Atmospheric boundary layer height inversion method based on UNet + + network and SimAM attention mechanism

Abstract

The invention relates to an atmospheric boundary layer height inversion method based on a UNet + + network and a SimAM attention mechanism, and the method comprises the steps: firstly, making an image segmentation data set according to the obtained micro-pulse laser radar data and microwave radiometer data, in the data set, enabling an original image to be a gray-scale map of a normalized relative back scattering signal, and enabling the original image to be a gray-scale map of a normalized relative back scattering signal; the annotation image is a binary image of the height of the atmospheric boundary layer obtained through inversion according to data of the microwave radiometer, and meanwhile, the original image and the annotation image need to be matched in time and space; then, taking UNet + + as a basic framework, introducing a SimAM attention mechanism into jump connection in the basic framework, constructing a deep learning model SimAM-UNet + +, and training the deep learning model by using a data set; and finally, the predicted atmospheric boundary layer height can be obtained through the trained deep learning model. According to the method, the SimAM attention mechanism is introduced into UNet + +, and effective inversion of the atmospheric boundary layer height under the complex weather condition is realized.

Description

Technical Field

[0001] The present invention relates to a method for retrieving the atmospheric boundary layer height, belonging to the technical field of computer vision. Background Art

[0002] The Arctic is one of the cold sources of the global climate, affecting the global-scale atmospheric circulation and climate variability. At the same time, it is also an important gathering place for greenhouse gases and atmospheric pollutants, and its feedback to environmental changes is extremely sensitive. In recent years, due to the large emissions of greenhouse gases, the warming rate in the Arctic region is more than twice that in the low-latitude regions, and this phenomenon is also known as "Arctic amplification". Through research, there are various reasons for the occurrence of Arctic amplification, such as the local radiation effect caused by greenhouse gases, the change of Arctic cloud cover, the change of water vapor content, etc., and these climate factors can be reflected in the change of the atmospheric boundary layer height.

[0003] The planetary boundary layer (PBL) is the lowest part directly related to the Earth's surface in the atmosphere. It is strongly affected by surface elements and can respond within a short time. Radiosondes and microwave radiometers are often used to detect the planetary boundary layer height (PBLH). However, these two instruments have disadvantages such as limited observation frequency, sparse spatial coverage, and low vertical resolution. Therefore, more scholars hope to use lidar with stable measurement data and high spatio-temporal resolution to achieve long-term observation of PBLH.

[0004] Different from radiosondes and microwave radiometers that determine PBLH from a thermodynamic perspective, lidar analyzes PBLH from the perspective of particle scattering. Aerosol is the main observation object of lidar, and its concentration will drop sharply at the top of the PBL, which is reflected in the data as a rapid attenuation of the backscatter signal. However, a rapid attenuation of the backscatter signal also occurs at the top of clouds and suspended aerosols, which is the most severe problem faced in retrieving PBLH using lidar.

[0005] Currently, traditional lidar-based methods for retrieving the atmospheric boundary layer height invert PBLH based on the data change at a certain time point. The more representative ones are the gradient method, the standard deviation method, the curve fitting method, and the wavelet covariance change method, etc. However, these methods are easily interfered by noise and are not applicable to complex weather conditions such as cloudy and foggy days. Summary of the Invention

[0006] To solve the problems existing in the prior art, the present invention proposes a deep learning method that introduces the SimAM attention mechanism into UNet++ to effectively retrieve the PBLH under complex weather conditions.

[0007] To achieve the above object, the technical solution proposed by the present invention is: an atmospheric boundary layer height inversion method based on the UNet++ network and the SimAM attention mechanism, comprising the following steps:

[0008] Step 1: Obtain micropulse lidar data and microwave radiometer data. Obtain a grayscale image of the normalized relative backscattering signal from the micropulse lidar data, and use this grayscale image as the original image; and retrieve the atmospheric boundary layer height data from the microwave radiometer data. Spatially and temporally match the atmospheric boundary layer height data with the original image to obtain an annotated image. The annotated image and the original image together form a dataset.

[0009] Step 2: Use UNet++ as the basic architecture, and introduce the SimAM attention mechanism into the skip connections inside the basic architecture to construct the deep learning model SimAM-UNet++.

[0010] Step 3: Use the dataset to train the deep learning model to obtain a trained deep learning model.

[0011] Step 4: Input the grayscale image to be inverted obtained from the micropulse lidar into the trained deep learning model to obtain the predicted atmospheric boundary layer height.

[0012] A further design of the above technical solution is: in the first step, according to the potential temperature profile in the microwave radiometer data, the atmospheric boundary layer height data is retrieved using the parcel method or the temperature gradient method.

[0013] Based on the potential temperature profile, if the potential temperature above the surface is higher than the surface potential temperature, the temperature gradient method is used for inversion, that is, the maximum gradient is retrieved on the potential temperature profile above the height of the surface inversion layer, and the height where the maximum gradient is located is the atmospheric boundary layer height; if the potential temperature above the surface is lower than the surface potential temperature, the parcel method is used for inversion, that is, the height at which the potential temperature profile first exceeds the surface potential temperature is used as the atmospheric boundary layer height.

[0014] In the fourth step, after cropping, denoising, and positioning processing of the output result of the deep learning model, the atmospheric boundary layer height is obtained.

[0015] The beneficial effects of the present invention are:

[0016] Based on only the micropulse lidar data, from the perspective of computer vision, this invention predicts the atmospheric boundary layer height in the way of image segmentation. In view of the phenomenon that weather conditions such as clouds, suspended aerosol layers and haze will interfere with the inversion of the atmospheric boundary layer height, this invention introduces the SimAM attention mechanism into UNet++, realizing the effective inversion of the atmospheric boundary layer height under complex weather conditions.

[0017] The deep learning model constructed by this invention is based on UNet++ as the basic architecture, and the SimAM attention mechanism is introduced in the skip connection. It can generate attention weights by calculating local self-similarity, effectively reducing the weights of noise features while maintaining the attention to target features, thus improving the stability and accuracy of the model under poor data conditions. Brief Description of the Drawings

[0018] Figure 1 is the flow chart of this invention;

[0019] Figure 2 is the schematic diagram of the production of the dataset images in this invention;

[0020] Figure 3 is the schematic diagram of splitting the grayscale image;

[0021] Figure 4 is the schematic diagram of the PBLH inversion method based on the microwave radiometer;

[0022] Figure 5 is the UNet++ structure diagram;

[0023] Figure 6 is the pytorch implementation code diagram of SimAM;

[0024] Figure 7 is the schematic diagram of the deep learning model of this invention;

[0025] Figure 8 is the schematic diagram of the use of the deep learning model;

[0026] Figure 9 is the schematic diagram of the image processing process;

[0027] Figure 10 is the scatter plot of the inversion effects of different methods. Detailed Embodiments

[0028] The following combines the drawings and specific embodiments to elaborate on this invention in detail.

[0029] Embodiment 1

[0030] An atmospheric boundary layer height inversion method in this embodiment, as Figure 1 shown, includes the following steps:

[0031] Step 1: Obtain the micropulse lidar data and the microwave radiometer data. Generate a grayscale image of the normalized relative backscattering signal from the micropulse lidar data, and use this grayscale image as the original image. Also, invert the microwave radiometer data to obtain the atmospheric boundary layer height data, and perform spatio-temporal matching on the atmospheric boundary layer height data and the grayscale image to obtain an annotated image. The annotated image and the original image together form a dataset.

[0032] Step 2: Use UNet++ as the basic architecture, and introduce the SimAM attention mechanism in the skip connections inside the basic architecture to construct the deep learning model SimAM-UNet++.

[0033] Step 3: Use the dataset to train the deep learning model to obtain a trained deep learning model.

[0034] Step 4: Input the grayscale image to be inverted obtained from the micropulse lidar into the trained deep learning model to obtain the predicted atmospheric boundary layer height.

[0035] Embodiment 2

[0036] The method for inverting the atmospheric boundary layer height in this embodiment is as Figure 1 shown, and includes the following steps:

[0037] Step 1: Make a dataset.

[0038] The dataset is made as Figure 2 shown. Based on the micropulse lidar data, draw the dataset image, that is, with time as the horizontal axis and height as the vertical axis, draw a grayscale map of the normalized relative backscattering signal (where "49" refers to the data points between 0.08 - 1.52 km, and "2880" represents the data volume for 24 hours). At the same time, based on the atmospheric boundary layer height inverted from the microwave radiometer data, draw an annotated image corresponding to the dataset image. This dataset contains all available data from the US ARM NSA weather station from September 2017 to August 2023. Among them, the data for the first four years are used as the training set, and the data for the last two years are used as the test set.

[0039] In this embodiment, first, based on the data of the micropulse lidar, a grayscale map of the normalized relative backscattering signal is drawn. At the same time, according to the PBLH result retrieved from the microwave radiometer data, an annotation image that is spatio-temporally matched with the grayscale map is drawn. In order to reduce the waste of computer computing power in subsequent experiments and ensure that each picture has sufficient information, in the present invention, the original 49×2880 image is evenly divided into three parts (where "49" refers to the data points between 0.08 - 1.52 km, and "2880" represents the data volume for 24 hours). The specific operation is as Figure 3 shown.

[0040] The PBLH retrieval method based on the microwave radiometer is as Figure 4 shown. The core idea of this method is to retrieve the PBLH using the parcel method (PM) or the temperature gradient method (TGM) respectively according to the shape of the potential temperature profile. If the potential temperature θ(z n ) above the ground surface is higher than the surface potential temperature θ(z0), then the TGM is used for retrieval, that is, the maximum gradient is retrieved on the potential temperature profile (and the retrieval range must be above the height of the surface inversion layer (SBI)), and the height where the maximum gradient is located is the PBLH; if the potential temperature θ(z n ) above the ground surface is lower than the surface potential temperature θ(z0), then the PM is used for retrieval, that is, the height at which the potential temperature profile first exceeds the surface potential temperature is defined as the PBLH.

[0041] Step 2: Construct a deep learning model.

[0042] In this embodiment, UNet++ is used as the basic architecture, and the SimAM attention mechanism is introduced in the internal skip connections. SimAM attention can generate attention weights by calculating local self-similarity, thereby highlighting details such as edges and textures in the image, enabling the model to still maintain attention to the target features even under poor data conditions. The specific structure of SimAM-UNet++ is as Figure 7 shown.

[0043] The structure of UNet++ in this embodiment is as Figure 5 shown, where black represents the original U-Net, green and blue represent the dense convolutional blocks on the skip paths, and red represents the deep supervision. UNet++ is essentially a deep-supervised encoder-decoder network. The encoder is responsible for converting the input image into features, and the decoder is responsible for restoring these features to the output result. Between the encoder and the decoder, the deep feature information and the shallow feature information are fused through skip connections, so as to combine features of different scales to improve the segmentation accuracy.

[0044] SimAM is a parameter-free attention mechanism applicable to convolutional neural networks. Its core idea is to infer the 3D attention weights of feature maps by optimizing the energy function (the energy function is shown in formula (1)), thereby improving the network's representation ability. It can analyze the differences between each pixel and its surrounding pixels to determine the importance of that pixel. Furthermore, it highlights details such as edges and textures in the image, treats them as important features and assigns higher weights, enabling the model to pay more attention to these details during the reconstruction process, thus enhancing the quality of the reconstructed image. The PyTorch implementation of SimAM is as Figure 6 shown below.

[0045]

[0046] Step 3: Train the deep learning model.

[0047] The deep learning model constructed in this embodiment is based on UNet++. The SimAM attention mechanism is introduced in the skip connections. It can generate attention weights by calculating local self-similarity, effectively reducing the weights of noise features while maintaining attention to target features, thereby improving the stability and accuracy of the model under poor data conditions. The structure of the deep learning model constructed in this embodiment is as Figure 7 shown. Since this model involves a large number of convolutional layers and a large number of datasets, training this model requires a good GPU. The GPU used in this embodiment for training the model is NVIDIA GeForce RTX 4090 with 24G video memory. The trained model can be used for the inversion of PBLH. The process is schematically shown as Figure 8 shown. After inputting the grayscale image obtained by the micropulse lidar into the model, the result shown in Figure 9 (a) will be obtained. After cropping, denoising, and positioning processing, the specific PBLH result is obtained, as shown in Figure 9 (d).

[0048] Cropping: During the model training process, the input image of 49×960 will be padded to 960×960 and stretched to 1024×1024. Therefore, the result obtained from the model cannot be directly used and needs to be cropped and scaled back to the original size.

[0049] Denoising: The denoising method adopted in the present invention is morphological opening (erosion followed by dilation). This is a simple and effective image processing method that can be used to eliminate small noise points, fill small holes, and smooth boundaries.

[0050] Positioning: The results after cropping and denoising are shown in Figure 9 (b) and Figure 9 (c). According to the boundary line of black and white pixels in the figure, the specific PBLH can be determined, that is, Figure 9The white line in (d).

[0051] In this embodiment, the deep learning model is trained with a dataset image for 50 rounds. During the training process, the weights of each round are saved. After the training is completed, the saved weights are applied to the test set, and the optimal weights are selected as the training result of the model according to the inversion effect. The model shows overfitting after 20 rounds, and the inversion effect gradually deteriorates and then stabilizes. The best number of training rounds obtained from the experiment is 16 rounds.

[0052] Step 4: Evaluate the inversion result.

[0053] Under the same test set, different inversion methods for the atmospheric boundary layer height are tested. The inversion effect of the trained SimAM-UNet++ method in this embodiment is as shown in Figure 10 (a). Compared with the UNet++ method in Figure 10 (b), the inversion result is more stable, and R 2 is also improved. While the two traditional methods shown in Figure 10 (c) and (d) are not applicable to the climate environment of ARM NSA in the United States.

[0054] Step 5: Application and deployment.

[0055] Applying the optimal prediction model to the actual micro-pulse lidar backscatter data can obtain hourly predictions of the atmospheric boundary layer height, and the prediction results can be used for other meteorological studies.

[0056] The technical solution of the present invention is not limited to the above embodiments, and all technical solutions obtained by equivalent replacement fall within the scope of protection required by the present invention.

Claims

1. An atmospheric boundary layer height inversion method based on the UNet++ network and the SimAM attention mechanism, characterized in that, It includes the following steps: Step 1: Obtain the micro-pulse lidar data and the microwave radiometer data. From the micro-pulse lidar data, obtain the grayscale image of the normalized relative backscattering signal as the original image; and inversely derive the atmospheric boundary layer height data from the microwave radiometer data, match the atmospheric boundary layer height data with the original image in space and time to obtain the labeled image, and the labeled image and the original image together form a data set; Step 2: Use UNet++ as the basic architecture, and introduce the SimAM attention mechanism in the skip connections inside the basic architecture to construct the deep learning model SimAM-UNet++; Step 3: Use the data set to train the deep learning model to obtain the trained deep learning model; Step 4: Input the grayscale image to be inverted obtained from the micro-pulse lidar into the trained deep learning model to obtain the predicted atmospheric boundary layer height.

2. The method for retrieving the atmospheric boundary layer height based on the UNet++ network and the SimAM attention mechanism according to claim 1, wherein: In the said Step 1, according to the potential temperature profile in the microwave radiometer data, the atmospheric boundary layer height data is inversely derived by using the parcel method or the temperature gradient method.

3. The method for retrieving the atmospheric boundary layer height based on the UNet++ network and the SimAM attention mechanism according to claim 2, wherein: Based on the potential temperature profile, if the potential temperature above the ground surface is higher than the potential temperature of the ground surface, the temperature gradient method is used for inversion, that is, the maximum gradient is retrieved on the potential temperature profile above the height where the surface inversion layer is located, and the height where the maximum gradient is located is the atmospheric boundary layer height; if the potential temperature above the ground surface is lower than the potential temperature of the ground surface, the parcel method is used for inversion, that is, the height at which the potential temperature profile first exceeds the potential temperature of the ground surface is used as the atmospheric boundary layer height.

4. The method for retrieving the atmospheric boundary layer height based on the UNet++ network and the SimAM attention mechanism according to claim 1, wherein: In the said Step 4, after performing cropping, denoising and positioning processing on the output result of the deep learning model, the atmospheric boundary layer height is obtained.

Images