spatial resolution of the feature map and increases the receptive field. After downsampling at each level, a multi-scale feature pyramid from coarse to fine is obtained.

[0076] In the decoder, upsampling levels are set up to correspond to the number of encoder levels. Each upsampling level restores the spatial resolution of the feature map through transpose convolution or interpolation operations.

[0077] At each upsampling level, the feature map recovered by the decoder at the current level is skipped and connected to the feature map at the corresponding spatial scale level in the encoder. The skip connection fuses the spatial detail information of the encoder feature map into the decoder feature map through feature concatenation operation, compensating for the high-frequency spatial information lost during the upsampling process.

[0078] The skip connections enable the decoder to retain the multi-scale spatial structure features extracted by the encoder while restoring spatial resolution, and output a first feature representation that includes coarse-grained semantic information and fine-grained spatial details.

[0079] The training process also includes:

[0080] An adaptive optimizer is used to iteratively update the parameters of the feature fusion network. The adaptive optimizer dynamically adjusts the learning step size by maintaining the first-order moment estimate and second-order moment estimate of each parameter, and introduces a weight decay mechanism to impose regularization constraints on the parameters to suppress overfitting.

[0081] A learning rate scheduling strategy is set up to warm up the model by gradually increasing the learning rate in the initial training phase, and then gradually decaying the learning rate to its peak value according to a predetermined pattern in the main training phase. This allows the model parameters to converge quickly in the early stage of training and be fine-tuned in the later stage.

[0082] During training, the combined loss function value and predefined performance evaluation index on the validation set are monitored in real time. Training is terminated when the validation set performance index does not improve within a preset number of consecutive rounds, and the network parameters with the best performance on the validation set are saved as the final model parameters.

[0083] Another specific embodiment of this application is as follows:

[0084] This patent proposes a deep learning-based PWV inversion method based on multi-source data fusion. It employs a self-developed HybridPWVNet model to achieve end-to-end automated processing from FY-4B AGRI multi-channel brightness temperature data to high-precision PWV products. Compared with existing technologies, the core innovation of this patent lies in:

[0085] This patent designs a specialized multi-branch neural network architecture, including: a brightness temperature data branch, a DEM terrain branch, a temporal feature encoding branch, and a spatial attention fusion module. Compared with the single network architecture in existing technologies, this patent can fully utilize the feature information of different modal data, significantly improving the inversion accuracy.

[0086] This patent introduces a spatial attention module in the feature fusion stage, enabling the model to adaptively focus on key meteorological areas such as active convection zones. Compared with the pixel-by-pixel processing method in the prior art, it can better utilize spatial context information.

[0087] This patent introduces physical constraints into the loss function, including range constraints and gradient smoothing priors, to ensure the physical authenticity of the inversion results. Compared with the pure data-driven methods in the prior art, this improves the reliability and interpretability of the model.

[0088] This patent selects Level 1 data from the Advanced Geosynchronous Orbit Radiometric Imager (AGRI) aboard the FY-4B satellite, including brightness temperature data for seven moisture-sensitive infrared and visible light channels, geolocation files (GEO), and cloud masking products (CLM). The data covers hourly observations from June 6 to August 31, 2024, spatially covering a key mid-latitude region of China (30°–50°N, 72°–112°E), with a raw spatial resolution of 4 km.

[0089] The preprocessing process includes: first, batch converting the original HDF format data into standardized NPZ format files; second, reprojecting the brightness temperature data of all times onto a unified latitude and longitude grid based on the GEO file; and finally, using the CLM product to perform cloud detection and removal, retaining only clear-sky pixels for inversion.

[0090] This patent uses ECMWF fifth-generation global atmospheric reanalysis data (ERA5) as the training label for the deep learning model. The data time is matched with satellite data, providing hourly data, with an original spatial resolution of 0.25°×0.25°. To achieve strict spatial alignment with FY-4B data, a bilinear interpolation method is used to downscale the ERA5 PWV grid data to a 4 km resolution satellite grid.

[0091] This patent uses data from a network of ground-based GNSS observation stations deployed within the study area as an independent verification source. Through rigorous time synchronization and spatial matching, a spatiotemporal matching dataset containing 6 stations and a total of 627 valid observation points was constructed for the objective evaluation of the final inversion product.

[0092] This patent uses a 4 km resolution digital elevation model (DEM) for the Chinese region as a static auxiliary input to explicitly characterize the modulating effect of topographic relief on the vertical distribution of water vapor and local circulation. The DEM data is resampled to a 4 km grid consistent with FY-4B and used as an independent input channel for the model.

[0093] The final structured deep learning dataset consists of eight input features (seven brightness temperature channels and one DEM elevation channel from FY-4B AGRI) and interpolated ERA5 PWV values ​​as labels. The dataset covers 2055 hourly time points, with a spatial size of 422×932 and a total data size of approximately 14.45 GB. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio. Quality control measures include removing outliers based on physical common sense, ensuring perfect timestamp matching, and guaranteeing the independence of the GNSS validation data.

[0094] Hybrid PWV Net is a multi-branch hybrid neural network architecture based on the U-Net backbone, specifically designed for retrieving atmospheric precipitable water (PWV) from FY-4B AGRI data. This model employs a multimodal data processing strategy, using multiple parallel input branches to process different types of data sources, and leveraging a feature fusion mechanism to achieve efficient integration and extraction of multi-source information.

[0095] The model has three input branches: the brightness temperature data branch receives multi-channel infrared and visible light data with dimensions [B,7,H,W], and directly inputs it into the U-Net backbone network for spatial feature extraction; the DEM data branch specifically processes single-channel elevation data with dimensions [B,1,H,W], and extracts the modulation features of terrain on water vapor distribution through a lightweight convolutional subnetwork (containing two 1×1 convolutional layers, which expand the number of channels from 1 to 32, and then further extract to 64 channels); the time branch is responsible for processing multi-temporal data (input dimension [B,8,H,W]), and captures temporal dynamic changes by flattening the spatial dimension and applying one-dimensional convolution, finally outputting a 32-channel feature map.

[0096] The U-Net backbone employs a classic encoder-decoder architecture. The encoder extracts multi-scale features through four levels of downsampling: the initial layer uses DoubleConv (a double convolutional block containing convolution, batch normalization, and ReLU activation) to convert the 7-channel input into 64-channel features; subsequently, it sequentially performs max pooling and DoubleConv operations, gradually expanding the number of channels to 128, 256, 512, until reaching the 1024-dimensional high-dimensional features at the bottom layer. The decoder performs upsampling through transposed convolutions, with each level making skip connections to the features of the corresponding layer in the encoder, gradually restoring the spatial resolution and ultimately outputting 64-channel high-resolution features.

[0097] Furthermore, the model introduces a spatial attention mechanism before feature fusion, generating spatial weight masks for U-Net features through 7×7 convolutions, enabling the model to adaptively focus on key research regions. Finally, the output layer maps 32-channel features to a 1-channel output through a 1×1 convolution operation, generating a PWV inversion result with the same spatial size as the input.

[0098] In the feature fusion stage, the model concatenates the 64-channel features output from the U-Net backbone, the 64-channel terrain features extracted from the DEM branch, and the 32-channel temporal features generated by the time branch to form a 160-channel fused feature tensor. This tensor is then further fused with multimodal information through a three-layer convolutional network (with the number of channels compressed to 64, 32, and 1 respectively), batch normalization, and ReLU activation.

[0099] The entire network uses DoubleConv as the basic building block, and batch normalization and ReLU activation are used to ensure training stability. At the same time, through multi-branch design and feature fusion mechanism, it effectively integrates spatial details, terrain modulation and temporal variation information, which significantly improves the accuracy and robustness of PWV inversion under complex weather conditions.

[0100] This patent employs the AdamW optimizer with a learning rate of 1e-3 and a weight decay coefficient of 1e-4 to prevent overfitting. The loss function is a weighted combination of MSE and MAE losses (weight coefficient α=0.7), with an additional gradient loss term to maintain spatial smoothness. Specifically, the total loss function is L_total = 0.7×MSE + 0.3×MAE + 0.1×L_gradient, where the gradient loss constrains spatial continuity by calculating the gradient difference between the predicted and true values ​​in the x and y directions.

[0101] The training batch size was set to 8, and the model underwent 65 training epochs on the GPU, employing the OneCycleLR learning rate scheduling strategy. The first 30% of training steps were used for learning rate warm-up, with a maximum learning rate of 1e-3. During training, key metrics such as loss, RMSE, and MAE on both the training and validation sets were monitored in real time, and an early stopping mechanism with performance thresholds and delayed start was implemented. Specifically, the first 50 epochs were a patient training period, unaffected by the early stopping mechanism; after 50 epochs, strict quality control was implemented, and only models with a validation RMSE less than 2.0 mm and a validation loss significantly reduced by more than 0.1 compared to the historical best were saved as the next generation's best model; if this standard was not met for 15 consecutive epochs, training would automatically stop. The training process integrated breakpoint resume functionality and detailed training log archiving.

[0102] Inference Process: The optimal model, once trained, can be used for batch processing of historical data or near real-time processing of new data. The inference module receives single or multiple FY-4B images and corresponding DEM data, outputs hourly PWV grid products in standard format, along with metadata and quality identification information.

[0103] Dual validation strategy: An internal and external dual validation strategy is adopted. Internal validation uses an independent test set, based on ERA5 reanalysis data, to systematically calculate statistical indicators such as RMSE, MAE, coefficient of determination (R²), and bias. External validation uses completely independent GNSS ground observation data as the ground truth, and quantifies the difference between the model inversion values ​​and the measured values ​​after spatiotemporal registration.

[0104] The HybridPWVNet model training process implemented in this patent exhibits good convergence characteristics. Figure 2 By monitoring validation set performance and employing an early stopping strategy, the model achieved optimal validation performance in the 56th epoch, at which point the validation set RMSE was 1.326 mm, the MAE was 0.980 mm, and the training loss decreased to 5.562. The training curves show that the model converged rapidly in the initial stages, with both training and validation set metrics exhibiting a stable downward trend, validating the rationality of the model architecture design and the effectiveness of the training strategy. This indicates that the model converged rapidly in the early stages of training (the first 20 epochs), with the training set RMSE rapidly decreasing from 15 mm to below 2 mm, and the validation set RMSE decreasing from 15 mm to around 5 mm. In the middle stages of training (20-50 epochs), the model entered a fine-tuning phase, with moderate fluctuations in the validation set metrics. This is a normal learning process for deep learning models and helps the model escape local optima. Finally, the optimal performance was achieved in the 56th epoch, at which point the training and validation set metrics converged, indicating that the model has good generalization ability and avoids overfitting.

[0105] In evaluations on independent test sets, compared to ERA5 data ( Figure 3 The inversion product exhibits excellent performance indicators: RMSE of 1.10 mm, MAE of 0.85 mm, coefficient of determination R² of 0.988, and average deviation of only 0.52 mm. Approximately 98.87% of the data errors are distributed within ±3 mm, with 72.35% of the sample errors between 0 and 3 mm, and 26.52% of the sample errors between -3 and 0 mm. Figure 4 In external validation with PWV data measured at completely independent GNSS ground stations ( Figure 5 The model achieved an R² of 0.926 between the inverted values ​​and the measured GNSS values, an RMSE of 3.549 mm, and a MAE of 2.748 mm. Dual validation using independent ERA5 test datasets and GNSS datasets fully demonstrates the model's high accuracy and stability.

[0106] This patented method demonstrates stability under both arid and humid conditions, exhibiting excellent adaptability. Analysis of the spatial error distribution of the inversion results shows that, regardless of whether in arid regions with low PWV values ​​(e.g., PWV < 5 mm) or in humid environments with abundant moisture (e.g., PWV > 20 mm), the model does not show any systematic overestimation or underestimation biases, and the error distribution is relatively uniform. Specific performance characteristics are as follows:

[0107] Arid conditions (PWV < 5 mm): In tests with 525,593 samples, the model demonstrated extremely high accuracy, with an RMSE of only 0.4873 mm, a MAE of 0.3707 mm, and a coefficient of determination (R²) of 0.5516. Although the R² was relatively low under extremely low water vapor conditions, the absolute error was controlled within 0.5 mm, fully demonstrating the stability and reliability of the model under arid conditions.

[0108] Under normal conditions (5 mm ≤ PWV ≤ 20 mm): In a large-scale test with 9,036,575 samples, the model performance was significantly improved, with an RMSE of 0.9471 mm, a MAE of 0.7393 mm, and a coefficient of determination R² of 0.9446. This result demonstrates that the model can accurately capture water vapor distribution characteristics under normal atmospheric conditions, providing reliable technical support for operational applications.

[0109] High humidity conditions (PWV > 20 mm): In high humidity environment tests with 4,078,293 samples, the model maintained excellent performance, with an RMSE of 1.4233 mm, a MAE of 1.1412 mm, and a high coefficient of determination (R²) of 0.9818. Notably, the highest R² was observed under high humidity conditions, indicating that the model has the strongest inversion capability for high water vapor content and can accurately identify and quantify heavy precipitation weather systems.

[0110] Overall, the model exhibits good adaptability under various atmospheric conditions. The error increases slightly with increasing PWV value, but the increase remains within a reasonable range. In particular, under high humidity conditions, the R² reaches 0.9818, indicating that the model has an outstanding ability to characterize complex meteorological conditions and provides an important technical means for extreme weather monitoring and early warning.

[0111] Compared to traditional physical inversion methods, the HybridPWVNet model implemented in this patent has achieved a breakthrough in PWV inversion accuracy, with an RMSE controlled within 1.10 mm, a MAE of 0.85 mm, and a coefficient of determination R² of 0.988. This accuracy represents a significant technological improvement in this field. According to relevant literature, the RMSE of traditional physical inversion methods under similar regions and conditions is typically between 4 and 6 mm, with even greater errors under complex weather or terrain conditions, especially under adverse conditions such as cloud cover and complex terrain, where the error of traditional methods can reach 6-8 mm (Biyan Chen, 2018 et al.). The advantages of this patent are mainly attributed to the powerful nonlinear fitting capability of the deep learning model and the effective fusion of multi-source data, achieving an accuracy improvement of approximately 70% compared to traditional methods. This improvement has significant technical implications in the field of satellite remote sensing PWV inversion.

[0112] Compared to machine learning methods that have emerged in recent years, this patent also demonstrates significant advantages. The traditional random forest method (Zhang Huanyu, 2021) achieves an RMSE of 2.47 mm in PWV inversion, which decreases to 1.5 mm after quality control. In contrast, this patented method achieves an accuracy of 1.10 mm without additional quality control. In the latest CNN-based study of water vapor PWV inversion (Cuixian Lu, 2023), also using ERA5 data as a benchmark, the inversion accuracy is RMSE 2.0 mm and MAE 1.4 mm. This patent achieves further accuracy improvements, reducing RMSE by 45% and MAE by 39%, fully demonstrating the technical advantages of the multi-branch hybrid architecture and spatial attention mechanism.

[0113] This patent, by introducing DEM branching and physical consistency constraints, maintains stable performance under three different atmospheric conditions: drought, normal, and high humidity. Particularly under high humidity conditions, R² reaches 0.9818, providing reliable technical support for extreme weather monitoring and early warning. This technological breakthrough not only improves the overall accuracy of PWV inversion but, more importantly, enhances the model's robustness and practicality under complex meteorological conditions, opening up a new technical path for the development of satellite remote sensing water vapor inversion technology.

[0114] A second aspect of the present invention provides an atmospheric precipitable water retrieval system based on multi-source satellite remote sensing data, comprising:

[0115] The first unit is used to acquire meteorological data from a geostationary satellite multispectral imager, reproject the meteorological data onto a unified spatial grid, and filter clear-sky pixels based on the cloud mask data of the meteorological data, removing cloud-covered areas to obtain valid input data.

[0116] The second unit is used to input the multi-band radiation brightness temperature data from the effective input data into the first feature extraction branch, and perform multi-scale feature extraction through an encoder-decoder structure. The encoder downsamples step by step to extract features at different scales, and the decoder restores the spatial resolution by upsampling and fuses the features of each level of the encoder with the corresponding level of the decoder through skip connections to obtain the first feature representation.

[0117] The third unit is used to input the digital elevation model data of the meteorological data into the second feature extraction branch, and extract the modulation features of topography on water vapor distribution through a convolutional neural network to obtain the second feature representation;

[0118] The fourth unit is used to input the spatiotemporal feature encoded data of the meteorological data into the third feature extraction branch, and extract the dynamic modulation features of spatiotemporal water vapor distribution through a one-dimensional convolutional neural network to obtain the third feature representation;

[0119] The fifth unit is used to concatenate the first feature representation, the second feature representation, and the third feature representation and input them into the feature fusion network. The feature fusion network maps the concatenated features into a single-channel inversion result through layer-by-layer dimensionality reduction. The training process uses a combined loss function composed of a mean square error term, a mean absolute error term, and a gradient smoothing term. The gradient smoothing term constrains the spatial continuity by calculating the gradient difference between the inversion result and the true value in the horizontal and vertical directions. The single-channel inversion result is output as a gridded product of atmospheric precipitable water.

[0120] A third aspect of the embodiments of the present invention,

[0121] An electronic device is provided, comprising:

[0122] processor;

[0123] Memory used to store processor-executable instructions;

[0124] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0125] Fourth aspect of the present invention,

[0126] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0127] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for retrieving atmospheric precipitable water based on multi-source satellite remote sensing data, characterized in that, include: Meteorological data from a geostationary satellite multispectral imager is acquired, and the meteorological data is reprojected onto a unified spatial grid. Based on the cloud mask data of the meteorological data, clear-sky pixels are filtered out, and cloud-covered areas are removed to obtain valid input data; The multi-band radiation brightness temperature data in the effective input data is input to the first feature extraction branch, and multi-scale feature extraction is performed through the encoder-decoder structure. The encoder downsamples step by step to extract features at different scales, and the decoder restores the spatial resolution by upsampling and fuses the features of each level of the encoder with the corresponding level of the decoder through skip connections to obtain the first feature representation. The digital elevation model data of the meteorological data is input into the second feature extraction branch, and the modulation features of topography on water vapor distribution are extracted through a convolutional neural network to obtain the second feature representation; The spatiotemporal feature encoding data of the meteorological data is input into the third feature extraction branch, and the dynamic modulation features of the time series on the water vapor distribution are extracted through a one-dimensional convolutional neural network to obtain the third feature representation; The first feature representation, the second feature representation, and the third feature representation are concatenated and input into the feature fusion network. A spatial attention mechanism is introduced to map the concatenated features into a single-channel inversion result through dimensionality reduction using convolutional layers. The training process uses a composite loss function, which includes mean squared error loss, mean absolute error loss, and gradient consistency loss. The gradient consistency constraint can adjust for overfitting. The single-channel inversion result is output as an atmospheric precipitable water product.

2. The method according to claim 1, characterized in that, The meteorological data includes at least one of the following: multi-band radiation brightness temperature data, cloud mask data, digital elevation model data, and atmospheric reanalysis gridded data.

3. The method according to claim 2, characterized in that, Reprojecting the meteorological data onto a unified spatial grid includes: The geographic coordinate information corresponding to each pixel in the radiation brightness temperature data is obtained by parsing the geographic location data; a geographic coordinate system of the target spatial grid is established, wherein the target spatial grid covers a predetermined geographic area and has a uniform spatial resolution; Based on the geographic coordinate information, the original pixels of the radiation brightness temperature data are mapped to the corresponding grid points of the target spatial grid through coordinate transformation, thereby completing the spatial registration of the radiation brightness temperature data; Spatial interpolation is performed on the atmospheric reanalysis grid data to resample the original spatial resolution grid data to a spatial resolution consistent with the target spatial grid, resulting in downscaled atmospheric reanalysis grid data; A resampling operation is performed on the digital elevation model data to adjust it to a spatial resolution and geographic range consistent with the target spatial grid. The radiation brightness temperature data, the downscaled atmospheric reanalysis grid data, and the resampled digital elevation model data are spatially aligned pixel-by-pixel on the target spatial grid to form a spatially registered multi-source dataset.

4. The method according to claim 1, characterized in that, After obtaining the first feature representation, the method further includes: A spatial attention mechanism is applied to the first feature representation, and a spatial weight mask is generated through convolution operation and weighted to enhance the feature response of the active convection region and the region with significant water vapor gradient, thus obtaining a weighted first feature representation.

5. The method according to claim 1, characterized in that, Multi-scale feature extraction is performed using an encoder-decoder structure. The encoder downsamples at different scales to extract features, while the decoder upsamples to restore spatial resolution and fuses features from each encoder level with corresponding features from the decoder via skip connections. Multiple downsampling levels are set in the encoder. Each downsampling level includes convolution and pooling operations. The convolution operation extracts the feature representation at the current scale, and the pooling operation reduces the spatial resolution of the feature map and increases the receptive field. After downsampling at each level, a multi-scale feature pyramid from coarse to fine is obtained. In the decoder, upsampling levels are set up to correspond to the number of encoder levels. Each upsampling level restores the spatial resolution of the feature map through transpose convolution or interpolation operations. At each upsampling level, the feature map recovered by the decoder at the current level is skipped and connected to the feature map at the corresponding spatial scale level in the encoder. The skip connection fuses the spatial detail information of the encoder feature map into the decoder feature map through feature concatenation operation, compensating for the high-frequency spatial information lost during the upsampling process. The skip connections enable the decoder to retain the multi-scale spatial structure features extracted by the encoder while restoring spatial resolution, and output a first feature representation that includes coarse-grained semantic information and fine-grained spatial details.

6. The method according to claim 1, characterized in that, The training process also includes: An adaptive optimizer is used to iteratively update the parameters of the feature fusion network. The adaptive optimizer dynamically adjusts the learning step size by maintaining the first-order moment estimate and second-order moment estimate of each parameter, and introduces a weight decay mechanism to impose regularization constraints on the parameters to suppress overfitting. A learning rate scheduling strategy is set up to warm up the model by gradually increasing the learning rate in the initial training phase, and then gradually decaying the learning rate to its peak value according to a predetermined pattern in the main training phase. This allows the model parameters to converge quickly in the early stage of training and be fine-tuned in the later stage. During training, the combined loss function value and predefined performance evaluation index on the validation set are monitored in real time. Training is terminated when the validation set performance index does not improve within a preset number of consecutive rounds, and the network parameters with the best performance on the validation set are saved as the final model parameters.

7. An atmospheric precipitable water inversion system based on multi-source satellite remote sensing data, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to acquire meteorological data from a geostationary satellite multispectral imager and reproject the meteorological data onto a unified spatial grid. Based on the cloud mask data of the meteorological data, clear-sky pixels are filtered out, and cloud-covered areas are removed to obtain valid input data; The second unit is used to input the multi-band radiation brightness temperature data from the effective input data into the first feature extraction branch, and perform multi-scale feature extraction through an encoder-decoder structure. The encoder downsamples step by step to extract features at different scales, and the decoder restores the spatial resolution by upsampling and fuses the features of each level of the encoder with the corresponding level of the decoder through skip connections to obtain the first feature representation. The third unit is used to input the digital elevation model data of the meteorological data into the second feature extraction branch, and extract the modulation features of topography on water vapor distribution through a convolutional neural network to obtain the second feature representation; The fourth unit is used to input the spatiotemporal feature encoded data of the meteorological data into the third feature extraction branch, and extract the dynamic modulation features of spatiotemporal water vapor distribution through a one-dimensional convolutional neural network to obtain the third feature representation; The fifth unit is used to concatenate the first feature representation, the second feature representation, and the third feature representation and input them into the feature fusion network. The feature fusion network maps the concatenated features into a single-channel inversion result through layer-by-layer dimensionality reduction. The training process uses a combined loss function composed of a mean square error term, a mean absolute error term, and a gradient smoothing term. The gradient smoothing term constrains the spatial continuity by calculating the gradient difference between the inversion result and the true value in the horizontal and vertical directions. The single-channel inversion result is output as a gridded product of atmospheric precipitable water.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images