Atmospheric humidity profile inversion method based on Unet3 + neural network

Abstract

The invention provides an atmospheric humidity profile inversion method based on a Unet3 + neural network. The atmospheric humidity profile inversion method comprises the following steps: constructing an atmospheric humidity profile inversion model based on Unet3 + MHS microwave data; performing radiometric calibration, projection conversion and quality control on the multi-channel brightness temperature observed by the original satellite; the method comprises the following steps: selecting long-time scale radiation brightness temperature data in a Chinese area range and specific humidity data of ERA-5 corresponding to time and space to make a data set; and performing humidity profile inversion according to the humidity profile inversion model of the brightness temperature data constructed in the step S1. The method has the advantages that the Unet3 + model in deep learning and microwave sensor data which are not affected by time and weather are combined, and the problems that the space distribution of the atmospheric humidity profile is small, and the time resolution is low can be further solved; the method provides atmospheric condition data with higher temporal and spatial resolution for numerical weather forecast research, and has good application value.

Description

Technical Field

[0001] The present invention belongs to the field of atmospheric humidity profile inversion of microwave hygrometers, and in particular relates to an atmospheric humidity profile inversion method based on a Unet3+ neural network. Background Art

[0002] The content and distribution of water vapor in the atmosphere have a significant impact on weather and climate. Atmospheric humidity directly influences both shortwave solar radiation and longwave radiation from the Earth's atmosphere, profoundly impacting the global radiative energy budget. Furthermore, atmospheric humidity is closely related to numerous meteorological phenomena, including the stability of atmospheric stratification, convective potential energy, precipitation processes, and cloud formation and evolution. Therefore, the observation and inversion of atmospheric humidity are not only crucial for the initialization and validation of numerical weather prediction models but also possess significant research value and practical application in numerous meteorological applications, including climate monitoring, weather analysis, and weather forecasting.

[0003] Currently, atmospheric temperature and humidity profiles are primarily acquired through satellite observations and radiosondes. These methods provide highly reliable and representative atmospheric temperature and humidity parameters. However, conventional radiosondes have very limited temporal and spatial coverage, particularly over oceans and plateaus, where data volumes are limited, failing to meet operational needs. Satellite-based observations offer advantages such as wide coverage and high temporal and spatial resolution, which to some extent offset the limitations of conventional radiosonde data. In particular, the continuous improvement of spectral and spatial resolution has opened up new possibilities for atmospheric observation. Microwave remote sensing, among other technologies, is highly regarded for its ability to penetrate clouds and haze. Self-calibrating microwave radiometers (MHSs), primarily carried on satellites such as the European Meteorological Observation Satellites (EUMETSAT Metop-A / B / C) and the Polar Operational Environmental Satellites (POES NOAA-18 / 19), can observe the Earth across five channels in the millimeter wave band (89–190 GHz) with a ±50° field of view. However, current satellite sensors cannot directly acquire air humidity; instead, they must invert it. In satellite remote sensing, the inversion of atmospheric parameters is a complex and highly sought-after mathematical and physical problem.

[0004] Traditional inversion methods can be categorized based on different algorithmic theoretical frameworks, including statistical regression and physical inversion methods. Statistical regression methods lack the physical meaning of radiative transfer and are therefore difficult to apply when data samples are limited. Physical inversion methods are also affected by observational conditions, such as underlying surface conditions and cloud distribution. Summary of the Invention

[0005] In view of this, the present invention aims to propose an atmospheric humidity profile inversion method based on Unet3+ neural network to solve the problem that traditional inversion methods are difficult to be widely used.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows: An atmospheric humidity profile inversion method based on Unet3+ neural network includes the following steps: S1. Construct an atmospheric humidity profile inversion model based on MHS microwave data from Unet3+; S2. Perform radiometric calibration, projection conversion, and quality control on the multi-channel brightness temperature observed by the original satellite; S3. Select the long-term radiation brightness temperature data of China and the ERA-5 specific humidity data of the corresponding time and space to create the dataset; S4, performing humidity profile inversion according to the humidity profile inversion model of brightness temperature data constructed in step S1; In step S1, an atmospheric humidity profile inversion model based on MHS microwave data using Unet3+ is constructed, including: S11, feature extraction; S12, feature decoding; S13, full-size skip connection; S14, activation function; S15, deep supervision; In step S2, radiometric calibration, projection conversion, and quality control are performed on the original satellite-observed multi-channel brightness temperature, including: S21, performing radiometric calibration on the original data to obtain brightness temperature results of multiple channels; S22. Construct a standard grid with a spatial resolution of 0.25°×0.25°, and fill the original data in swath format into the standard grid according to longitude and latitude; S23. Perform quality control on the data after radiometric calibration and reprojection, including removing edge effects and outliers.

[0007] Furthermore, in step S11, feature extraction includes: Use the Unet3+ encoder to downsample the image and obtain features of different scales; The encoder consists of several groups of downsampling modules, each of which includes two convolutional layers and one pooling layer.

[0008] Furthermore, in step S12, feature decoding includes: The decoder of Unet3+ uses convolution, maximum pooling, and deconvolution operations to fuse feature maps of different sizes, and finally convolutes them with filters to output the fused feature map; The decoder consists of convolutional layers, pooling layers, and deconvolution layers, and distributes deep supervision.

[0009] Furthermore, in step S13, the full-size skip connection includes: Smaller encoders Downsampling, the same scale Direct convolution, large-scale decoder Perform upsampling and fuse multi-scale features.

[0010] Furthermore, in step S14, the activation function includes: After each convolutional layer, the ReLu activation function is used to build a nonlinear mapping relationship between the input and output. The expression is as follows: .

[0011] Furthermore, in step S15, deep supervision includes: The network loss is calculated by comparing the output and input of each decoding block. The loss function expression is as follows: ; Where, is the true value in the training data sample, is the model prediction value, for The average value of for The average value of yes The variance of yes The variance of yes and The covariance of , , Take 0.01, Take 0.03, is the dynamic range of the image pixel values.

[0012] Furthermore, in step S3, data set creation includes: S31, cropping the brightness temperature after calibration, reprojection, and quality control in step S2, marking the imaging time, setting the time window to ±30 minutes, and rounding the time; S32, selecting label data according to the rounded time of the microwave hygrometer, selecting the specific humidity of the ERA-5 multi-pressure layer corresponding to the hygrometer imaging time, and cutting it into the same spatial range as after the hygrometer processing; S33. Divide the dataset into training set, validation set and test set.

[0013] Furthermore, in step S4, the humidity profile inversion includes: According to the humidity profile inversion model constructed in step S1, the model parameters are continuously adjusted until the loss curve of the model is fitted and the value approaches 0, and the humidity profile inversion error is within 1.5g / kg.

[0014] Compared with the prior art, the atmospheric humidity profile inversion method based on the Unet3+ neural network described in the present invention has the following beneficial effects: (1) This invention combines the Unet3+ model in deep learning with microwave sensor data that is not affected by time and weather, which can further solve the problems of the limited spatial distribution and low temporal resolution of atmospheric humidity profiles. (2) This invention provides atmospheric condition data with higher temporal and spatial resolution for numerical weather forecast research and has good application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings, which constitute part of the present invention, are provided to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are provided to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings: Figure 1 This is a schematic diagram of the Unet3+ network structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the root mean square error distribution of the 2021 full-year inversion results compared with sounding observations, and the ERA-5 compared with sounding observations as described in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0017] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0018] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0019] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0020] like Figures 1 to 2 As shown in the figure, the atmospheric humidity profile inversion method based on the Unet3+ neural network includes the following steps: S1. Construct an atmospheric humidity profile inversion model based on MHS microwave data using Unet3+; S2. Perform radiometric calibration, projection conversion, and quality control on the multi-channel brightness temperature observed by the original satellite; S3. Select the long-term radiation brightness temperature data of China and the ERA-5 specific humidity data of the corresponding time and space to create the dataset; S4. Perform humidity profile inversion based on the humidity profile inversion model of brightness temperature data constructed in step S1.

[0021] The specific implementation is as follows: Step 1: Construction of humidity profile inversion model based on Unet3+ microwave hygrometer like Figure 1 The following is a diagram showing the framework of the Unet3+ model used in this invention, which mainly consists of an encoder, a decoder, and a full-size skip connection module. The implementation process is as follows: (1) Encoder The present invention uses the basic structure of Unet3+, that is, the encoder-decoder structure. The left part of the network is the encoder structure, which is composed of convolutional layers and pooling layers. The convolutional layer adopts a 3×3 convolutional layer (channel number 64→128→256→512→1024), and the pooling layer adopts 2×2 maximum pooling. Its function is to use convolution and maximum pooling operations to downsample the image, extract image features, and generate feature maps of different scales.

[0022] (2) Decoder The right side of the network is the decoder, which consists of convolutional, pooling, and deconvolutional layers. It uses full-scale skip connections and employs convolution, max pooling, and deconvolution operations to fuse feature maps of different sizes. Finally, it applies convolution to these maps using filters and outputs the fused feature map, gradually restoring the feature map to its original size. Deep supervision is also implemented within each decoder, comparing the input and output of each decoder to calculate the loss, thereby learning hierarchical representations from the aggregated feature maps at all scales.

[0023] (3) Full-size skip connections structure The whole process of feature map, feature map The full-size connection mainly comes from three parts: Smaller encoders and ,This part retains the detail information, that is, it integrates fine-grained semantics. and To downsample, downsample by 4 times and 2 times respectively so that the feature maps are unified to the same size, and then convolution is performed (convolution kernel size 3×3) with 64 channels.

[0024] Feature maps of encoders of the same scale , this layer directly performs 3×3 convolution and 64-channel operation.

[0025] In larger decoders and ,This part integrates coarse-grained semantic information. and ( ) are upsampled by a factor of 2 and a factor of 4 respectively to keep the feature map size consistent. The upsampling method used is bilinear interpolation.

[0026] After each convolutional layer, the ReLu activation function is used to build a nonlinear mapping relationship between the input and output. The calculation formula is as follows: ; Deep supervision is implemented in each layer of the deep supervision decoder. The network loss is calculated by comparing the output and input of each decoding block to accurately invert the humidity profile. The loss function is calculated as follows: ; Where, is the true value in the training data sample, is the model prediction value, for The average value of for The average value of yes The variance of yes The variance of yes and The covariance of , , Generally, 0.01 is used. Generally, 0.03 is taken. is the dynamic range of the image pixel values.

[0027] Step 2: Rapid projection conversion and quality control of satellite observation data Since the microwave hygrometer is carried on a polar-orbiting satellite and uses a cross-track scanning method, the edge pixels of the scan line are affected by the long signal transmission path, resulting in relatively low quality. In addition, the data has not been quality-controlled, and there are abnormal observation pixels. To ensure the accuracy and reliability of the inversion model, data quality control is required. Secondly, since the projection method, grid point coordinates, and imaging time between the label data and the observed brightness temperature are not completely consistent, the two need to be matched in time and space. The specific steps are as follows: (1) The original observation data is stored as DN values. When performing humidity inversion, brightness temperature is required as input. Therefore, the original data is first radiometrically calibrated to obtain multi-channel brightness temperature results. (2) The original data is in swath scan format, which is inconsistent with the corresponding label data projection method and the longitude and latitude of the corresponding grid. To ensure the accuracy and reliability of the dataset, a standard grid is constructed with a spatial resolution of 0.25°, and the original data in swath format are filled into the standard grid according to the longitude and latitude; (3) Perform quality control on the data after radiometric calibration and reprojection. Since the data are L1 brightness temperatures that have not undergone quality control, quality control is performed on the data before constructing the dataset, including removing edge effects and eliminating outliers.

[0028] The edge effect removal means that the radiation observed on both sides of the scanning line is affected by the observation path and the quality is low. Five pixels on both sides of each scanning line are removed.

[0029] To eliminate outliers, we need to collect statistics on the minimum, maximum, mean, and standard deviation of each channel of the observed data, and then make statistics on each indicator of each channel to eliminate outliers.

[0030] Step 3: Select ERA-5 label data for the corresponding period and create a dataset (1) Crop the brightness temperature after radiometric calibration, reprojection, and quality control in step 2, and mark the imaging time: the time window is set to ±30 minutes, and the time is rounded; (2) The label data were selected according to the rounded time of the microwave hygrometer. The humidity data of ERA-5 multi-pressure layers (300, 400, 500, 600, 700, 800, 850, 900, 1000 hPa) at the corresponding time were selected and cut into the same spatial range as that after the hygrometer processing.

[0031] (3) A total of 21,793 data points were obtained from the time range of 1998 to 2007 and 2021, and the data set was divided into training set, validation set, and test set in a ratio of 8:1:1. Random rotation (±10°), translation (±5%), and brightness perturbation (±2K) were applied to the training set to improve the robustness of the model; Step 4: Invert humidity profile According to the humidity profile inversion model of the Unet3+ microwave hygrometer constructed in step 1, the model parameters are continuously adjusted until the loss curve of the model is fitted and the value approaches 0, and the humidity profile inversion error is within 1.5g / kg.

[0032] The Unet3+ humidity profile inversion model was trained using the microwave hygrometer dataset constructed above. By continuously adjusting model parameters, the model achieved efficient and accurate humidity profile inversion. This model accurately inverts atmospheric humidity profiles for different regions and at different times, providing spatially continuous and wide-ranging humidity information at all altitudes, providing strong data support for atmospheric monitoring and numerical weather forecasting.

[0033] Figure 2 This is the root mean square error (RMSE) distribution diagram of the inversion results of the present invention compared with the sounding observations for the whole year of 2021, and the ERA-5 compared with the sounding observations; Table 1 shows the accuracy results of the test set of each altitude layer of the present invention (data for the whole year of 2021).

[0034] Table 1 ; The present invention proposes a deep learning model for humidity profile inversion based on Unet3+ microwave hygrometer, and reprojects and controls the quality of the original observed brightness temperature data. It can achieve high-precision all-day and all-weather atmospheric humidity profile inversion. Based on the Unet3+ neural network method, the observed brightness temperature data of the microwave hygrometer is used as input. The inverted humidity profile is compared with the sounding data. The RMSE of the entire layer is 0.91 g / kg, and the RMSE of each altitude layer is within 1.35 g / kg.

[0035] The advantages and beneficial effects of the present invention are as follows: (1) This invention combines the Unet3+ model in deep learning with microwave sensor data that is not affected by time and weather, which can further solve the problems of the limited spatial distribution and low temporal resolution of atmospheric humidity profiles. (2) This invention provides atmospheric condition data with higher temporal and spatial resolution for numerical weather forecast research and has good application value.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The atmospheric humidity profile inversion method based on Unet3+ neural network is characterized by: The following steps are involved: S1. Construct an atmospheric humidity profile inversion model based on MHS microwave data from Unet3+; S2. Perform radiometric calibration, projection conversion, and quality control on the multi-channel brightness temperature observed by the original satellite; S3. Select the long-term radiation brightness temperature data of China and the ERA-5 specific humidity data of the corresponding time and space to create the dataset; S4, performing humidity profile inversion according to the humidity profile inversion model of brightness temperature data constructed in step S1; In step S1, an atmospheric humidity profile inversion model based on MHS microwave data using Unet3+ is constructed, including: S11, feature extraction; S12, feature decoding; S13, full-size skip connection; S14, activation function; S15, deep supervision; In step S11, feature extraction includes: Use the Unet3+ encoder to downsample the image and obtain features of different scales; The encoder consists of several groups of downsampling modules, each of which includes two convolutional layers and one pooling layer; In step S12, feature decoding includes: The decoder of Unet3+ uses convolution, maximum pooling, and deconvolution operations to fuse feature maps of different sizes, and finally convolutes them with filters to output the fused feature map; The decoder consists of convolutional, pooling, and deconvolution layers, with distributed deep supervision. In step S13, full-size skip connection is performed, including: Smaller encoders Downsampling, the same scale Direct convolution, large-scale decoder Perform upsampling and fuse multi-scale features.

2. The atmospheric humidity profile inversion method based on the Unet3+ neural network according to claim 1, characterized in that: In step S14, the activation function includes: After each convolutional layer, the ReLu activation function is used to build a nonlinear mapping relationship between the input and output. The expression is as follows: 。 3. The atmospheric humidity profile inversion method based on the Unet3+ neural network according to claim 1, characterized in that: In step S15, deep supervision includes: The network loss is calculated by comparing the output and input of each decoding block. The loss function expression is as follows: ； Where, is the true value in the training data sample, is the model prediction value, for The average value of for The average value of yes The variance of yes The variance of yes and The covariance of , , Take 0.01, Take 0.03, is the dynamic range of the image pixel values.

4. The atmospheric humidity profile inversion method based on the Unet3+ neural network according to claim 1, characterized in that: In step S2, radiometric calibration, projection conversion, and quality control are performed on the original satellite-observed multi-channel brightness temperature, including: S21, performing radiometric calibration on the original data to obtain brightness temperature results of multiple channels; S22. Construct a standard grid with a spatial resolution of 0.25°×0.25°, and fill the original data in swath format into the standard grid according to longitude and latitude; S23. Perform quality control on the data after radiometric calibration and reprojection, including removing edge effects and outliers.

5. The atmospheric humidity profile inversion method based on the Unet3+ neural network according to claim 1, characterized in that: In step S3, the dataset is prepared, including: S31, cropping the brightness temperature after calibration, reprojection, and quality control in step S2, marking the imaging time, setting the time window to ±30 minutes, and rounding the time; S32, selecting label data according to the rounded time of the microwave hygrometer, selecting the specific humidity of the ERA-5 multi-pressure layer corresponding to the hygrometer imaging time, and cutting it into the same spatial range as after the hygrometer processing; S33. Divide the dataset into training set, validation set and test set.

6. The atmospheric humidity profile inversion method based on the Unet3+ neural network according to claim 1, characterized in that: In step S4, the humidity profile is inverted, including: According to the humidity profile inversion model constructed in step S1, the model parameters are continuously adjusted until the loss curve of the model is fitted and the value approaches 0, and the humidity profile inversion error is within 1.5g / kg.

Images