A method for monitoring daily changes of snow cover based on an improved spatiotemporal fusion model

Abstract

The present invention discloses a method for monitoring daily changes of snow accumulation based on an improved spatiotemporal fusion model, and belongs to the technical field of satellite remote sensing image processing and application. The purpose of the present invention is to solve the problem that the existing snow accumulation products have low spatial resolution and cannot accurately obtain the details of changes such as snowfall and snowmelt. The present invention uses 2 Landsat8 satellite remote sensing images and 17 MOD09GA spectral reflectance images as data sources, and uses the STDFA spatiotemporal fusion model to obtain preliminary daily spatiotemporal fusion images. Since the spatiotemporal fusion results of the STDFA model have spectral distortion, the present invention proposes to use the Matching-Pix2pixGAN network to improve the spatiotemporal fusion results obtained by the STDFA fusion model, and finally obtain the daily spatiotemporal fusion results without spectral distortion, and extract the NDSI index of the final spatiotemporal fusion results to realize the recognition of snow accumulation, thereby realizing the daily change monitoring of snow accumulation for 16 consecutive days. The method of the present invention can accurately realize the daily change monitoring of snow accumulation, capture the details of the changes in snowfall and snowmelt, and provide reliable analysis data for snowmelt runoff analysis, agricultural irrigation and ecological environment protection.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite remote sensing image processing and application. Background Art

[0002] Snow cover is a basic type of land cover and is widely used in the study of hydrology and climate-human-environment interactions. On Earth, the average maximum area covered by snow is 47 million square kilometers, making it one of the main components of the cryosphere. As a major component of the cryosphere, snow covers an important role in the Earth's climate system by affecting the surface energy budget, water cycle, primary productivity and surface gas exchange. Due to its high albedo and heat storage characteristics, it greatly affects the radiation balance. The accumulation and ablation of snow cover are closely related to temperature. Snowmelt in mid-latitudes affects local or regional climate and hydrological processes, so snow cover is a very important indicator for observing climate change. The interaction between snow and vegetation, the melting of snow cover helps the survival and fertility of vegetation, and has a profound impact on soil frost and plant chilling damage. Obtaining accurate information on snow cover is of great significance for climate change research, water cycle and hydrological resource investigation.

[0003] At present, snow monitoring is mainly achieved through three methods: hydrological model inversion, ground station measurement and remote sensing image analysis. Hydrological model is a reliable snow estimation method. Based on the input of climate, topography and other environmental information, numerical models can be used to simulate snow. Although snow models can predict snow changes under specific climate scenarios in the future, snow models have certain limitations and hinder their development, such as the limitations of model application areas, meteorological observation conditions and complex snow properties. The method based on ground station measurement is a snow monitoring process that relies on artificial methods. This method is not only time-consuming and slow to measure, but also unable to conduct detailed snow surveys on a large spatial scale. Remote sensing images can monitor large areas of snow in real time, and remote sensing data can be used to observe snow instead of traditional artificial ground station measurements. Nowadays, various optical satellite data are used to identify the extent of snow cover. Li et al. used MODIS data from the Qinghai-Tibet Plateau from 2001 to 2014 to study the spatiotemporal variation characteristics and trends of snow cover on the Qinghai-Tibet Plateau. (See C. Li, F. Su, D. Yang, K. Tong, F. Meng, and B. Kan, “Spatiotemporal variation of snow cover over the Tibetan Plateau based on MODIS snow product, 2001–2014,” Int. J. Climatol., vol. 38, pp. 708–728, 2018); Jin et al. used MOD10A2 and MYD10A2 data to analyze the changes in snow cover on the Loess Plateau and studied the spatiotemporal characteristics of snow cover in the Loess Plateau. (See X. Jin, C. Q. Ke, Y. Y. Xu, X. C. Li, “Spatial and temporal variations of snow cover in the Loess Plateau, China,” Int. J. Climatol., vol. 35, no. 8, pp. 1721–1731, Jun. 2015, doi: 10.1002 / joc.4086.); Gascoin et al. used Sentinel-2 and Landsat 8 satellite data to obtain high spatial resolution snow monitoring results, which can capture the details of snow changes.(See S. Gascoin, M. Grizonnet, M. Bouchet, G. Salgues, and O. Hagolle, "Theia snow collection: High-resolution operational snow cover maps from sentinel-2 and landsat-8 data," Earth Syst. Sci. Data., vol. 11, no. 2, pp. 493–514, 2019.).

[0004] So far, domestic and foreign scholars have proposed many methods for snow monitoring, but there are still some obvious defects: (1) Due to the low spatial resolution of satellite data used for snow monitoring, the spatial resolution of the snow monitoring results obtained is low, and it is impossible to monitor the details of snow changes in a certain area. (2) Due to the low temporal resolution of satellite data used for snow monitoring, it is impossible to obtain snow monitoring results on a daily basis. (3) Due to the limitations of orbital parameters and sensors, it is challenging to obtain remote sensing images with high temporal and spatial resolutions, so it is particularly difficult to obtain snow monitoring results with high temporal and spatial resolutions.

[0005] The spatiotemporal fusion method can enhance the temporal resolution and spatial resolution of the data, and generate effective data with both high temporal resolution and high spatial resolution. The spatiotemporal fusion data obtained by using the spatiotemporal fusion model can realize daily snow monitoring. The STDFA-Matching-Pix2pixGAN algorithm proposed in the present invention is an improved spatiotemporal fusion algorithm, which can improve the data quality of the spatiotemporal fusion results, improve the spectral distortion in the spatiotemporal fusion data, and further improve the accuracy of snow recognition. Summary of the invention

[0006] In order to solve the problems that existing snow products cannot meet the requirements of high spatial resolution and high temporal resolution at the same time, and cannot simultaneously realize daily snow observation and monitor the details of snow changes, the present invention uses the STDFA-Matching-Pix2pixGAN algorithm to improve the spatiotemporal fusion model, obtain spatiotemporal fusion data with high temporal and spatial resolution, and obtain the snow cover of the study area by calculating the normalized difference snow index (NDSI) of the spatiotemporal fusion data.

[0007] The technical solution adopted by the present invention is as follows:

[0008] The method for monitoring daily changes of snow cover based on the improved spatiotemporal fusion model has the following specific steps:

[0009] Step 1: Image preprocessing: Preprocess the image at the initial stage (T0) and the final stage (T k) to perform radiometric calibration and atmospheric correction on the two Landsat8 multispectral data corresponding to T0~T k The MOD09GA spectral reflectance data of different strip numbers every day are spliced, and then the processed Landsat8 data and MOD09GA data are aligned, and finally the experimental area is cropped and selected.

[0010] (a) Radiometric calibration method: Radiometric calibration is the process of converting the digital quantization value (DN) of an image into physical quantities such as radiation brightness, reflectivity or surface temperature. The radiometric calibration parameters are generally placed in the metadata file. The radiometric calibration tool (Radiometric Calibration) of the complete remote sensing image processing platform (ENVI) can automatically read the parameters from the metadata file to complete the radiometric calibration.

[0011] (b) Atmospheric correction method: Use the atmospheric correction tool (FLASSH) of the complete remote sensing image processing platform (ENVI) to eliminate the radiation error caused by the atmospheric influence and invert the true surface reflectivity of the object.

[0012] (c) MOD09GA data splicing method: Use MODIS Reprojection Tool (MRT) software to splice data with different strip numbers to form complete data covering the experimental area.

[0013] (d) Image registration method: Images taken from different sensors and different angles will be offset. Image registration is the process of matching and superimposing two images taken from different angles or different sensors to achieve image unification. Landsat8 data is selected as the reference image, and the image registration tool (ImageRegistration) of the complete remote sensing image processing platform (ENVI) is used to realize the automatic registration of Landsat8 satellite data and MOD09GA spectral reflectance data.

[0014] (e) Image cropping is used to obtain the region of interest (ROI).

[0015] Step 2: Generate spatiotemporal fusion data using the STDFA model: Spatiotemporal fusion technology can fuse remote sensing data with high temporal and spatial resolution. STDFA is a spatiotemporal fusion method based on spectral decomposition, with the initial (T0) and final (T k ) of Landsat8 data and MOD09GA data and the fusion period (T i , i = 1 ~ k-1) MOD09GA data is the input data of the STDFA model. The STDFA model consists of the following four steps:

[0016] (a) For the initial (T0) and final (T k ) to classify the two Landsat8 data and determine the types of land features;

[0017] (b) Calculate the abundance of each end member in each low spatial resolution pixel of MOD09GA data using equation (1)

[0018]

[0019] is the location of the MOD09GA low spatial resolution pixel, It is a type C feature in position The abundance, For location where n is the number of Landsat8 pixels belonging to class c in the MOD09GA pixels, and n is the total number of Landsat8 pixels contained in the MOD09GA pixels.

[0020] (c) According to formula (2), calculate the time series average reflectivity of each type of ground object.

[0021]

[0022] The conditions include: and

[0023] in, is the location of the MOD09GA low spatial resolution pixel, is the c-type feature at position obtained by formula (1). The abundance, MOD09GA low spatial resolution pixels in T i The surface reflectivity at time It is a c-type feature in T i The average reflectivity at the moment, M is the total number of categories, is the residual term.

[0024] (d) According to formula (3), a fused image with high spatial and temporal resolution is obtained.

[0025]

[0026] Where (x, y) is the location of the Landsat8 high spatial resolution pixel, r(x, y, T0) and r(x, y, T i ) are the Landsat8 high spatial resolution pixels at position (x, y) at time T0 and T i The surface reflectivity at time and is the c-type feature at time T0 and T i The average reflectivity at that time.

[0027] Step 3: Constructing the STDFA-Matching-Pix2pixGAN network: After step 2, the spatiotemporal fusion result of the STDFA model has spectral distortion, which will affect the subsequent snow monitoring. The present invention improves the spectral distortion by constructing the STDFA-Matching-Pix2pixGAN network that can improve the STDFA spatiotemporal model, and obtains a distortion-free fused image with high spatiotemporal resolution. The STDFA-Matching-Pix2pixGAN network is composed of three parts:

[0028] (a) Pre-matching process: Taking the spatiotemporal fusion image output by the STDFA model in step 2 as input, pre-matching uses histogram matching to reduce the source image (STDFA fusion image with spectral distortion) and the target image (T0 or T k The spectral difference between the corresponding Landsat8 satellite image). Taking the Landsat8 target image (y) and the STDFA fused image with spectral distortion (x) as input source data, the pre-matching module first uses the histogram matching algorithm to obtain the preliminary improved result X.

[0029] (b) Pix2pixGAN process: The output result (X) of the pre-matching module in (a) and the Landsat8 target image (y) are sent to the Pix2pixGAN network in pairs for training. Pix2pixGAN first constructs a generator (called G) to generate the required target high-resolution image G(X). Then G(X) and y are input into the discriminator D together. The discriminator D will continuously compare the difference between G(X) and y to determine whether G(X) still has spectral distortion. If the output image G(X) is judged to still have spectral distortion, the result is fed back to the generator to continue to optimize the generated image G(X) until the optimal generated result Pix2pixGAN image (P) without distortion is obtained. The loss function of the network is as shown in Formula 4.

[0030]

[0031] Among them, G * is the loss function of Pix2pixGAN, G represents the generator, and D represents the discriminator. Through the loop mechanism of Pix2pixGAN, the spectral distortion of the STDFA spatiotemporal fusion result can be improved, so that the fused image has high spatial resolution characteristics.

[0032] (c) Post-matching process: The post-matching module uses the spectral information of the MOD09GA data (m) on the same day to make the Pix2pixGAN image (P) have similar spectral information to m through histogram matching, further improving the spatiotemporal fusion results.

[0033] The present invention introduces the Matching-Pix2pixGAN network after the original STDFA spatiotemporal fusion model to form the STDFA-Matching-Pix2pixGAN spatiotemporal fusion algorithm. Through the method of the present invention, spatiotemporal fusion data with a spatial resolution of 30m that improves spectral distortion day by day can be obtained.

[0034] Step 4: Use the NDSI index to monitor snow accumulation: The present invention uses the normalized difference snow index (NDSI) to identify the snow accumulation in the experimental area. NDSI refers to the ratio of the difference between the reflectivity of the visible light band and the short-wave infrared band to the sum of the reflectivity of these two bands. NDSI uses the high reflectivity of snow in the green band and its strong absorption in the short-wave infrared to effectively distinguish snow from other objects. The calculation formula of NDSI is as shown in formula (5).

[0035]

[0036] Among them, b4 is the green band and b6 is the shortwave infrared band. NDSI can be used to identify snow cover and non-snow cover, so the snow cover in the experimental area is identified by calculating the NDSI index of the spatiotemporal fusion remote sensing image of the STDFA-Matching-Pix2pixGAN network, and the daily snow cover results of the experimental area with a spatial resolution of 30m for 16 consecutive days can be obtained.

[0037] Beneficial effects of the present invention:

[0038] The present invention uses Landsat8 satellite data and MOD009GA spectral reflectance data as data sources, uses the improved spatiotemporal fusion algorithm of STDFA-Matching-Pix2pixGAN, improves the spectral distortion problem in the spatiotemporal fusion result of the original STDFA model, obtains a spatiotemporal fusion image with a daily spatial resolution of 30m without spectral distortion, identifies snow cover in the experimental area by calculating the NDSI index of the spatiotemporal fusion result, and finally obtains a snow cover result with high spatiotemporal resolution. The snow cover area obtained by the method of the present invention has an error of no more than 0.84% ​​with the actual snow cover area. The method of the present invention can improve the spatiotemporal resolution of snow products, realize daily observation of snow cover change details, and provide valuable data for regional snow cover dynamic monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 These are the research area and research data used in the present invention.

[0040] Figure 2 It is a flow chart of the method for monitoring daily changes of snow accumulation based on the improved spatiotemporal fusion model of the present invention.

[0041] Figure 3 This is a spatiotemporal fusion result diagram obtained by the present invention using the STDFA model.

[0042] Figure 4 This is a spatiotemporal fusion result diagram obtained by the present invention using the STDFA-Matching-Pix2pixGAN algorithm.

[0043] Figure 5 This is a snow cover result map of the spatiotemporal fusion result obtained by the present invention using the STDFA-Matching-Pix2pixGAN algorithm. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further explained and illustrated in the form of specific embodiments below.

[0045] Two Landsat8 satellite data from December 12, 2018 and December 28, 2018 and 17 MOD09GA spectral reflectance data from December 12, 2018 to December 28, 2018 were used as data sources (Table 1), and red (R), green (G), blue (B), near infrared (NIR), shortwave infrared 1 (SW1) and shortwave infrared 2 (SW2) were used as experimental bands (Table 2), and Landsat8 data and MOD09GA data were image preprocessed. The experimental area is located in the southeast of Jilin Province. There is a large area of ​​snow in the experimental area in winter, which is conducive to snow monitoring ( Figure 1). The STDFA-Matching-Pix2pixGAN spatiotemporal fusion algorithm proposed in the present invention first uses the STDFA model to obtain preliminary spatiotemporal fusion images for 16 consecutive days, and then uses the Matching-Pix2pixGAN network to improve the spatiotemporal fusion results obtained by the STDFA model, and finally obtains a spatiotemporal fusion image of the experimental area with a daily spatial resolution of 30m without spectral distortion. By calculating the NDSI of the spatiotemporal fusion image to identify the snow cover of the experimental area, the snow cover situation of the experimental area with high spatial and temporal resolution can be obtained, and the daily snow cover change monitoring of the experimental area for 16 consecutive days can be realized, and the details of the snow cover change can be monitored. The accuracy of the STDFA-Matching-Pix2pixGAN spatiotemporal fusion results was verified using the Sentinel-2 satellite data on December 24, 2018 and the Landsat8 data on December 28, 2018. The snow cover of the Landsat8 data on December 28, 2018 was used to verify the accuracy of the snow cover results of the STDFA-Matching-Pix2pixGAN spatiotemporal fusion results. The error between the snow cover obtained by the present invention and the actual snow cover is no more than 0.84%. Refer to the overall flow chart of this example ( Figure 2 ).

[0046] Table 1

[0047]

[0048] Table 2

[0049]

[0050] Step 1: Image preprocessing

[0051] The initial stage (T0) and the final stage (T 16 ) were used to perform radiation calibration and atmospheric correction on the two Landsat8 multispectral data corresponding to the MOD09GA spectral reflectance data of different strip numbers for 17 days during the spatiotemporal fusion period, and then the processed Landsat8 data and MOD09GA data were aligned, and finally the experimental area was cropped and selected.

[0052] (a) Radiometric calibration method: Radiometric calibration is the process of converting the digital quantization value (DN) of an image into physical quantities such as radiation brightness, reflectivity or surface temperature. The radiometric calibration parameters are generally placed in the metadata file. The radiometric calibration tool (Radiometric Calibration) of the complete remote sensing image processing platform (ENVI) can automatically read the parameters from the metadata file to complete the radiometric calibration.

[0053] (b) Atmospheric correction method: Use the atmospheric correction tool (FLASSH) of the complete remote sensing image processing platform (ENVI) to eliminate the radiation error caused by the atmospheric influence and invert the true surface reflectivity of the object.

[0054] (c) MOD09GA data splicing method: Use MODIS Reprojection Tool (MRT) software to splice data with different strip numbers to form complete data covering the experimental area.

[0055] (d) Image registration method: Images taken from different sensors and different angles will be offset. Image registration is the process of matching and superimposing two images taken from different angles or different sensors to achieve image unification. Landsat8 data is selected as the reference image, and the image registration tool (ImageRegistration) of the complete remote sensing image processing platform (ENVI) is used to realize the automatic registration of Landsat8 satellite data and MOD09GA spectral reflectance data.

[0056] (e) Image cropping is used to obtain the region of interest (ROI).

[0057] Step 2: Generate spatiotemporal fusion data using the STDFA model:

[0058] The STDFA algorithm was used to obtain the spatiotemporal fusion results of high spatial and temporal resolution in the experimental area for 16 consecutive days. The input data of STDFA include the initial (T0) and final (T 16 ) of Landsat8 data and MOD09GA data and the fusion period (T i , i=1~15)MOD09GA data, the experimental bands include red, green, blue (RGB) bands as well as near infrared (NIR) and short wave infrared (SW1 and SW2) bands. Through the STDFA model, we can get T1 to T 16 16 high-speed spatiotemporal fusion data with 6 spectra with a spatial resolution of 30m and a temporal resolution of 1 day ( Figure 3 ), but the fused data of the STDFA model has spectral distortion, which affects the subsequent snow recognition and monitoring.

[0059] Step 3: Build the STDFA-Matching-Pix2pixGAN network:

[0060] Due to the spectral deviation of the fused data obtained by the STDFA model in step 2, the present invention proposes to add the STDFA-Matching-Pix2pixGAN network after the STDFA model to form an improved space-time fusion algorithm of STDFA-Matching-Pix2pixGAN to improve the spectral distortion of the space-time fusion results. The data of Landsat8 and MOD09GA without spectral distortion are used as the target images, and the images with spectral distortion obtained by STDFA are used as the source images. The source images and the target images together constitute the training data set of Matching-Pix2pixGAN. After many experiments, the optimal number of iterations of Matching-Pix2pixGAN is determined to be 400 times. By adding matching modules before and after Pix2pixGAN, the difference between the spectrally distorted images and the images without spectrally distorted images is significantly reduced, and the quality of the space-time fusion images is improved. Finally, high space-time fusion data with a spatial resolution of 30m and a temporal resolution of 1 day in the experimental area are obtained without spectral distortion, and the snow accumulation change process can be clearly observed day by day ( Figure 4 ).

[0061] Step 4: NDSI calculation and snow cover change detection:

[0062] Through step 3, the daily high spatiotemporal resolution fused images from December 12, 2018 to December 28, 2018 were obtained, and the NDSI of the fused images obtained by the method of the present invention was calculated according to formula (5). NDSI can identify snow pixels, and by calculating the number of snow pixels, the daily snow cover conditions in the experimental area for 16 consecutive days can be obtained ( Figure 5 ), observe the details of snow cover changes, and calculate the snow-covered area and the proportion of snow-covered area to the total area in the experimental area (Table 3).

[0063] Table 3

[0064]

[0065] The present invention utilizes the learning characteristics of the deep learning Pix2pixGAN network to make the spatiotemporal fusion image of the present invention have high spatial resolution characteristics, and adds matching modules to the front and back ends of Pix2pixGAN respectively, forming a Matching-Pix2pixGAN network and combining it with the STDFA model. Through the front and back two matching modules, the spatiotemporal fusion image of the present invention has the spectral characteristics of Landsat8 satellite data and MOD09GA spectral reflectance data, thereby obtaining distortion-free high spatiotemporal fusion data, providing high-quality basic data for subsequent snow monitoring. Subsequently, the NDSI of the STDFA-Matching-Pix2pixGAN spatiotemporal fusion result is calculated to identify the snow coverage of the experimental area. Finally, the present invention can realize the daily monitoring of the snow changes in the experimental area for 16 consecutive days, and can capture the details of the snow changes. The snow monitoring results obtained by the present invention have an error of less than 0.84% ​​with the actual snow coverage.

[0066] Experimental results: In order to illustrate the accuracy of the snow monitoring results obtained using the snow daily variation monitoring method based on the improved spatiotemporal fusion model, the final period T 16 The snow cover mapping of real Landsat8 data is used as verification data, and compared with the snow area products from December 12, 2018 to December 28, 2018 in the China MODIS daily cloudless 500m snow area product data set of the National Glacier, Frozen Soil and Desert Science Data Center (NCDC). Although the snow product of NCDC can realize daily snow monitoring, it is affected by its low spatial resolution, and the accuracy of NCDC snow area products is low and it is impossible to monitor the details of snow changes. The snow daily change monitoring method based on the improved spatiotemporal fusion model of the present invention can improve the problem of spectral distortion in the original spatiotemporal fusion model. At the same time, the method of the present invention has less dependence on cloudless data, can overcome the interference of clouds in the fusion intermediate period, and generate daily high spatiotemporal resolution fusion data for 16 consecutive days. Compared with the existing snow products, this method has better continuous monitoring capabilities, can track the snow change process in the experimental area in detail, is conducive to the monitoring of ecological environment and climate change, and provides reliable reference data for snowmelt runoff simulation and agricultural production irrigation.

Claims

1. A method for monitoring daily changes in snow cover based on an improved spatiotemporal fusion model, characterized in that: The specific steps of this method are as follows: Step 1: Image preprocessing: Preprocess the image at the initial stage T0 and the final stage T k The corresponding two Landsat8 multispectral data were calibrated for radiation and atmospheric correction, and the T0~T k The MOD09GA spectral reflectance data of different strip numbers every day were spliced, and then the processed Landsat8 data and MOD09GA data were aligned, and finally the experimental area was cropped and selected; Step 2: Generate spatiotemporal fusion data using the STDFA model: k Landsat8 data and MOD09GA data and the fusion period T i The MOD09GA data of (i=1~k-1) is the input data of the STDFA model to generate spatiotemporal fusion data; Step 3: Construct the STDFA-Matching-Pix2pixGAN network: By constructing the STDFA-Matching-Pix2pixGAN network that can improve the STDFA spatiotemporal model, the spectral distortion is improved through iteration, and finally a distortion-free high spatiotemporal resolution fused image is obtained; Step 4: Use the NDSI index to monitor snow accumulation: Use the normalized difference snow index NDSI to identify the snow accumulation in the experimental area; The calculation formula of NDSI is as shown in formula (5); Among them, b4 is the green band and b6 is the short-wave infrared band; In step 3, the STDFA-Matching-Pix2pixGAN network consists of three parts: (a) Pre-matching process: T0 or T k The corresponding Landsat8 data is used as the target image (y) and the spatiotemporal fusion image (x) output by the STDFA model obtained in step 2 is used as the input source data. The pre-matching module first uses the histogram matching algorithm to obtain the preliminary improvement result (X), so that the preliminary improvement result (X) has the spectral characteristics of the Landsat8 data as the target image (y); (b) Pix2pixGAN process: The output result (X) of the pre-matching module in step (a) and the target image (y) of the adjacent date are combined into a paired training data set and sent to the Pix2pixGAN network for training; Pix2pixGAN first constructs a generator, called G, to generate the required target high-resolution image G(X); then G(X) and y are input into the discriminator D together, and the discriminator D continuously compares the difference between G(X) and y to determine whether G(X) still has spectral distortion; if the output image G(X) is judged to still have spectral distortion, the result is fed back to the generator to continue to optimize the generated image G(X) until the optimal generated result Pix2pixGAN image (P) without distortion is obtained; the loss function of the network is as shown in Formula 4; Among them, G * is the loss function of Pix2pixGAN, G represents the generator, and D represents the discriminator. Through the loop mechanism of Pix2pixGAN, the spectral distortion of the STDFA spatiotemporal fusion result can be improved, so that the fused image has high spatial resolution characteristics. (c) Post-matching process: The post-matching module uses the spectral information of the MOD09GA data (m) on the same day to make the Pix2pixGAN image (P) have similar spectral information as the MOD09GA data (m) through histogram matching, further improving the spatiotemporal fusion results.

2. The method for monitoring daily changes of snow accumulation based on an improved spatiotemporal fusion model according to claim 1 is characterized in that: Step 1: Image preprocessing includes the following processing steps: (a) Radiometric calibration method: Radiometric calibration is the process of converting the digital quantization value DN of an image into radiometric brightness value, reflectivity or surface temperature; The radiation calibration parameters are placed in the metadata file. The radiation calibration tool of ENVI, a complete remote sensing image processing platform, can automatically read the parameters from the metadata file to complete the radiation calibration. (b) Atmospheric correction method: Use the atmospheric correction tool FLASSH of the complete remote sensing image processing platform ENVI to eliminate the radiation error caused by the atmospheric influence and invert the true surface reflectivity of the ground object; (c) Method of MOD09GA data splicing: Use MODIS Reprojection Tool software to splice data with different strip numbers to form complete data covering the experimental area; (d) Image registration method: Images taken from different sensors and at different angles will be offset. Image registration is the process of matching and superimposing two images taken from different angles or different sensors to achieve image unification. Landsat8 data is selected as the reference image, and the image registration tool of ENVI, a complete remote sensing image processing platform, is used to realize automatic registration of Landsat8 satellite data and MOD09GA spectral reflectance data. (e) Image cropping is used to obtain the region of interest (ROI).

3. The method for monitoring daily changes of snow accumulation based on an improved spatiotemporal fusion model according to claim 1, characterized in that: The STDFA model described in step 2 consists of the following four steps: (a) For the initial (T0) and final (T k ) to classify the two Landsat8 data and determine the types of land features; (b) Calculate the abundance of each end member in each low spatial resolution pixel of MOD09GA data using equation (1) is the location of the MOD09GA low spatial resolution pixel, It is a type C feature in position The abundance, For location is the number of Landsat8 pixels belonging to category c in the MOD09GA pixels, and n is the total number of Landsat8 pixels contained in the MOD09GA pixels; (c) According to formula (2), calculate the time series average reflectivity of each type of ground object; The conditions include: and in, is the location of the MOD09GA low spatial resolution pixel, is the c-type feature at position obtained by formula (1) The abundance, MOD09GA low spatial resolution pixels in T i The surface reflectivity at time, It is a c-type feature in T i The average reflectivity at the moment, M is the total number of categories, is the residual term; (d) Obtaining a fused image with high spatial and temporal resolution according to formula (3); Where (x, y) is the location of the Landsat8 high spatial resolution pixel, r(x, y, T0) and r(x, y, T i ) are the Landsat8 high spatial resolution pixels at position (x, y) at time T0 and T i The surface reflectivity at time, and is the c-type feature at time T0 and T i The average reflectivity at that time.

Images