Arctic accumulated snow thickness estimation method based on long short-term memory network

Abstract

The invention provides a remote sensing method for improving the thickness estimation precision of accumulated snow on the surface of north pole sea ice. According to the method, firstly, the monthly-scale actually-measured accumulated snow thickness (ASD) is obtained through the active microwave technology; then, by analyzing AMSR-2 brightness temperature data provided by the American snow and ice data center, the polarization ratio (PR) and the spectral gradient ratio (GR) of different channels are calculated; meanwhile, daily snow density data (sden) of the American snow and ice center is converted into monthly snow density data. Thirdly, performing space matching on the processed data and measured data of one-year ice and multi-year ice provided by OSI SAF, and extracting key variables based on importance features; by selecting appropriate variables and machine learning and deep learning models for training, a model combination with the best correlation, the minimum deviation and the lowest root mean square error (RMSE) is screened out to estimate the surface accumulated snow thickness of one-year ice and multi-year ice. Besides, different sea ice types are distinguished by using the sea ice type data of the OSI SAF, and a specific model is applied to each type, so that long-time sequence data of the thickness of the accumulated snow on the surface of the north pole sea ice is obtained. According to the method, by integrating active and passive microwave remote sensing observation data and an advanced deep learning method, the remote sensing estimation precision of the sea ice surface accumulated snow thickness is remarkably improved.

Description

Technical Field

[0001] The invention relates to a remote sensing estimation method for snow thickness on the sea ice surface, belonging to the technical field of remote sensing geoscience application. Background Art

[0002] In the context of global climate change, changes in the thickness and volume of Arctic sea ice have a profound impact on the Earth's energy balance and climate system. The thickness of snow on the sea ice surface is a key factor connecting the atmosphere and sea ice. Its low thermal conductivity and high reflectivity significantly affect the heat exchange process. In addition, snow thickness is a key parameter for inverting sea ice thickness using altimeter data, and is crucial for the conversion of sea ice freeboard height and the application of re-tracking methods. Therefore, high-precision snow thickness measurements are of great significance for the accurate estimation of Arctic sea ice thickness and volume, as well as the safe navigation planning of scientific research vessels such as Xuelong and Xuelong 2.

[0003] At present, the main methods for measuring the thickness of snow on the sea ice surface include buoy measurement, station measurement, cruise measurement, aerial observation, and microwave remote sensing technology. Among them, although the first four methods can provide high data accuracy, due to time and space limitations, it is difficult to analyze long-term series and large-scale snow depth changes. In contrast, satellite remote sensing technology can provide large-scale, long-term observation data, which has obvious advantages in the inversion of sea ice thickness and the analysis of the spatiotemporal variation characteristics of snow thickness.

[0004] Active microwave remote sensing can calculate snow depth from space, while the brightness temperature data estimated by passive microwave remote sensing has a certain correlation with the snow thickness on the sea ice surface. The spectral gradient ratio and polarization gradient ratio of different frequencies and polarization modes are calculated, and the best combination is found by machine learning. This is used as input data to determine the deep learning model for snow thickness estimation. Different snow thickness models are used for different months and different sea ice types to meet the estimation requirements of surface snow thickness of different sea ice types and seasons, improve the snow thickness inversion algorithm, and thus improve the snow thickness inversion accuracy. High-precision snow thickness has important application value for accurately inverting sea ice thickness and volume and predicting climate change. Summary of the invention

[0005] The technical problem to be solved by the present invention is: in view of the problems of low temporal and spatial representativeness of sample data (field observation data) and low snow depth inversion accuracy (especially for multi-year ice) in the current passive microwave remote sensing empirical method for inverting snow depth, this study uses deep learning, combined with satellite and ground synchronous observation data, to establish an active and passive microwave remote sensing joint inversion algorithm for snow depth on the surface of sea ice in the entire Arctic, which greatly improves the estimation accuracy of snow thickness on the surface of sea ice.

[0006] In order to solve the above technical problems, the technical solution proposed in the present invention is: the first step is to calculate the sea ice freeboard height using CryoSat-2 satellite data, and calculate the snow freeboard height in combination with ICESat-2 satellite data to obtain the measured snow thickness (ASD) on a monthly scale;

[0007] The second step is to calculate the polarization ratio (PR) and gradient ratio (GR) on a daily scale by analyzing the brightness temperature image data provided by the AMSR-2 satellite, and then sum and average these daily PR and GR data to obtain the monthly PR and GR;

[0008] The third step is to convert the daily snow density data (sden) provided by the US Snow and Ice Center into monthly snow density data;

[0009] Step 4: spatially match the ASD snow thickness with the monthly PR, GR and snow density data to extract the data of the same-name points;

[0010] Step 5: Use the random forest algorithm to train and predict all PR, GR, snow density and ASD snow thickness data. By performing permutation feature importance analysis, calculate the correlation, bias, root mean square error (RMSE) and mean absolute error (MAE) of different variable combinations, and select the variable combination with the best evaluation index as the input parameter. Then, input the optimal input parameter data into the long short-term memory network (LSTM) model for training to determine the best prediction model for one-year ice and multi-year ice in each month (months: 1, 2, 3, 4, 10, 11, 12).

[0011] Step 6: Use OSI SAF sea ice type data to distinguish one-year ice from multi-year ice. Use the optimal input parameter data to train the LSTM model for prediction, and use different prediction models to calculate the surface snow thickness of different types of sea ice in each month. Use Kriging interpolation to smooth and denoise the preliminary snow thickness data to obtain the final estimation result of sea ice surface snow thickness.

[0012] The remote sensing estimation method for snow thickness on the sea ice surface of the present invention also has the following characteristics:

[0013] 1. In the first step, the data coverage of ASD snow thickness should cover most of the Arctic to ensure that the data has sufficient representativeness and coverage, thereby ensuring the accuracy of the estimation results.

[0014] 2. In the second step, the brightness temperature data (TB) used is in nc format. The TB raw data is processed by ARCGIS software and converted into tif format. The PR and GR values ​​on a monthly scale are calculated according to the calculation formulas of PR and GR. The converted tif format data is further converted into Shapefile format, and resampling is performed every 25 km. The time resolution is set to monthly, and the projection uniformly adopts the Arctic azimuth projection. The converted Shapefile file contains PR / GR values, longitude and latitude information.

[0015] 3. In the third step, the original format of the snow density data is nc. The snow density data is processed by ARCGIS software and converted into Shapefile format. Resampling is performed every 25 km. The time resolution is set to monthly, and the projection uniformly adopts the Arctic azimuth projection. The converted Shapefile file contains snow density, longitude and latitude information.

[0016] 4. In the fourth step, the Shapefile data of ASD snow thickness and the tif format data of PR, GR, and snow density are spatially matched to extract the points with the same name, thereby obtaining a Shapefile file containing PR, GR, snow density, and ASD snow thickness information.

[0017] 5. In the fifth step, a random forest algorithm is used to rank the features by importance, and the optimal input model parameters are determined. Then, a long short-term memory network (LSTM) is used for training to obtain an estimation model for the snow thickness on the sea ice surface.

[0018] 6. In the sixth step, OSI SAF sea ice type data, which includes sea ice type and corresponding confidence information, is used to distinguish one-year ice from multi-year ice. The optimal input parameter data is used to train and predict the LSTM model, and different prediction models are used to calculate the surface snow thickness of different types of sea ice in each month. The Kriging interpolation method is used to smooth and denoise the preliminary snow thickness data to obtain the final estimation result of the sea ice surface snow thickness.

[0019] The beneficial effects of the present invention are:

[0020] Obtaining high-precision snow depth on the sea ice surface is of great significance to the material energy balance of the sea ice surface, the ocean-sea ice-atmosphere heat exchange, and the accurate inversion of sea ice thickness and volume. The present invention realizes the joint inversion of the surface snow thickness of different sea ice types in different months using active and passive microwave remote sensing, and adds a snow density parameter to improve the inversion accuracy. Based on the relationship between ASD snow thickness data and brightness temperature and snow density, a model is built to determine the calculation method of snow thickness in each month for multi-year ice and one-year ice, and the large-scale and long-time series of sea ice surface snow thickness is quickly and accurately obtained. The specific beneficial effects are as follows:

[0021] First, the present invention establishes the relationship between snow depth and brightness temperature based on the snow depth estimated by the satellite altimeter. The snow depth calculated by the satellite altimeter includes data from at least six growing seasons from 2018 to 2024, which solves the problem of limitations in the spatiotemporal distribution of measured data and improves the applicability of passive microwave remote sensing data in inverting snow depth in terms of spatiotemporal distribution.

[0022] Second, the brightness temperature data used in the present invention is AMSR-2 data, which is the successor of AMSR-E data. Together with AMSR-E, it can construct a brightness temperature data set from 2002 to the present. Brightness temperature images can be obtained free of charge, with strong real-time performance, convenient acquisition and processing;

[0023] Third, the present invention takes into account the effect of snow density on snow thickness. By spatially matching snow density data with ASD data and modeling the relationship between snow density, PR, GR and snow thickness, the snow thickness model on the sea ice surface is determined. This method not only greatly improves the prediction accuracy of the model, but also has the advantages of good practicality, easy understanding and implementation;

[0024] Fourth, the present invention comprehensively considers the snow thickness models of different sea ice types (such as multi-year ice and one-year ice) in different months. This method has high inversion accuracy, is more consistent with the actual situation of snow thickness, and improves the applicability and accuracy of the model.

[0025] Fifth, the data extraction and calculation process of the present invention is automatically implemented through Python programming, which greatly reduces the workload and manual errors and improves work efficiency. This method can be applied to large-scale, long-term series snow thickness analysis, effectively obtain the interannual and seasonal changes of snow thickness, and provide a powerful tool for polar climate change research. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below in conjunction with the accompanying drawings.

[0027] Figure 1 It is a flow chart of the remote sensing estimation method of snow thickness on the sea ice surface.

[0028] Figure 2is the spatial distribution map of ASD snow thickness used for modeling.

[0029] Figure 3 It is the latitude and longitude coordinates of ASD and snow thickness information.

[0030] Figure 4 It is the PR, GR and longitude and latitude information on a monthly scale.

[0031] Figure 5 It is the snow density and its longitude and latitude information on a monthly scale.

[0032] Figure 6 It is the snow thickness, PR, GR, and snow density information after ASD extracts the same-name points.

[0033] Figure 7 This is the spatial distribution of snow thickness on the sea ice surface on January 1, 2016. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to the accompanying drawings to make the steps and effects of the present invention clearer.

[0035] The example data used in this paper include AMSR-E / AMSR2 UnifiedL3Daily 25km Brightness Temperatures&Sea Ice Concentration Polar Grids, Version 1 products with a spatial resolution of 25km×25km, sea ice density data provided by the US Snow and Ice Center, and OSI SAF sea ice type data. The coverage area of ​​the example data image is the Northern Hemisphere, and the time span is from January 1, 2012 to April 30, 2021. ASD snow thickness data is calculated by active microwave (CryoSat-2 and ICESat-2), and the acquisition time is from October 2018 to April 2021.

[0036] Figure 1 This is a flow chart of the remote sensing estimation method for snow thickness on the sea ice surface. The specific implementation steps of this method include the following:

[0037] The first step is to select ASD snow depth data for multi-year ice and one-year ice. The selected ASD data meet the working period including the ice-forming period (October-April). A total of 21 sets of data were selected for the establishment of the snow accumulation model. The spatial distribution is as follows Figure 2 The specific snow thickness and longitude and latitude information are shown in Figure 3 shown.

[0038] The second step is to analyze the brightness temperature image data provided by the AMSR-2 satellite to calculate the polarization ratio (PR) and gradient ratio (GR) on a daily scale. Then, these daily PR and GR data are summed and averaged to obtain the monthly PR, GR values ​​and longitude and latitude information. Figure 4 shown.

[0039] The third step is to convert the daily snow density data of Sden provided by the US Snow and Ice Center into monthly snow density data; use ARCGIS to operate the original snow density data, convert it into shape format, perform resampling every 25km, the time resolution is monthly, and the projection is unified as the Arctic azimuth projection. The converted Shape file contains snow density, longitude and latitude information, as shown below: Figure 5 shown.

[0040] Step 4: Extract the same-name points of ASD snow thickness, PR, GR and snow density to obtain a Shape file containing snow thickness, snow density, PR, GR and their longitude and latitude information, such as Figure 6 shown.

[0041] Step 5: Use OSI SAF sea ice type data to distinguish one-year ice from multi-year ice. Eliminate all data with snow thickness less than 0, train and predict all PR, GR, snow density and ASD snow thickness through random forest, calculate the correlation, deviation, root mean square error and mean absolute error of different variable combinations according to the importance of permutation features, and use the best combination of evaluation indicators as the optimal input parameters. Then, the optimal input parameters are entered into the LSTM model for training to determine the best prediction model for one-year ice and multi-year ice in each month (months: 1, 2, 3, 4, 10, 11, 12).

[0042] Step 6: Perform LSTM training and prediction on the optimal input parameter data for different sea ice types, use different prediction models to calculate the snow thickness on the sea ice surface of different types in each month, and use Kriging interpolation method to smooth and denoise the preliminary snow thickness to obtain the final snow thickness on the sea ice surface. Figure 7 The figure shows the spatial distribution of snow thickness on January 1, 2016, which reflects the characteristics of high snow thickness north of the Canadian Archipelago and Greenland and low snow thickness north of Russia.

[0043] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A method for estimating the thickness of Arctic snow based on a long short-term memory network, comprising the following steps: The first step is to use CryoSat-2 satellite data to calculate the sea ice freeboard height, and combine it with ICESat-2 satellite data to calculate the snow freeboard height to obtain the measured snow thickness (ASD) on a monthly scale. The second step is to calculate the polarization ratio (PR) and gradient ratio (GR) on a daily scale by analyzing the brightness temperature image data provided by the AMSR-2 satellite, and then sum and average these daily PR and GR data to obtain the monthly PR and GR; The third step is to convert the daily snow density data (sden) provided by the National Snow and Ice Data Center into monthly snow density data; Step 4: spatially match the ASD snow thickness with the monthly PR, GR and snow density data to extract the data of the same-name points; Step 5: Use the random forest algorithm to train and predict the PR, GR, snow density and ASD snow thickness data of all monthly scales. By performing permutation feature importance analysis, calculate the correlation, bias, root mean square error (RMSE) and mean absolute error (MAE) of different variable combinations, and select the variable combination with the best evaluation index as the input parameter. Then, input the optimal input parameter data into the long short-term memory network (LSTM) model for training to determine the best prediction model for one-year ice and multi-year ice in each month (months: 1, 2, 3, 4, 10, 11, 12). Step 6: Use OSI SAF sea ice type data to distinguish one-year ice from multi-year ice. Use the optimal input parameter data to train the LSTM model for prediction, and use different prediction models to calculate the surface snow thickness of different types of sea ice in each month. Then use the Kriging interpolation method to smooth and denoise the preliminary snow thickness data to obtain the final estimation result of the sea ice surface snow thickness.

2. The remote sensing estimation method for snow thickness on sea ice surface according to claim 1 is characterized in that: In the first step, the data coverage of ASD snow thickness should cover most areas of the Arctic to ensure that the data has sufficient representativeness and coverage, thereby ensuring the accuracy of the estimation results.

3. The remote sensing estimation method for snow thickness on sea ice surface according to claim 1, characterized in that: In the second step, the brightness temperature data (TB) used is in nc format. The TB raw data is processed by ARCGIS software and converted into tif format. The PR and GR values ​​on a monthly scale are calculated according to the calculation formulas of PR and GR. The converted tif format data is further converted into Shapefile format, and resampling is performed every 25 km. The time resolution is set to monthly, and the projection uniformly adopts the Arctic azimuth projection. The converted Shapefile file contains PR / GR values, longitude and latitude information.

4. The remote sensing estimation method for snow thickness on sea ice surface according to claim 1 is characterized in that: In the third step, the original format of the snow density data is nc, and the snow density data is processed by ARCGIS software and converted into Shapefile format, and resampling is performed every 25km, the time resolution is set to monthly, and the projection uniformly adopts the Arctic azimuth projection. The converted Shapefile file contains snow density, longitude and latitude information.

5. The remote sensing estimation method for snow thickness on sea ice surface according to claim 1, characterized in that: In the fourth step, the Shapefile data of ASD snow thickness and the tif format data of PR, GR, and snow density are spatially matched to extract points with the same name, thereby obtaining a Shapefile file containing PR, GR, snow density, and ASD snow thickness information.

6. The remote sensing estimation method for snow thickness on sea ice surface according to claim 1 is characterized in that: In the fifth step, a random forest algorithm is used to rank the features by importance, and the optimal input model parameters are determined. Then, a long short-term memory network (LSTM) is used for training to obtain an estimation model for the snow thickness on the sea ice surface.

7. The remote sensing estimation method for snow thickness on sea ice surface according to claim 1 is characterized in that: In the sixth step, OSI SAF sea ice type data, which includes sea ice type and corresponding confidence information, is used to distinguish one-year ice from multi-year ice. The optimal input parameter data is used to train and predict the LSTM model, and different prediction models are used to calculate the surface snow thickness of different types of sea ice in each month. The Kriging interpolation method is used to smooth and denoise the preliminary snow thickness data to obtain the final estimation result of the sea ice surface snow thickness.

Images