A coupling method for improving the resolution and accuracy of surface temperature retrieval in desert areas

Abstract

The application discloses a coupling method for improving surface temperature inversion resolution and precision in a desert area, and comprises the following steps: (1) obtaining fine classification of a surface type based on a trained surface fine classification model, and then calculating first high-resolution surface specific radiation rate data; (2) calculating adjusted Landsat emissivity, and then calculating second high-resolution surface specific radiation rate data; (3) using linear weighted fusion to give different weights to the first and second high-resolution surface specific radiation rates to obtain high-resolution surface specific radiation rate data; (4) calculating atmospheric water vapor content; (5) generating a high-resolution thermal infrared image based on a generative adversarial network algorithm; (6) calculating high-resolution radiation brightness and brightness temperature according to DN values of the high-resolution thermal infrared image, and simultaneously performing surface temperature inversion based on all high-resolution parameter data and a single-channel algorithm to finally obtain high-precision and high-resolution surface temperature data.

Description

Technical Field

[0001] This invention belongs to the field of land surface temperature inversion, and specifically relates to a coupling method for improving the resolution and accuracy of land surface temperature inversion in desert areas. Background Technology

[0002] On a global scale, ecological and environmental issues are increasingly prominent, becoming a core topic of common concern for all countries. Land surface temperature (LST) is an extremely important physical quantity that reflects the thermal balance between the atmosphere and the land surface, serving as a crucial indicator for measuring the Earth's natural ecological environment. Furthermore, LST carries a wealth of environmental information, such as dynamic changes in land cover and water evaporation rates. This information is vital for in-depth research into ecological and environmental evolution and the development of protection strategies. Its acquisition plays a pivotal role in various aspects of global climate change monitoring, disaster prevention, weather forecasting, and urban development.

[0003] Traditional LST acquisition methods mainly rely on temperature measurement equipment at ground measurement stations. Although this method can provide some high-precision data and obtain good temperature inversion results, its high cost and data limitations (such as uneven distribution) limit its application.

[0004] Currently, with the rapid development of satellite remote sensing technology, obtaining LST (Landsight and Surface Temperature) using remote sensing has become a new approach. Compared with traditional measurement methods, remote sensing technology has many advantages. It can cover a wider measurement area, provide more comprehensive LST information, and save on high costs. However, commonly used single-channel, multi-channel, or split-window algorithms also have many disadvantages. First, there is the inaccuracy of parameters. Parameters such as surface emissivity and atmospheric water vapor content are difficult to obtain and are easily affected by factors such as changes in land cover type and atmospheric conditions, resulting in uncertain and low-precision data. Second, real physical models are very complex, but inversion algorithms simplify the model to simplify calculations and improve efficiency, which may also lead to errors. Finally, the existing algorithms have limited application scenarios. Single-channel algorithms are more suitable for simple atmospheric conditions and land surface types, while multi-channel or split-window algorithms are more suitable for complex scenarios. Therefore, it is necessary to develop an inversion method that can be tailored to specific scenarios to effectively improve inversion accuracy. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by providing a coupling method to improve the resolution and accuracy of surface temperature inversion in desert areas, which can effectively improve the accuracy and resolution of surface temperature inversion results in desert areas.

[0006] To address the above technical problems, this invention provides the following technical solution: a coupling method for improving the resolution and accuracy of surface temperature inversion in desert regions, comprising the following steps:

[0007] S1. Construct and train a surface fine classification model U-BiFormer using several remote sensing images of desert areas and corresponding data with fine classification labels for land surface types. Use the surface fine classification model U-BiFormer to perform fine classification of land surface types on the remote sensing images of desert areas, and then calculate the first high-resolution surface emissivity data based on the fine classification of land surface types.

[0008] S2. Classify the pixels in the image according to the normalized vegetation index to obtain the vegetation coverage; calculate the bare soil emissivity of the ASTER data in the preset band by combining the reflectivity of the ASTER data, and obtain the emissivity of the Landsat series satellite thermal infrared band by conversion. Then, after vegetation adjustment and interpolation, the adjusted Landsat emissivity is obtained. Finally, the second high resolution surface emissivity data is calculated.

[0009] S3. Using a linear weighted fusion method, the first high-resolution surface emissivity data and the second high-resolution surface emissivity data are assigned different weights, and the third high-resolution surface emissivity data is obtained after weighted calculation.

[0010] S4. Using low-resolution NCEP atmospheric water vapor content data, perform spatial and temporal linear interpolation on the data to obtain high-resolution atmospheric water vapor content data, and then calculate the atmospheric function.

[0011] S5. Based on the adversarial neural network algorithm, the SRGAN model is trained on the preprocessed thermal infrared dataset. Through adversarial training between the generator and the discriminator, the generator is able to generate super-resolution thermal infrared images.

[0012] S6. Based on the DN value of the super-resolution thermal infrared image, the high-resolution radiance and luminance temperature are calculated. Finally, combined with the third high-resolution surface emissivity data and atmospheric functions, a single-channel algorithm is used to retrieve the surface temperature, ultimately obtaining high-precision, high-resolution surface temperature data.

[0013] Furthermore, the aforementioned step S1 includes the following sub-steps:

[0014] S1.1 Acquire high-resolution visible light data and select clear images from remote sensing images of desert areas; crop the images and add fine-grained classification labels for land surface types; then annotate the images and divide them into training, validation, and test sets;

[0015] S1.2. Add a U-shaped decoder to the BiFormer model to construct a deep learning-based fine-grained land classification network. Then set the initial parameters and use the CE+Focal hybrid loss function when training the deep learning-based fine-grained land classification network, as follows:

[0016] Loss = Loss CE +Loss Focal

[0017] Among them, the cross-entropy loss function Loss CE With focus loss function Loss Focal The formulas for calculating a single pixel in multi-category tasks are as follows:

[0018]

[0019]

[0020] Where, y i p represents the predicted true label of the sample. i FL(y) represents the probability that the model predicts a pixel belongs to the i-th category. i ,p i ) represents the focus loss of the i-th category, α is a balancing factor used to adjust the weights between the positive and negative categories, and γ is an adjustment parameter used to control the weights of easy and difficult samples;

[0021] S1.3 Set the training algorithm parameters and adjust the weight connection parameters according to the learning results of the loss function until the preset accuracy is reached to obtain the U-BiFormer surface fine classification model. Use the U-BiFormer surface fine classification model to perform fine classification of surface types on remote sensing images of desert areas to obtain the corresponding surface types, including towns, farmland, desert, water bodies, and mountains.

[0022] S1.4. Different surface emissivity values ​​are obtained for different land surface types after classification to obtain the first high-resolution surface emissivity data:

[0023]

[0024] Furthermore, the aforementioned step S2 includes the following sub-steps:

[0025] S2.1 Obtain low-resolution reflectance data for bands 13 and 14 in the ASTER dataset;

[0026] S2.2 The Normalized Difference Vegetation Index (NDVI) is calculated using near-infrared (NIR) and infrared (R) data, as shown in the following formula:

[0027]

[0028] S2.3. Based on the NDVI thresholds for soil and vegetation, the NDVI thresholds are used to classify pixels in the image using the vegetation index, dividing the pixels into pure water bodies, mixed pixels, and pure vegetation, and calculating the vegetation cover PV:

[0029]

[0030] S2.4 Determine the surface emissivity. When the NDVI is less than or equal to the soil threshold, the pixel is considered to belong to water; when the NDVI is greater than or equal to the vegetation threshold, the pixel is considered to belong to vegetation. For pixels classified as mixed pixels, the previously calculated PV value is used for further calculation to obtain the bare soil emissivity of ASTER bands 13 and 14, as shown in the following formula:

[0031]

[0032] Where, ε ASTER,bare ε represents the emissivity of bare soil in ASTER. ASTER ε represents the emissivity of the preset band. ASTER,veg This indicates the emissivity of ASTER's pure vegetation.

[0033] S2.5. Convert the bare soil emissivity of ASTER bands 13 and 14 to the thermal infrared emissivity of Landsat series satellites, as follows:

[0034] ε L10,bare =c0ε A13,bare +c1ε A14,bare +c2

[0035] Where, ε L10,bare ε represents the bare soil emissivity in band 10 of the Landsat TIRS sensor. A13,bare and ε A14,bare These represent the bare soil emissivity of ASTER bands 13 and 14 obtained using calculations;

[0036] S2.6. Based on the bare soil emissivity, pure vegetation emissivity, and PV, the vegetation-adjusted Landsat emissivity is calculated as follows:

[0037] ε=PV·ε L10,veg +(1-PV)·ε L10,bare

[0038] Where ε represents the surface emissivity of the 10th band, which is used as input for subsequent single-channel algorithms. L10,veg ε represents the pure vegetation emissivity of Landsat band 10. L10,bare This represents the bare soil emissivity of Landsat band 10;

[0039] S2.7. In the single-channel algorithm, the following formula is used for calculation:

[0040]

[0041] S2.8. Interpolate the acquired low-resolution surface emissivity data to the same resolution as the visible light data to obtain the second high-resolution surface emissivity ε2.

[0042] Furthermore, the aforementioned step S3 includes the following sub-steps:

[0043] S3.1. Using a linear weighted fusion method, assign different weights to the first high-resolution surface emissivity data and the second high-resolution surface emissivity data.

[0044] S3.2. Based on the set weights, the final high-resolution surface emissivity ε used in the single-channel algorithm is calculated according to the following formula:

[0045]

[0046] Where ε i The surface emissivity obtained in step i, w i This refers to the weight corresponding to the i-th step.

[0047] Furthermore, the aforementioned step S4 includes the following sub-steps:

[0048] S4.1. Based on NCEP atmospheric water vapor content data, time interpolation is performed on the data to obtain the water vapor content at the time of data capture, using the following formula:

[0049]

[0050] Where w represents the water vapor content at the time Landsat acquired the data. and t represents the NCEP water vapor content before and after the Landsat image capture time, t1 and t2 represent the NCEP atmospheric water vapor content before and after the Landsat image capture time, respectively.

[0051] S4.2. Spatial interpolation is performed on the atmospheric water vapor content data to interpolate the data to the target area. Then, the atmospheric functions f1, f2, and f3 are calculated using the obtained atmospheric water vapor content data, as shown in the following formulas:

[0052]

[0053] Among them, c ij It is a constant related to the sensor.

[0054] Furthermore, the aforementioned step S5 includes the following sub-steps:

[0055] S5.1 Construct a thermal infrared image dataset and preprocess it. The dataset includes low-resolution and high-resolution thermal infrared image data of different thermal infrared bands, as well as the corresponding labels.

[0056] S5.2 Construct an SRGAN model including a generator and a discriminator. Input a low-resolution thermal infrared image into the SRGAN model and train it adversarially to output a high-resolution thermal infrared image.

[0057] S5.3 Construct loss functions: content loss, adversarial loss, and perceptual loss. Content loss ensures that the generated image is as consistent as possible with the high-resolution image; adversarial loss improves the realism of the image by training the generator through a discriminator; perceptual loss focuses on the similarity of image detail features based on a pre-trained VGG network.

[0058] Content loss function:

[0059] Adversarial loss function:

[0060] Perceptual loss function: Among them, W i,j H is the width of the feature map. i,j It is the height of the feature map, -φ i,j It is the feature map obtained after the j-th convolutional activation before the i-th maxpooling layer in the VGG19 network, r 2 Indicates the scaling factor. This represents the loss function of the standard VGG, where I is the number of pixels, HR represents the real image, and LR represents the scaled image.

[0061] S5.4 Set the training parameters for the generator and discriminator, and start adversarial training of the SRGAN model. The model optimizes the weight parameters of the generator, generates high-resolution thermal infrared images, and refines the texture and radiation features of the crack region.

[0062] Furthermore, the aforementioned step S6 includes the following sub-steps:

[0063] S6.1 Convert the DN value of the thermal infrared band image to an absolute radiance value. The calculation formula is as follows:

[0064]

[0065] Where gain is the gain constant and bias is the bias constant;

[0066] S6.2 Calculate the luminance temperature based on the radiance value. The calculation formula is as follows:

[0067]

[0068] Where K1 and K2 are calibration constants for the thermal infrared band.

[0069] S6.3. Combining the third-highest resolution surface emissivity data and atmospheric functions, the surface temperature is calculated using a single-channel algorithm, as shown in the following formula:

[0070]

[0071] Where b is a quantity related to the equivalent wavelength; f1, f2, and f3 are atmospheric functions.

[0072] Compared to existing technologies, the beneficial technical effects of the present invention using the above technical solution are as follows: Firstly, the present invention establishes a fine-grained land surface classification model based on the U-BiFormer deep learning model, realizing automatic identification of land surface types corresponding to any visible light remote sensing data, solving the problem of low accuracy in current land surface emissivity data. Then, by combining the land surface emissivity calculation method in general single-channel algorithms with linear weighted fusion, the accuracy of land surface emissivity is further improved. Next, an SRGAN model is trained based on an adversarial neural network algorithm. Through adversarial training between the generator and discriminator, the generator is able to generate high-quality, detailed super-resolution thermal infrared images. Finally, based on high-resolution parameter data, a single-channel algorithm is used to calculate high-resolution, high-precision land surface temperature data, providing a new approach to improving the accuracy and resolution of land surface temperature inversion.

[0073] This invention is primarily applied to surface temperature retrieval in desert regions. By modifying the fine-grained surface classification model and the method for calculating surface emissivity, it can also be applied to other surface types. The obtained high-resolution, high-precision surface temperature data provides a theoretical foundation for global climate change monitoring, disaster prevention, weather forecasting, and urban development, demonstrating strong practicality and high application and promotion value. Attached Figure Description

[0074] Figure 1 Schematic diagram of the invention process.

[0075] Figure 2 Test image of the results of a fine-grained land surface classification model based on U-BiFormer.

[0076] Figure 3 Flowchart of a super-resolution reconstruction model based on adversarial neural network algorithm.

[0077] Figure 4Test results of a super-resolution reconstruction model based on an adversarial neural network algorithm. Detailed Implementation

[0078] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0079] In this invention, various aspects of the invention are described with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. Embodiments of the invention are not limited to those shown in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in this invention are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed can be used alone or in any suitable combination with other aspects of the invention disclosed.

[0080] refer to Figure 1 This invention discloses a coupling method for improving the resolution and accuracy of surface temperature inversion in desert regions, characterized by comprising the following steps:

[0081] S1. Construct and train a surface fine classification model U-BiFormer using several remote sensing images of desert areas and corresponding data with fine classification labels for land surface types. Use the surface fine classification model U-BiFormer to perform fine classification of land surface types on the remote sensing images of desert areas, and then calculate the first high-resolution surface emissivity data based on the fine classification of land surface types.

[0082] S2. Classify the pixels in the image according to the normalized vegetation index to obtain the vegetation coverage; calculate the bare soil emissivity of the ASTER data in the preset band by combining the reflectivity of the ASTER data, and obtain the emissivity of the Landsat series satellite thermal infrared band by conversion. Then, after vegetation adjustment and interpolation, the adjusted Landsat emissivity is obtained. Finally, the second high resolution surface emissivity data is calculated.

[0083] S3. Using a linear weighted fusion method, the first high-resolution surface emissivity data and the second high-resolution surface emissivity data are assigned different weights, and the third high-resolution surface emissivity data is obtained after weighted calculation.

[0084] S4. Using low-resolution NCEP atmospheric water vapor content data, perform spatial and temporal linear interpolation on the data to obtain high-resolution atmospheric water vapor content data, and then calculate the atmospheric function.

[0085] S5. Based on the adversarial neural network algorithm, the SRGAN model is trained on the preprocessed thermal infrared dataset. Through adversarial training between the generator and the discriminator, the generator is able to generate super-resolution thermal infrared images.

[0086] S6. Based on the DN value of the super-resolution thermal infrared image, the high-resolution radiance and luminance temperature are calculated. Finally, combined with the third high-resolution surface emissivity data and atmospheric functions, a single-channel algorithm is used to retrieve the surface temperature, ultimately obtaining high-precision, high-resolution surface temperature data.

[0087] Further, as a preferred embodiment of the present invention, in step S1, based on the U-BiFormer model, the outputs of all stages of the decoder are used to predict the segmentation map, the software environment required for running the U-BiFormer model is configured, and the dataset prepared in the previous step is input into the model for training. After training on a large amount of data, a U-BiFormer-based land surface classification model is obtained. The model is then used to classify the land surface of a visible light remote sensing image of a desert area, and the results are as follows. Figure 2 As shown in the diagram. The green areas represent farmland, the yellow areas represent desert, the blue areas represent water bodies, and the black areas represent towns. This example does not involve mountainous terrain. Specifically, step S1 includes the following sub-steps:

[0088] S1.1. High-resolution visible light data were acquired from the Gaofen-1 satellite, and clear images were selected from more than 1,700 remote sensing images of the African region. The images were cropped to 512*512 pixels to facilitate model training. Then, 1,000 images for each land surface category were selected and labeled using the Labelme image annotation tool. The training set, validation set, and test set were divided into three sets in an 8:1:1 ratio.

[0089] S1.2. Based on the BiFormer model, a U-shaped decoder is added to construct a deep learning-based fine-grained land surface classification network. This network is a novel U-shaped Transformer model, U-BiFormer. Based on BiFormer, a U-shaped decoder is used, and the output of the decoder at all stages is used to predict the segmentation map. This can improve the model's ability to capture details and contextual information in the image, enabling the model to better segment similar categories. The hybrid attention module unique to the U-shaped decoder is improved by increasing the proportion of the current stage features in the hybrid features, allowing the decoder to focus more on refining the current stage features and improving the model's segmentation effect on similar categories.

[0090] Next, configure the software environment required for running the U-BiFormer model, and input the dataset prepared in step S1.1 into the model for training. The initial training parameters are shown in Table 1 below:

[0091] Table 1

[0092]

[0093] When training a deep learning network for fine-grained land classification, a hybrid loss function of CE and Focal is used, as follows:

[0094] Loss = Loss CE +Loss Focal

[0095] Among them, the cross-entropy loss function Loss CE With focus loss function Loss Focal The formulas for calculating a single pixel in multi-category tasks are as follows:

[0096]

[0097] Where, y i p represents the predicted true label of the sample. i FL(y) represents the probability that the model predicts a pixel belongs to the i-th category. i ,p i ) represents the focus loss of the i-th category, α is a balancing factor used to adjust the weights between the positive and negative categories, and γ is an adjustment parameter used to control the weights of easy and difficult samples.

[0098] S1.4 Set the training algorithm parameters as shown in Table 2. Adjust the weighted connection parameters according to the learning results of the loss function until the preset accuracy is reached to obtain the U-BiFormer surface fine classification model. Use the U-BiFormer surface fine classification model to perform fine classification of surface types on remote sensing images of desert areas to obtain the corresponding surface types, including towns, farmland, desert, water bodies, and mountains.

[0099] Table 2

[0100]

[0101]

[0102] S1.5. Take different surface reflectance values ​​for different classified surface types to obtain the first high-resolution surface reflectance data:

[0103]

[0104] In this embodiment, the environment settings for the U-BiFormer algorithm, a fine-grained land surface classification model, are as follows:

[0105] Enter in the terminal

[0106] python"u-biformer / tools / train.py"

[0107] "u-biformer / configs / u-biformer / u-biformer____-512x512.py"

[0108] The model can be trained using training scripts.

[0109] Third-party packages required for Python

[0110] 1. numpy

[0111] 2.scipy

[0112] 3. Pillow

[0113] 4. cython

[0114] 5. matplotlib

[0115] 6.mmcv>=2.0.0rc1,<2.1.0

[0116] 7.mmengine >= 0.4.0, < 1.0.0

[0117] 8. prettytable

[0118] 9.torch

[0119] 10.torchvision

[0120] 11.scikit-image

[0121] 12.keras>=2.0.8

[0122] 13. OpenCV-Python

[0123] 14.h5py

[0124] 15.imgaug

[0125] 16.codecov

[0126] 17.flake8

[0127] 18. interrogate

[0128] 19.pytest

[0129] 20.xdoctest>=0.10.0

[0130] 21.Yapf

[0131] Furthermore, as a preferred embodiment of the present invention, step S2 includes the following sub-steps:

[0132] S2.1 Obtain low-resolution reflectance data for bands 13 and 14 in the ASTER dataset;

[0133] S2.2 The Normalized Difference Vegetation Index (NDVI) is calculated using near-infrared (NIR) and infrared (R) data, as shown in the following formula:

[0134]

[0135] S2.3. Based on the NDVI thresholds for soil and vegetation, the NDVI thresholds are used to classify pixels in the image using the vegetation index, dividing the pixels into pure water bodies, mixed pixels, and pure vegetation, and calculating the vegetation cover PV:

[0136]

[0137] S2.4 Determine the surface emissivity. When the NDVI is less than or equal to the soil threshold, the pixel is considered to belong to water, and the emissivity is set to 0.995; when the NDVI is greater than or equal to the vegetation threshold, the pixel is considered to belong to vegetation, and the emissivity is set to 0.986. For pixels classified as mixed pixels, the previously calculated PV value is used for further calculation to obtain the bare soil emissivity of ASTER bands 13 and 14, as shown in the following formula:

[0138]

[0139] Where, ε ASTER,bare ε represents the emissivity of bare soil in ASTER. ASTER ε represents the emissivity of the preset band. ASTER,veg ε represents the emissivity of ASTER pure vegetation. ASTER,veg =0.986;

[0140] S2.5. Convert the bare soil emissivity of ASTER bands 13 and 14 to the thermal infrared emissivity of Landsat series satellites, as follows:

[0141] ε L10,bare =c0ε A13,bare +c1ε A14,bare +c2

[0142] Where, ε L10,bare ε represents the bare soil emissivity in band 10 of the Landsat TIRS sensor. A13,bare and ε A14,barec0 = 0.6820, c1 = 0.2578, and c2 = 0.0584 represent the bare soil emissivity of ASTER bands 13 and 14, respectively, and are regression coefficients.

[0143] S2.6. Based on the bare soil emissivity, pure vegetation emissivity, and PV, the vegetation-adjusted Landsat emissivity is calculated as follows:

[0144] ε=PV·ε L10,veg +(1-PV)·ε L10,bare

[0145] Where ε represents the surface emissivity of the 10th band, which is used as input for subsequent single-channel algorithms. L10,veg ε represents the pure vegetation emissivity of Landsat band 10. L10,veg =0.986, ε L10,bare This represents the bare soil emissivity of Landsat band 10.

[0146] S2.7. In the single-channel algorithm, the following formula is used for calculation:

[0147] ε2

[0148] S2.8. Interpolate the acquired low-resolution surface emissivity data to the same resolution as the visible light data to obtain the second high-resolution surface reflectivity ε2.

[0149] Furthermore, as a preferred embodiment of the present invention, step S3 includes the following sub-steps:

[0150] S3.1. Use a linear weighted fusion method to assign different weights to the first high-resolution surface reflectance data and the second high-resolution surface reflectance data.

[0151] S3.2. Based on the set weights, the final high-resolution surface emissivity ε used in the single-channel algorithm is calculated according to the following formula:

[0152]

[0153] Where ε i w refers to the surface emissivity obtained in step i. i The weights corresponding to the i-th step are w1 = 0.3 and w2 = 0.7.

[0154] Furthermore, as a preferred embodiment of the present invention, step S4 includes the following sub-steps:

[0155] S4.1. Download atmospheric water vapor content data from the NCEP website, perform time interpolation on the data to obtain the water vapor content at the time the data was captured, using the following formula:

[0156]

[0157] Where w represents the water vapor content at the time Landsat acquired the data. and t represents the NCEP water vapor content before and after the Landsat image capture time, t1 and t2 represent the NCEP atmospheric water vapor content before and after the Landsat image capture time, respectively.

[0158] S4.2. Spatial interpolation is performed on the atmospheric water vapor content data to interpolate the data to the target area. Then, the atmospheric functions f1, f2, and f3 are calculated using the obtained atmospheric water vapor content data, as shown in the following formulas:

[0159]

[0160] Among them, c ij It is a constant related to the sensor.

[0161] The specific values ​​are shown in Table 3 below.

[0162] Table 3

[0163]

[0164] Further, refer to Figure 3 In a preferred embodiment of the present invention, step S5 includes the following sub-steps:

[0165] S5.1 Constructing and preprocessing a thermal infrared image dataset: The dataset includes over 20,000 thermal infrared images of 30×30 low resolution and 15×15 high resolution, primarily containing data from different thermal infrared bands to ensure data diversity and generalization ability; then, image preprocessing is performed; for the collected thermal infrared images, noise removal, image normalization, and appropriate magnification, rotation, and cropping are performed to ensure consistent image quality after processing; label files for training are generated according to the corresponding low-resolution and high-resolution images, and the data is divided into an 8:2 training set and test set. Subsequently, the training set is further divided using k-fold cross-validation to ensure the stability of model performance;

[0166] S5.2 Construct an SRGAN model including a generator and a discriminator. Input a low-resolution thermal infrared image into the SRGAN model and output a high-resolution thermal infrared image through adversarial training: Configure the software environment required for running the SRGAN model, and input the dataset prepared in step S5.1 into the SRGAN model for training. The model consists of a generator and a discriminator. Through adversarial training, the generator gradually learns to reconstruct a high-resolution image from a low-resolution image. Content Loss is calculated in the generator to measure the similarity in content between the generated image and the target high-resolution image. The generated high-resolution image is input into the discriminator model and compared with a real high-resolution image to help the generator model learn how to better generate near-realistic high-resolution images. Adversarial Loss is calculated in the discriminator to evaluate the difference between the generated image and the real image, thereby guiding the generator model to improve. The entire process optimizes the output of the generator model through feedback from content loss and adversarial loss, gradually improving the low-resolution image to a high-resolution image. A thermal infrared image of the study area is used for result testing. The test results are as follows: Figure 4 As shown in the figure, (a) shows the thermal infrared image before reconstruction, and (b) shows the thermal infrared image after reconstruction.

[0167] The environment settings for the algorithm in this embodiment are as follows:

[0168] `python train.py --mode=train"` # Train the model based on the dataset and obtain the model parameters.

[0169] "python train.py --mode=eval" # Perform inference or testing, generate high-resolution images. Python requires the installation of third-party packages.

[0170] 1. numpy

[0171] 2.cv2

[0172] 3. SSL

[0173] 4. TensorLayerX 0.5.8

[0174] 5.torch 1.10.1+cu111

[0175] 6. IPython [all] Python 3.7.3

[0176] The initial training parameters are shown in Table 4 below:

[0177] Table 4 Initial Training Parameters

[0178]

[0179]

[0180] S5.4 Construct loss functions: content loss, adversarial loss, and perceptual loss. Content loss ensures that the generated image is as consistent as possible with the high-resolution image; adversarial loss improves the realism of the image by training the generator through a discriminator; perceptual loss, based on a pre-trained VGG network, focuses on the similarity of image detail features to achieve realistic detail representation.

[0181] Content loss function:

[0182] Adversarial loss function:

[0183] Perceptual loss function:

[0184] Among them, W i,j H is the width of the feature map. i,j It is the height of the feature map, -φ i,j It is the feature map obtained after the j-th convolutional activation before the i-th maxpooling layer in the VGG19 network, r 2 G represents the scaling factor. θG This represents the loss function of the standard VGG, where I is the number of pixels, HR represents the real image, and LR represents the scaled image.

[0185] S5.5 Set the training parameters for the generator and discriminator, and begin adversarial training of the SRGAN model. The model gradually optimizes the weight parameters of the generator to generate high-resolution thermal infrared images and refines the texture and radiation features of the crack area, making the reconstructed image visually realistic and rich in detail.

[0186] Furthermore, as a preferred embodiment of the present invention, step S6 includes the following sub-steps:

[0187] S6.1 Convert the DN value of the thermal infrared band image to an absolute radiance value. The calculation formula is as follows:

[0188]

[0189] Where gain = 0.0003342 is the gain constant, and bias = 0.1 is the bias constant;

[0190] S6.2 Calculate the luminance temperature based on the radiance value. The calculation formula is as follows:

[0191]

[0192] Wherein, K1 and K2 are calibration constants for the thermal infrared band, K1 = 774.8853 W / m·μm·sr and K2 = 1321.0789 K;

[0193] S6.3. Combining the third-high resolution surface reflectance data and atmospheric functions, a single-channel algorithm is used to calculate the surface temperature, resulting in a high-precision, high-resolution surface temperature result, as shown in the following formula:

[0194]

[0195] Where b is a quantity related to the equivalent wavelength, b = 1324; f1, f2, and f3 are atmospheric functions.

[0196] Based on the single-channel algorithm, an innovation was made in the calculation of surface emissivity during the algorithm process. Surface emissivity was obtained using two methods: (1) Based on the U-BiFormer surface fine classification model, the visible light remote sensing data was finely classified, and then different surface emissivity values ​​were set for different surface types; (2) The normalized vegetation index was calculated using infrared and near-infrared data, and then the surface emissivity data was inferred by combining the reflectance of ASTER 13 and 14 bands. Finally, the linear weighted fusion method was used to assign weights of 0.3 and 0.7 to the results obtained by the two methods, respectively, and the final surface emissivity data was obtained by weighted calculation, which improved the accuracy of this parameter and thus improved the accuracy of the surface temperature inversion results.

[0197] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A coupling method for improving the resolution and accuracy of surface temperature inversion in desert regions, characterized in that, Includes the following steps: S1. Construct and train a surface fine classification model U-BiFormer using several remote sensing images of desert areas and corresponding data with fine classification labels for land surface types. Use the surface fine classification model U-BiFormer to perform fine classification of land surface types on the remote sensing images of desert areas, and then calculate the first high-resolution surface emissivity data based on the fine classification of land surface types. S2. Classify the pixels in the image according to the normalized vegetation index to obtain the vegetation coverage; calculate the bare soil emissivity of the ASTER data in the preset band by combining the reflectivity of the ASTER data, and obtain the emissivity of the Landsat series satellite thermal infrared band by conversion; then, after vegetation adjustment and interpolation, calculate the adjusted Landsat emissivity; finally, calculate the second high-resolution surface emissivity data; S3. Use the linear weighted fusion method to assign different weights to the first high-resolution surface emissivity data and the second high-resolution surface emissivity data, and obtain the third high-resolution surface emissivity data after weighted calculation. S4. Using low-resolution NCEP atmospheric water vapor content data, perform spatial and temporal linear interpolation on the data to obtain high-resolution atmospheric water vapor content data, and then calculate the atmospheric function. S5. Based on the adversarial neural network algorithm, the SRGAN model is trained on the preprocessed thermal infrared dataset. Through adversarial training between the generator and the discriminator, the generator is able to generate super-resolution thermal infrared images. S6. Based on the DN value of the super-resolution thermal infrared image, the high-resolution radiance and luminance temperature are calculated. Finally, combined with the third high-resolution surface emissivity data and atmospheric functions, a single-channel algorithm is used to retrieve the surface temperature, ultimately obtaining high-precision, high-resolution surface temperature data.

2. The coupling method for improving the resolution and accuracy of surface temperature inversion in desert areas according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.1 Acquire high-resolution visible light data and select clear images from remote sensing images of desert areas; crop the images and add fine-grained classification labels for land surface types; then annotate the images and divide them into training, validation, and test sets; S1.

2. Add a U-shaped decoder to the BiFormer model to construct a deep learning-based fine-grained land classification network. Then set the initial parameters and use the CE+Focal hybrid loss function when training the deep learning-based fine-grained land classification network, as follows: ， Among them, the cross-entropy loss function Loss CE With focus loss function Loss Focal The formulas for calculating a single pixel in multi-category tasks are as follows: ， ， ， in, This represents the true label of the predicted sample. This represents the probability that the model predicts a pixel belongs to the i-th category. This represents the focus loss for the i-th category. It is a balancing factor used to adjust the weights between the positive and negative classes, and γ is an adjustment parameter used to control the weights of easy and difficult samples; S1.3 Set the training algorithm parameters and adjust the weight connection parameters according to the learning results of the loss function until the preset accuracy is reached to obtain the U-BiFormer surface fine classification model. Use the U-BiFormer surface fine classification model to perform fine classification of surface types on remote sensing images of desert areas to obtain the corresponding surface types, including towns, farmland, desert, water bodies, and mountains. S1.

4. Different surface emissivity values ​​are obtained for different land surface types after classification to obtain the first high-resolution surface emissivity data: 。 3. The coupling method for improving the resolution and accuracy of surface temperature inversion in desert areas according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1 Obtain low-resolution reflectance data for bands 13 and 14 in the ASTER dataset; S2.2 The Normalized Difference Vegetation Index (NDVI) is calculated using near-infrared (NIR) and infrared (R) data, as shown in the following formula: ， S2.

3. Based on the NDVI thresholds for soil and vegetation, the NDVI thresholds are used to classify pixels in the image using the vegetation index, dividing the pixels into pure water bodies, mixed pixels, and pure vegetation, and calculating the vegetation cover PV: ， S2.4 Determine the surface emissivity. When the NDVI is less than or equal to the soil threshold, the pixel is considered to belong to water; when the NDVI is greater than or equal to the vegetation threshold, the pixel is considered to belong to vegetation. For pixels classified as mixed pixels, the previously calculated PV value is used for further calculation to obtain the bare soil emissivity of ASTER bands 13 and 14, as shown in the following formula: ， in, This represents the emissivity of bare soil in ASTER. Indicates the emissivity of the preset band. This indicates the emissivity of ASTER's pure vegetation. S2.

5. Convert the bare soil emissivity of ASTER bands 13 and 14 to the thermal infrared emissivity of Landsat series satellites, as follows: ， in, This represents the bare soil emissivity in band 10 of the Landsat TIRS sensor. and These represent the bare soil emissivity of ASTER bands 13 and 14 obtained using calculations; S2.

6. Based on the bare soil emissivity, pure vegetation emissivity, and PV, the vegetation-adjusted Landsat emissivity is calculated as follows: ， in, This represents the surface emissivity of the 10th band, which serves as the input for subsequent single-channel algorithms. This represents the pure vegetation emissivity in Landsat band 10. This represents the bare soil emissivity of Landsat band 10; S2.

7. In the single-channel algorithm, the following formula is used for calculation: ， S2.

8. Interpolate the acquired low-resolution surface emissivity data to the same resolution as the visible light data to obtain the second-highest resolution surface emissivity. .

4. The coupling method for improving the resolution and accuracy of surface temperature inversion in desert areas according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Using a linear weighted fusion method, assign different weights to the first high-resolution surface emissivity data and the second high-resolution surface emissivity data. S3.

2. Based on the set weights, calculate the final high-resolution surface emissivity used in the single-channel algorithm according to the following formula. : ， in Refers to the first The surface emissivity obtained from the step, Refers to the first The weights corresponding to each step.

5. The coupling method for improving the resolution and accuracy of surface temperature inversion in desert areas according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Based on NCEP atmospheric water vapor content data, time interpolation is performed on the data to obtain the water vapor content at the time of data capture, using the following formula: ， in, This indicates the water vapor content at the time Landsat acquired the data. and These represent the NCEP water vapor content before and after the Landsat image capture time, respectively. Indicates the time when the Landsat image was captured. and These represent the times corresponding to the NCEP atmospheric water vapor content data records before and after the Landsat image capture time, respectively. S4.

2. Spatial interpolation is performed on the atmospheric water vapor content data to interpolate the data to the target area, and then the atmospheric function is calculated using the obtained atmospheric water vapor content data. The formula is as follows: ， in, It is a constant related to the sensor.

6. The coupling method for improving the resolution and accuracy of surface temperature inversion in desert areas according to claim 1, characterized in that, Step S5 includes the following sub-steps: S5.1 Construct a thermal infrared image dataset and preprocess it. The dataset includes low-resolution and high-resolution thermal infrared image data of different thermal infrared bands, as well as the corresponding labels. S5.2 Construct an SRGAN model including a generator and a discriminator. Input a low-resolution thermal infrared image into the SRGAN model and train it adversarially to output a high-resolution thermal infrared image. S5.3 Construct loss functions: content loss, adversarial loss, and perceptual loss. Content loss ensures that the generated image is as consistent as possible with the high-resolution image; adversarial loss improves the realism of the image by training the generator through a discriminator; perceptual loss focuses on the similarity of image detail features based on a pre-trained VGG network. Content loss function: , Adversarial loss function: , Perceptual loss function: , in, It is the width of the feature map. It is the height of the feature map. It is the feature map obtained after the j-th convolutional activation before the i-th maxpooling layer in the VGG19 network. Indicates the scaling factor. The loss function represents the standard VGG, where I is the number of pixels, HR represents the real image, and LR represents the scaled image. S5.4 Set the training parameters for the generator and discriminator, and start adversarial training of the SRGAN model. The model optimizes the weight parameters of the generator, generates high-resolution thermal infrared images, and refines the texture and radiation features of the crack region.

7. The coupling method for improving the resolution and accuracy of surface temperature inversion in desert areas according to claim 1, characterized in that, Step S6 includes the following sub-steps: S6.1 Convert the DN value of the thermal infrared band image to an absolute radiance value. The calculation formula is as follows: , in, It is a gain constant. It is a bias constant; S6.2 Calculate the luminance temperature based on the radiance value. The calculation formula is as follows: , in, It is the calibration constant for the thermal infrared band. S6.

3. Combining the third-highest resolution surface emissivity data and atmospheric functions, the surface temperature is calculated using a single-channel algorithm, as shown in the following formula: ， in, It is a quantity related to the equivalent wavelength; , This represents the high-resolution surface emissivity in a single-channel algorithm.

Images