Spatiotemporal prediction method for snow depth of SAR satellite remote sensing based on ADC-GRU network and FF-DNN model

Abstract

The application discloses a SAR satellite remote sensing snow depth spatio-temporal prediction method based on an ADC-GRU network and an FF-DNN model and belongs to the technical field of satellite remote sensing image processing and application.The application aims at the demand that existing snow depth products cannot simultaneously meet high spatial resolution and high temporal resolution, combines historical monitoring data of a weather station and other auxiliary data, and proposes a time series snow depth prediction model based on an ADC-GRU network (an abnormal data detection and correction-gate recurrent unit network); further, the time series snow depth prediction model based on the ADC-GRU network is combined with SAR satellite remote snow depth inversion results based on an FF-DNN model, the time series of a ground time series algorithm model and the advantage of radar satellite remote sensing large-range observation snow depth are combined, the SAR satellite remote sensing snow depth spatio-temporal prediction method based on the ADC-GRU network and the FF-DNN model is established, and daily high spatio-temporal resolution snow depth results are obtained.

Description

Technical Field

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

[0002] Snow cover is a precious freshwater resource and an indispensable part of natural climate research, profoundly impacting the global water cycle, crop growth, and energy balance. Accurately estimating snow depth helps in the rational use of snow resources, disaster prevention, and biodiversity conservation. This is crucial for human production activities and global climate change mitigation.

[0003] Traditional methods for measuring snow depth typically involve using fixed instruments at ground-based meteorological stations or conducting field experiments with portable equipment. These methods usually require significant manpower, resources, and time. In recent years, an increasing number of scholars have utilized various mathematical models and physical methods, combined with ground data, to obtain meteorological parameters such as snow depth. Anderson et al. proposed a complex snow cover model for meteorological research by considering the changes in snow cover itself (see Anderson E A. A point energy and mass balance model of a snow cover [R]. NOAA Technical Report NWS, 19, Office of Hydrology. National Weather Service, Silver Spring, MD, 1976: 150.). Loth constructed a one-dimensional multilayer snow cover model based on the balance of mass and energy, such as temperature and water vapor diffusion, melting and freezing, attenuation of shortwave radiation, and retention of liquid water (see Loth B, GRAFH F. Snow cover model for global climate simulation [J]. Journal of Geophysical Res, 1993, 8(D6): 10451-10464.). Sun Shufen et al., from the perspective of climate research, established a moderately complex snow cover model based on the complex snow cover model, and analyzed and conducted sensitivity tests on key parameters of key snow cover processes (see Sun Shufen, Li Jingyang. Sensitivity Study of Snow Cover Model Parameterization Scheme for Climate Research [J]. Atmospheric Science, 2002, 26(4):558-576.). Deng Wei et al. analyzed relevant factors in different regions, determined precipitation phase by establishing corresponding forecast indicators, and predicted snow depth (see Deng Wei, Li Huien, Wang Huanxiao, et al. A Snow Depth Prediction Method and System: CN201911405094.X [P]. CN113126182A [2023-10-07].). In general, snow depth prediction schemes based on conventional ground observation methods have some reference value for snow cover forecasting. However, traditional models have limited ability to extract temporal features, and due to the difficulty in obtaining surface information in most areas, it is difficult to obtain the spatiotemporal distribution characteristics of large-scale snow cover using only conventional ground observation methods. Satellite remote sensing has a wide monitoring range and can quickly obtain comprehensive information on large-scale snow cover. Synthetic Aperture Radar (SAR) is not affected by cloud and fog weather and has high spatial resolution. Machine learning and deep learning algorithms are not limited by input data to a certain extent. In recent years, methods combining deep learning algorithms and satellite remote sensing have been gradually used to invert snow depth or snow water equivalent.J Luo et al. used a new algorithm based on machine learning to obtain snow cover in the eastern forest area of ​​the Qilian Mountains in northwestern China using surface reflectance and ground real-time data (see Luo J, Dong C, Lin K, et al. Mapping snow cover in forests using optical remote sensing, machine learning and time-lapse photography[J]. Remote Sensing of Environment, 2022, 275:113017.DOI:10.1016 / j.rse.2022.113017.). Yang et al. studied local snow-covered areas in China, mainly using a random forest model to retrieve snow depth, and developed several pixel-based passive microwave snow depth inversion algorithms (see Yang J, Jiang L, Pan J, et al. Comparison of Machine Learning-Based Snow Depth Estimates and Development of a New Operational Retrieval Algorithm over China[J]. Remote Sensing, 2022, 14(12):2800.). Hu et al. used passive microwave data and artificial neural networks, support vector regression and random forest regression algorithms to explore the characteristics of snow cover in the Northern Hemisphere, evaluated the potential of the three machine learning methods in snow depth estimation, and constructed a fused snow depth dataset for 2002-2004 based on this (see Hu Y, Che T, Dai L, et al. Snow depth fusion based on machine learning methods for the Northern Hemisphere[J]. Remote Sensing,2021,13(7):1250.).

[0004] SAR satellite remote sensing data supports high spatial resolution snow cover monitoring; however, due to the discontinuity of the revisit cycle for the study area, daily high-resolution snow cover observations cannot be obtained. This invention proposes a SAR satellite remote sensing snow depth spatiotemporal prediction method based on an ADC-GRU network and an FF-DNN model. This method combines the temporal characteristics of ground-based time-series algorithms with the wide-area advantage of radar and satellite remote sensing, enabling the acquisition of snow depth data with both high temporal and spatial resolution, thus achieving daily high spatial resolution snow depth spatiotemporal prediction. Summary of the Invention

[0005] This invention addresses the limitation of existing snow depth products in simultaneously meeting the requirements of high spatial and temporal resolution. By combining historical monitoring data from meteorological stations and other auxiliary data, a time-series snow depth prediction model based on an ADC-GRU network (anomaly detection and correction-gate cyclic unit network) is proposed. Furthermore, this invention combines the time-series snow depth prediction model based on the ADC-GRU network with the snow depth inversion results from SAR satellite remote sensing based on the FF-DNN model. By leveraging the temporal characteristics of ground-based time-series algorithm models and the advantages of large-scale snow depth observation by radar satellite remote sensing, a spatiotemporal prediction method for SAR satellite remote sensing snow depth based on the ADC-GRU network and the FF-DNN model is established to obtain daily high spatiotemporal resolution snow depth results.

[0006] The specific steps of the technical solution adopted in this invention are as follows:

[0007] Step 1: Data Preprocessing. Using land classification data (https: / / globallandcover.com / ), the study area was divided into farmland and forest land cover types. Simultaneously, the data required for snow cover inversion and snow cover prediction were preprocessed, primarily including Sentinel-1 data, meteorological data, geographic information data, and forest characteristic data.

[0008] (a) Sentinel-1 SAR data (https: / / scihub.copernicus.eu / dhus / ) preprocessing: Sentinel-1 IW GRD level data was selected for preprocessing. The main processing steps included: orbit correction, thermal noise removal, radiometric calibration, speckle filtering, and terrain correction. After preprocessing, the data was cropped and stitched according to the selected study area.

[0009] The key characteristic data include radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE, and backscattering coefficients Sigma for different polarization modes within the study area. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH-Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VV );

[0010] (b) Meteorological data were obtained from the ERA5 climate reanalysis dataset (https: / / cds.climate.copernicus.eu) and Rp5 data (https: / / rp5.ru / docs / about / cn). This dataset is the fifth generation of global climate and atmosphere reanalysis results calculated by the European Centre for Weather Forecasts by combining observational and model data from around the world. It has a spatial resolution of 11 km and a temporal resolution of hourly.

[0011] Meteorological data preprocessing involved cropping and stitching the data according to the selected study area, followed by resampling. After processing, daily, pixel-by-pixel meteorological data for the study area were obtained, including 2m temperature, snow density, snow albedo, 2m dew point temperature, and atmospheric pressure.

[0012] The geographic information data used was SRTMDEM 90M data (https: / / EarthExplorer.usgs.gov), which provides geographic information data as raw 90m resolution elevation data. Geographic information data preprocessing involved cropping and stitching according to the selected study area, followed by resampling. This process yielded pixel-by-pixel geographic information data for the study area, including elevation, slope, and aspect.

[0013] Forest feature data includes forest cover data and forest canopy height data. The national forest canopy height data (https: / / www.plantplus.cn / cn / dataset / ) provides forest canopy height values ​​for China, with a spatial resolution of 30 m / pixel. Pixel-by-pixel forest canopy height is obtained after preprocessing. Forest cover data is provided by the annual product of MOD44B VCF (https: / / search.earthdata.nasa.gov). It provides tree cover, non-tree cover, and bare vegetation cover, with a spatial resolution of 250 m / pixel. Pixel-by-pixel forest cover is obtained after preprocessing.

[0014] Step 2: Establish a network anomaly detection and correction-gate cyclic unit network (ADC-GRU) time series snow depth prediction model based on historical meteorological station data; this model consists of the following three parts:

[0015] (a) Input data: For the forest and farmland study areas, meteorological data, geographic information data and time auxiliary data corresponding to the ground stations are obtained as input data. The meteorological data are 2m temperature, 2m dew point temperature and atmospheric pressure; the geographic information data are longitude, latitude, elevation, slope and aspect data; the time auxiliary data are date data; among them, the meteorological data adopts Rp5 data and the geographic information data adopts SRTMDEM 90M data after processing in step one.

[0016] (b) Snow Depth Anomaly Data Detection and Correction Module (ADC): Historical weather station snow depth observation data often contain some data anomalies or missing data. First, an empirical threshold SD is determined based on the snow depth changes in the study area. thd When the difference in snow depth data ΔSD between observation points at nearby times exceeds an empirical threshold, it is considered that there is abnormal data. The abnormal data is then corrected using snow depth data from the current observation point at nearby times, using the following formula:

[0017] ΔSD=SD(t-1)-SD(t+1) (1)

[0018]

[0019] Wherein, SD' is the correction result, SD(t-1) is the snow depth value on the day before the anomaly occurred, and SD(t+1) is the snow depth value on the day after the anomaly occurred.

[0020] (c) Time-series snow depth prediction GRU network: Corrected historical snow depth data is further input into the GRU network to construct a time-series snow depth prediction model based on meteorological station data. GRU can be applied to time-series prediction and is an improved LSTM neural network. It can preserve information in long-term sequences without being cleared or removed over time due to irrelevance to the prediction. GRU can solve the gradient vanishing problem in traditional RNN networks, effectively reducing computational complexity and improving computational efficiency. Compared with LSTM, which has input gate, forget gate, and output gate, GRU only has two storage units, namely the update gate (Z). t ) and reset gate (r t The update gate is used to update the hidden state and determine the amount of information passed to the model. The reset gate acts directly on the pre-hidden state, combining the new input information with the previous information. These two gate vectors determine the final output information. The training process uses the backpropagation algorithm to update the parameters and iterates sequentially until the loss function converges. The calculation formulas (3)-(6) for specific units of the GRU model are shown.

[0021] Z t =σ(W zx x t +W zh h t-1 +b z (3)

[0022] r t =σ(W rx x t +W rh h t-1 +b r (4)

[0023]

[0024]

[0025] Z t It is the update gate vector at time t, r t h is the reset gate vector at time t. t-1 It was the hidden state from a previous moment, X t It represents the input features at the current moment. (C) t It is the output vector of the hidden layer at time t, and the reset gate control h. t-1 and x t The degree of influence on the candidate set. This is the updated candidate vector. Additionally, σ is the sign function, and W... zx W zh W rh W rh W ch Wcx It is the weight matrix between each connection vector, b z b r b c These are the corresponding bias parameters. t It is the output state of the hidden layer at time t, and also the input state at the next time step.

[0026] Step 3: Using Sentinel-1 SAR data and auxiliary data as input, construct the FF-DNN model; the construction of the FF-DNN model consists of the following three steps:

[0027] (a) Data preprocessing: For farmland and forest land cover types, the corresponding Sentinel-1 data, ERA5 meteorological data, SRTMDEM 90M geographic information data, and forest cover and forest canopy height data were obtained for the study area.

[0028] The original characteristics of farmland types include: radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE obtained based on Sentinel-1 data, and backscattering coefficients Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (GammaVH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VV ); 2m temperature, snow density and snow albedo were obtained based on ERA5 meteorological data; elevation, slope and aspect were obtained based on SRTMDEM 90M.

[0029] The original characteristics of the forest type include: radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE obtained based on Sentinel-1 data, and backscattering coefficient Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VVBased on ERA5 meteorological data, 2m temperature, snow density, and snow albedo were obtained; based on SRTMDEM 90M, elevation, slope, and aspect were obtained; based on national forest canopy height data, forest canopy height data were obtained; based on MOD44BVCF data, forest coverage was obtained.

[0030] (b) Standardize all input data for each feature dimension using formula (7).

[0031]

[0032] Where x represents the input data, and μ and σ represent the mean and variance of each feature data, respectively.

[0033] (c) Optimize all input features using a feature selection method based on the Pearson correlation coefficient, selecting the best feature inputs and retaining the features with the strongest correlation to snow depth but the weakest cross-correlation, thus avoiding feature redundancy. The Pearson correlation coefficient measures the correlation between data; the closer its absolute value is to 1, the stronger the correlation between data. The specific selection process is as follows:

[0034] (1) The formula for calculating the Pearson correlation coefficient between two variables X and Y is as follows:

[0035]

[0036] Where X and Y are feature variables. Let Y be the average of the feature variables, and n be the number of values ​​each variable can take. First, let Y be the snow depth (SD), X be the input feature, and form a set P1 = {P...} of all results. SD-X1 P SD -X2, ..., P SD-XM}

[0037] (2) Sort the feature values ​​in set P1 in descending order, and retain the feature values ​​of the top K values ​​as set P2. P2 = {P SD-X1 ',P SD-X2 ',...,P SD-Xk The threshold K is determined specifically through experience and experimental precision.

[0038] (3) Calculate the Pearson correlation coefficient of any two variables in set P2, and retain the features with a Pearson correlation coefficient of less than 0.9 and a higher correlation with snow depth as the final optimized features.

[0039] (d) The optimized features are input into a DNN network to construct a snow depth inversion model and obtain the snow depth values ​​corresponding to the sentinel data in the study area. A Deep Neural Network (DNN) is a feedforward neural network based on an error backpropagation algorithm. It can effectively handle nonlinear regression problems and process and analyze large-scale data. Internally, it includes an input layer, hidden layers, and an output layer. Adjacent layers are fully connected, but nodes within the same layer are not connected. The training process of a DNN can be mainly divided into two parts: forward propagation of the input information and error-based backpropagation. During training, the input features are first propagated forward along with the information, while the error is propagated backward. The loss function is calculated, and the weights and biases in the network are continuously adjusted to reduce the loss, ultimately reducing the error of the output result.

[0040] Step 4: Based on network anomaly detection and correction - gated recurrent unit network and FF-DNN model, perform spatiotemporal prediction of snow depth using SAR satellite remote sensing data:

[0041] For different land cover types, combining the time-series snow depth prediction model based on the ADC-GRU network and the SAR satellite remote sensing snow depth inversion results based on the FF-DNN model, the time-series snow depth prediction model of meteorological station data is applied to the spatiotemporal snow depth prediction model of SAR remote sensing data to obtain high temporal and spatial resolution spatiotemporal snow cover prediction results. The specific implementation consists of the following two parts: (a) and (b):

[0042] (a) Using the FF-DNN model constructed in step three and SAR satellite remote sensing images, snow depth in the study area was retrieved to determine the initial (T0) and final (T1) snow depth of the study area. k Sentinel-1 SAR data and auxiliary data are used as input data for the FF-DNN model. The FF-DNN model then obtains the initial (T0) and final (T1) time data of the study area. k The snow depth values ​​are calculated as follows: Initial (T0) snow depth is used as input data for the spatiotemporal snow depth prediction module, and final (T...) snow depth is used as input data for the final (T...) snow depth prediction module. k The snow depth value is used as the output verification data of the snow depth spatiotemporal prediction module.

[0043] When the study area is farmland: the selected features include radar local incidence angle (LIA) and Beta obtained based on Sentinel-1 data. VV Gamma VH and Beta VH / Beta VV Temperature, snow density, and snow albedo were obtained using ERA5 meteorological data at 2m depth; elevation and slope were obtained using SRTMDEM 90M data.

[0044] When the study area is a forest type: the selected features include radar local incidence angle LIA, ellipsoidal incidence angle IAFE, and backscattering coefficient Gamma for different polarization modes obtained from Sentinel-1 data. VH Beta VV and Sigma VH / Sigma VV Beta VH / Beta VV Sigma VH -Sigma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ); 2m temperature, snow density, and snow albedo obtained using ERA5 meteorological data; elevation, slope, and aspect obtained using SRTMDEM90M data; forest canopy height data obtained using national forest canopy height data; and forest cover obtained using MOD44B VCF data.

[0045] (b) Implement spatiotemporal prediction of SAR satellite remote sensing snow depth based on ADC-GRU network and FF-DNN model. The specific implementation process is as follows:

[0046] (1) Input data preparation: The data mainly includes snow depth, meteorological data, geographic information data and time data of the initial (T0) SAR satellite remote sensing images obtained based on the FF-DNN model.

[0047] Meteorological data were obtained from the ERA5 data processing in step one, including 2m temperature, 2m dew point temperature, and atmospheric pressure. Geographic information data were obtained from the SRTMDEM 90M data processing in step one, including elevation, slope, and aspect data. Time data included manually labeled date, year, month, and day data.

[0048] (2) For different land cover types, the ADC-GRU model proposed in step 2 is used to predict the snow depth in the study area in time series. The initial (T0) snow depth value is sent to the trained ADC-GRU model to predict the snow depth value on the second day (T1). At the same time, this prediction result is returned as the input information for the next prediction process.

[0049] (3) Iterate continuously to obtain data from the initial stage (T0) to the final stage (T... k The daily snow depth prediction results for the research area.

[0050] The beneficial effects of this invention are:

[0051] The purpose of this invention is to address the inability of existing snow depth products to simultaneously meet the requirements of high spatial and temporal resolution. By combining historical monitoring data from meteorological stations and other auxiliary data, a time-series snow depth prediction model based on an ADC-GRU (Anomaly Detection and Correction-Gate Cyclic Unit Network) network is proposed. Furthermore, this invention combines the ADC-GRU network-based time-series snow depth prediction model with snow depth inversion results from SAR satellite remote sensing imagery based on an FF-DNN model. By leveraging the temporal sequence of ground-based time-series algorithms and the advantages of large-scale snow depth observation by radar satellite remote sensing, a spatiotemporal prediction method for SAR satellite remote sensing snow depth based on the ADC-GRU network and FF-DNN model is established, acquiring daily high spatiotemporal resolution snow depth data. Compared to existing snow depth products, this method has better continuous monitoring capabilities and relatively higher resolution, enabling detailed tracking of snow cover changes in the study area. This is beneficial for ecological environment and climate change monitoring, and provides reliable reference data for snowmelt runoff simulation and meteorological research. Attached Figure Description

[0052] Figure 1 This is a flowchart of time series snow depth prediction based on ADC-GRU network.

[0053] Figure 2 This is a flowchart of a spatiotemporal prediction method for SAR satellite remote sensing snow depth based on ADC-GRU network and FF-DNN model.

[0054] Figure 3 This is a spatiotemporal prediction map of snow depth in farmland areas of Dunhua City, Jilin Province, China, based on SAR satellite remote sensing using an ADC-GRU network and an FF-DNN model.

[0055] Figure 4 This is a spatiotemporal prediction map of snow depth in the forest area of ​​Dunhua City, Jilin Province, China, based on SAR satellite remote sensing using the ADC-GRU network and FF-DNN model. Detailed Implementation

[0056] The technical solution of the present invention will be further explained and described below with reference to specific embodiments.

[0057] The specific steps of the embodiment are as follows:

[0058] Step 1: Data Preprocessing. Dunhua City, Jilin Province, is divided using land classification data, with regions of interest defined by farmland and forest features. Simultaneously, the data required for snow cover inversion and snow cover prediction (Table 1) is preprocessed, primarily including Sentinel-1 SAR radar data, ERA5 data, SRTMDEM 90M data, MOD44B VCF data, and forest canopy height data from January 13th and 18th, 2017. The specific processing method is as follows:

[0059] (a) Preprocessing of snow depth inversion image data:

[0060] (1) Sentinel-1 Data Preprocessing: Two Sentinel-1 IW GRD-level data frames from January 13, 2017 and January 18, 2017 were preprocessed. The preprocessing mainly included: orbit correction, thermal noise removal, radiometric calibration, speckle filtering, and terrain correction. After preprocessing, the data were cropped and stitched according to the selected study area. Key characteristic data included Radar Local Incident Angle (LIA), Product Local Incident Angle (PLIA), Ellipsoidal Incident Angle (IAFE), and the backscattering coefficient Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV (Beta) VH -Beta VV ) / (Beta VH +Beta VV ).

[0061] (2) Meteorological data preprocessing: The ERA5 meteorological data corresponding to Dunhua City, Jilin Province on January 13, 2017 and January 18, 2017 were cropped and spliced ​​according to the selected areas and resampled. The data mainly include 2m temperature, snow density, snow albedo, and 2m dew point temperature.

[0062] (3) Geographic Information Data Preprocessing: The SRTMDEM 90M geographic information data corresponding to Dunhua City, Jilin Province on January 13, 2017 and January 18, 2017 were cropped and stitched according to the selected areas and then resampled. The data mainly includes elevation, slope, and aspect.

[0063] (4) Forest feature data preprocessing: The forest coverage data and forest canopy height data corresponding to Dunhua City, Jilin Province on January 13, 2017 and January 18, 2017 were cropped and spliced ​​according to the selected regions and then resampled.

[0064] (b) Preprocessing of snow depth prediction image data:

[0065] (1) Meteorological data preprocessing: The ERA5 meteorological data corresponding to Dunhua City, Jilin Province from January 13, 2017 to January 18, 2017 were cropped, spliced, and resampled according to the selected areas. The data mainly include 2m temperature, 2m dew point temperature, and atmospheric pressure.

[0066] (2) Geographic Information Data Preprocessing: The SRTMDEM 90M geographic information data corresponding to Dunhua City, Jilin Province from January 13, 2017 to January 18, 2017 were cropped, stitched together, and resampled according to the selected regions. The data mainly includes elevation, slope, and aspect.

[0067] Table 1 Snow cover inversion and spatiotemporal prediction data

[0068]

[0069] Step 2: Establish an ADC-GRU network (Anomaly Detection and Correction - Gate Cyclic Unit Network) time series snow depth prediction model based on historical meteorological station data. Figure 1The model is built in the following three steps:

[0070] (a) For forest and farmland areas in Northeast China, longitude, latitude, date, year, month, day, 2m temperature, 2m dew point temperature, atmospheric pressure, elevation, slope, and aspect data were obtained from 7 farmland meteorological stations and 6 forest meteorological stations from November 2015 to December 2018. The data were then fed into the prediction data generation and partitioning module for data preparation to construct the ADC-GRU network. The meteorological data of 2m temperature, 2m dew point temperature, and atmospheric pressure were obtained using Rp5 data (https: / / rp5.ru / docs / about / cn), and the geographic information data of elevation, slope, and aspect were obtained using SRTMDEM 90M data (https: / / EarthExplorer.usgs.gov).

[0071] (b) Snow Depth Anomaly Data Detection and Correction Module (ADC): Historical meteorological station data from November 2015 to December 2018 were used as the raw data for model construction. Historical meteorological station snow depth observation data often contain some data anomalies or missing data. First, an empirical threshold SD was determined based on the snow depth changes in the study area. thd In this example, it is set to 10cm. When the difference in snow depth data between adjacent observation points, ΔSD > SD... thd If the data is abnormal, it is considered to be abnormal. The abnormal data is then corrected using snow depth data from the nearest observation time, as shown in the following formula:

[0072] ΔSD=SD(t-1)-SD(t+1) (1)

[0073]

[0074] Wherein, SD' is the correction result, SD(t–1) is the snow depth value on the day before the anomaly occurred, and SD(t+1) is the snow depth value on the day the anomaly occurred.

[0075] (c) Time Series Snow Depth Prediction GRU Network: GRU can be applied to time series prediction. It is an improved LSTM neural network that can preserve information from long-term sequences without it being cleared over time or removed because it is irrelevant to the prediction. GRU can solve the gradient vanishing problem in traditional RNN networks, effectively reducing computational complexity and improving operational efficiency. Compared with LSTM, which has an input gate, a forget gate, and an output gate, GRU only has two storage units, namely the update gate (Z). t ) and reset gate (r tThe update gate is used to update the hidden state and determine the amount of information passed to the model. The reset gate is used to directly act on the pre-hidden state, combining the new input information with the previous information. These two gate vectors determine the final information output as the gated loop unit. The training process uses the backpropagation algorithm to update the parameters and iterates sequentially until the loss function converges. The calculation formulas (3)-(6) for the specific units of the GRU model are shown.

[0076] Z t =σ(W zx x t +W zh h t-1 +b z (3)

[0077] r t =σ(W rx x t +W rh h t-1 +b r (4)

[0078]

[0079]

[0080] Z t It is the update gate vector at time t, r t h is the reset gate vector at time t. t-1 It was the hidden state from a previous moment, X t It represents the input features at the current moment. (C) t It is the output vector of the hidden layer at time t, and the reset gate control h. t-1 and x t The degree of influence on the candidate set. This is the updated candidate vector. Additionally, σ is the sign function, and W... zx W zh W rh W rh W ch W cx It is the weight matrix between each connection vector, b z b r b c These are the corresponding bias parameters. t It is the output state of the hidden layer at time t, and also the input state at the next time step.

[0081] During the experiment, the corrected historical snow depth data was further input into the GRU network to construct a time-series snow depth prediction model based on meteorological station data. An ADC-GRU time-series snow depth prediction model for Northeast China was constructed using data from 7 farmland meteorological stations and 6 forest meteorological stations from November 2015 to December 2018, after anomaly processing.

[0082] Step 3: Snow Depth Inversion Method Based on SAR Satellite Remote Sensing Imagery using the FF-DNN Model. A snow depth inversion model suitable for forest and farmland areas in Dunhua City, Jilin Province, was constructed using meteorological station data and field experimental data from 2015 to 2019. Using Sentinel-1 data and auxiliary data as input, high spatial resolution snow depth results were obtained through the FF-DNN model. The FF-DNN model construction consists of the following three steps:

[0083] (a) Data preprocessing: For farmland and forest areas, Sentinel-1 data, ERA5 meteorological data, SRTMDEM 90M geographic information data, and forest coverage and forest canopy height data from 2015 to 2019 were obtained. The data were sent to the inversion data preprocessing module for data preparation and used to construct the FF-DNN network.

[0084] The original characteristics of farmland types include: radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE obtained based on Sentinel-1 data, and backscattering coefficients Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (SigmaVH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VV Temperature, snow density, and snow albedo at 2m depth obtained from ERA5 meteorological data; elevation, slope, and aspect obtained from SRTMDEM 90M.

[0085] The original characteristics of the forest type include: radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE obtained based on Sentinel-1 data, and backscattering coefficient Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV) and (Beta VH -Beta VV ) / (Beta VH +Beta VV ); 2m temperature, snow density, and snow albedo obtained from ERA5 meteorological data; elevation, slope, and aspect obtained from SRTMDEM 90M; forest canopy height data obtained from national forest canopy height data; forest cover obtained from MOD44B VCF data.

[0086] (b) Normalize the input data of the model. Use the z-score standardization method to standardize all input data for each feature dimension.

[0087]

[0088] Where x is the input data, and μ and σ are the mean and variance of each feature data, respectively.

[0089] (c) Optimize all input features using a feature selection method based on the Pearson correlation coefficient, selecting the best feature inputs and retaining the features with the strongest correlation to snow depth but the weakest cross-correlation, thus avoiding feature redundancy. The Pearson correlation coefficient measures the correlation between data; the closer its absolute value is to 1, the stronger the correlation between data. The specific selection process is as follows:

[0090] (1) The formula for calculating the Pearson correlation coefficient between two variables X and Y is as follows:

[0091]

[0092] Where X and Y are feature variables. Let Y be the average of the feature variables, and n be the number of values ​​each variable can take. First, let Y be the snow depth (SD), X be the input feature, and form a set P1 = {P...} of all results. SD-X1 P SD -X2, ..., P SD-XM}

[0093] (2) Sort the feature values ​​in set P1 in descending order, and retain the feature values ​​of the top K values ​​as set P2. P2 = {P SD-X1 ',P SD-X2 ',...,P SD-Xk Specifically, the threshold K is determined through experience and experimental precision; in this invention, K is 0.2.

[0094] (3) Calculate the Pearson correlation coefficient of any two variables in set P2, and retain the features with a Pearson correlation coefficient of less than 0.9 and a higher correlation with snow depth as the final optimized features. The final retained results are shown in Table 1.

[0095] (d) The optimized features are input into the DNN network to construct a snow depth inversion model. A Deep Neural Network (DNN) is a feedforward neural network based on the error backpropagation algorithm. It can effectively handle nonlinear regression problems and process and analyze large-scale data. Internally, it includes an input layer, hidden layers, and an output layer. Adjacent layers are fully connected, while nodes within the same layer are not connected. The training process of a DNN can be mainly divided into two parts: forward propagation of input information and error-based backpropagation. During training, the input features are first propagated forward with the information, while the error is propagated backward. The loss function is calculated, and the weights and biases in the network are continuously adjusted to reduce the loss, ultimately reducing the error of the output result. After hyperparameter tuning, the FF-DNN model constructed in the farmland area has 13 input features, with the number of neurons in the hidden layers being 32, 64, 64, and 32, respectively. The FF-DNN model constructed in the forest area has 21 input features, with the number of neurons in the hidden layers being 32, 128, 64, and 64, respectively. Specific features are shown in Table 1.

[0096] Step 4: Spatiotemporal prediction method for snow depth based on SAR satellite remote sensing imagery using ADC-GRU network and FF-DNN model ( Figure 2 This study targets farmland and forest areas in Dunhua City, Jilin Province. Combining a time-series snow depth prediction model based on an ADC-GRU network and snow depth retrieval results from SAR satellite remote sensing images based on an FF-DNN model, the time-series snow depth prediction model based on meteorological station data is applied to the spatiotemporal snow depth prediction model based on SAR remote sensing data. This yields the spatiotemporal prediction results of snow cover in Dunhua City, Jilin Province, from January 13th to January 18th, 2017. The specific implementation consists of the following two parts: (a) and (b).

[0097] (a) Snow depth inversion was performed in Dunhua City, Jilin Province, using the FF-DNN model constructed in step three and SAR satellite remote sensing imagery. Sentinel-1 SAR data and auxiliary data from Dunhua City, Jilin Province on January 13th and 18th, 2017, were used as input data for the FF-DNN model. The FF-DNN model then obtained the snow depth values ​​for Dunhua City, Jilin Province on January 13th and 18th, 2017. The results from January 13th, 2017, were used as input data for the snow depth spatiotemporal prediction module, and the results from January 18th, 2017, were used as output validation data for the snow depth spatiotemporal prediction module.

[0098] When the study area is farmland: the features selected by the FF-DNN model include radar local incidence angle (LIA) and Beta obtained based on Sentinel-1 data. VV Gamma VH and Beta VH / Beta VV Temperature, snow density, and snow albedo were obtained using ERA5 meteorological data at 2m depth; elevation and slope were obtained using SRTMDEM 90M data.

[0099] When the study area is a forest type: the features selected by the FF-DNN model include the radar local incidence angle LIA, ellipsoidal incidence angle IAFE, and backscattering coefficient Gamma for different polarization modes obtained from Sentinel-1 data. VH Beta VV and Sigma VH / Sigma VV Beta VH / Beta VV Sigma VH -Sigma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ); 2m temperature, snow density, and snow albedo obtained using ERA5 meteorological data; elevation, slope, and aspect obtained using SRTMDEM 90M data; forest canopy height data obtained using national forest canopy height data; and forest cover obtained using MOD44B VCF data.

[0100] (b) Implement spatiotemporal snow depth prediction based on ADC-GRU network and SAR satellite remote sensing imagery. The specific implementation process is as follows:

[0101] (1) Input data preparation: The data mainly includes snow depth inversion data on January 13, 2017, ERA5 meteorological data on January 13, 2017, SRTMDEM 90M geographic information data, and time data (date, year, month, day) of manually labeled information. The processed data is then sent to the prediction data generation and partitioning module for data preparation.

[0102] (2) For different land cover types, the ADC-GRU model proposed in step 2 is used to predict the snow depth in Dunhua City, Jilin Province. The snow depth inversion data on January 13, 2017 is sent into the trained ADC-GRU model to predict the snow depth value on the next day (January 14, 2017). At the same time, the prediction result is returned as the input information for the next prediction process.

[0103] (3) Iterate continuously to obtain the daily snow depth prediction results from January 13, 2017 to January 18, 2017. Figure 3 , Figure 4 ).

[0104] Experimental Results: This invention proposes a time-series snow depth prediction model based on an ADC-GRU network, combining historical monitoring data and auxiliary data from meteorological stations. Taking Northeast China as an example, a time-series snow depth prediction model for meteorological stations is first constructed. The average MAE of the prediction model for farmland areas is 0.61 cm, the RMSE is 1.12 cm, and the R... 2 The mean value was 0.98; the mean MAE of the forest prediction model was 0.75 cm, the mean RMSE was 1.45 cm, and the R... 2 It is 0.96.

[0105] Experimental results show that the time-series snow depth prediction model based on the ADC-GRU network can obtain high-precision daily snow depth prediction results at meteorological stations. Furthermore, by combining the time-series snow depth prediction model based on the ADC-GRU network with the SAR satellite remote sensing snow depth inversion results based on the FF-DNN model, a spatiotemporal prediction method for SAR satellite remote sensing snow depth based on the ADC-GRU network and the FF-DNN model is established to obtain daily high spatiotemporal resolution snow depth data. Taking farmland and forest areas in Dunhua City, Jilin Province as examples, the snow depth results inverted by the FF-DNN model on January 18, 2017, are used as ground truth verification data. Compared with the spatiotemporal prediction results of snow depth on January 18, 2017, the average MAE of the predicted snow depth and the average RMSE of the actual snow depth in farmland are 0.548 cm and 0.724 cm, respectively; while the average MAE of the predicted snow depth and the actual snow depth in forestland are 3.508 cm and 4.290 cm, respectively.

Claims

1. A spatiotemporal prediction method for SAR satellite remote sensing snow depth based on ADC-GRU network and FF-DNN model, characterized in that, The method includes the following steps: Step 1: Data Preprocessing Using land classification data, the study area was divided into farmland and forest land cover types. At the same time, the data required for snow cover inversion and snow cover prediction were preprocessed. The data mainly included Sentinel-1 SAR data, meteorological data, geographic information data and forest characteristic data. Sentinel-1 SAR data mainly includes radar local incidence angle (LIA), product local incidence angle (PLIA), ellipsoidal incidence angle (IAFE) in the study area, and backscattering coefficients (Sigma) for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VV ); Meteorological data were obtained using the ERA5 climate reanalysis dataset. After processing, daily and pixel-by-pixel meteorological data for the study area were obtained, including 2m temperature, snow density, snow albedo, 2m dew point temperature, and atmospheric pressure. Geographic information data was obtained using SRTMDEM 90M data, which was processed to obtain pixel-by-pixel geographic information data for the study area, including elevation, slope, and aspect. Forest feature data included forest cover data and forest canopy height data. The forest cover data was obtained pixel-by-pixel by processing the annual product of MOD44B VCF, and the forest canopy height data was obtained pixel-by-pixel by processing the national forest canopy height data. Step 2: Establish a time series snow depth prediction model based on historical meteorological station data using Network Anomaly Detection and Correction-Gate Cyclic Unit Network (ADC-GRU). (a) Input data: For the forest and farmland study areas, meteorological data, geographic information data and time auxiliary data corresponding to farmland meteorological stations and forest meteorological stations are obtained as input data; The meteorological data includes 2m temperature, 2m dew point temperature, and atmospheric pressure; the geographic information data includes longitude, latitude, elevation, slope, and aspect data; and the time-related data includes date data. Among these, the meteorological data uses Rp5 data, and the geographic information data uses SRTMDEM 90M data processed in step one. (b) Snow depth anomaly data detection and correction processing module (ADC): First, the empirical threshold SD for snow depth is determined based on the changes in snow depth in the study area. thd When the difference in snow depth data between observation points at adjacent times ΔSD > SD thd If so, it is considered that there is abnormal data; ΔSD=SD(t-1)-SD(t+1) (1) Where SD(t–1) is the snow depth value on the day before the anomaly occurred, and SD(t+1) is the snow depth value on the day after the anomaly occurred. Then, snow depth data from the current observation period were used to correct the anomalous data; Wherein, SD' is the correction result; (c) Time-series snow depth prediction GRU network: The corrected historical snow depth data is further input into the GRU network to construct a time-series snow depth prediction model based on meteorological station data. The model training process uses the backward error propagation algorithm to update the parameters and iterates sequentially until the loss function converges. The calculation formulas (3)-(6) for specific units of the GRU model are shown. Z t =σ(W zx x t +W zh h t-1 +b z ) (3) r t =σ(W rx x t +W rh h t-1 +b r ) (4) Z t It is the update gate vector at time t, r t It is the reset gate vector at time t; It is the updated candidate vector, C t W is the output vector of the hidden layer at time t; σ is the sign function, and W is the output vector of the hidden layer at time t. zx W zh W rx W rh W ch W cx h is the weight matrix between each connection vector. t It is the output state of the hidden layer at time t, and also the input state at the next time step; x t It is the input feature at the current moment; b z b r b c These are the corresponding bias parameters; Step 3: Using Sentinel-1 SAR data and auxiliary data as input, construct the FF-DNN model; the construction of the FF-DNN model consists of the following three steps: (a) Data preprocessing: For farmland and forest land cover types, the Sentinel-1 SAR data, ERA5 meteorological data, SRTMDEM 90M geographic information data, forest cover and forest canopy height data were obtained for the study area after processing in step one; The original characteristics of farmland types include: radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE obtained based on Sentinel-1 data, and backscattering coefficients Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VV Temperature, snow density, and snow albedo at 2m depth obtained from ERA5 meteorological data; elevation, slope, and aspect obtained from SRTMDEM 90M. The original characteristics of the forest type include: radar local incidence angle LIA, product local incidence angle PLIA, ellipsoidal incidence angle IAFE obtained based on Sentinel-1 data, and backscattering coefficient Sigma for different polarization modes. VH Gamma VH Beta VH Sigma VV Gamma VV Beta VV Sigma, and data augmentation based on original features. VH / Sigma VV Gamma VH / Gamma VV Beta VH / Beta VV Sigma VH -Sigma VV Gamma VH -Gamma VV Beta VH -Beta VV Sigma VH +Sigma VV Gamma VH +Gamma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ), (Gamma VH -Gamma VV ) / (Gamma VH +Gamma VV ) and (Beta VH -Beta VV ) / (Beta VH +Beta VV ); 2m temperature, snow density, and snow albedo obtained from ERA5 meteorological data; elevation, slope, and aspect obtained from SRTMDEM 90M; forest canopy height data obtained from national forest canopy height data; forest cover obtained from MOD44B VCF data; (b) Standardize all input data for each feature dimension using formula (7); Where x* represents standardized feature data, x represents input feature data, and μ and σ are the mean and variance of each feature data, respectively. (c) Optimize all input features using a feature selection method based on Pearson correlation coefficient, selecting the best feature input and retaining the features with the strongest correlation to snow depth but the weakest cross-correlation, thus avoiding feature redundancy. The Pearson correlation coefficient measures the correlation between data; the closer its absolute value is to 1, the stronger the correlation between data. The specific selection process is as follows: (1) The formula for calculating the Pearson correlation coefficient between two variables X and Y is as follows: Where X and Y are feature variables. Let n be the average of the characteristic variables, and n be the number of values ​​that each variable can take; firstly, Let Y be the snow depth (SD), and X be the input feature. Construct a set P1 = {P...} of all results. SD-X1 P SD -X2, ..., P SD-XM}； (2) Sort the feature values ​​in set P1 in descending order, and retain the feature values ​​of the top K values ​​as set P2; P2 = {P SD-X1 ',P SD-X2 ',...,P SD-Xk The specific threshold K is determined through experience and experimental precision. (3) Calculate the Pearson correlation coefficient of any two variables in set P2, and retain the features with a Pearson correlation coefficient of less than 0.9 and a higher correlation with snow depth as the final optimized features; (d) The optimized features are input into a deep neural network (DNN) to construct a snow depth inversion model and obtain the snow depth value corresponding to the Sentinel-1 SAR data of the study area. During the training process, the input features are first propagated forward with the information, and the error is propagated backward. The loss function is calculated, and the weights and biases in the network are continuously adjusted to reduce the loss and ultimately reduce the error of the output results. Step 4: Based on network anomaly detection and correction - gated recurrent unit network and FF-DNN model, perform spatiotemporal prediction of snow depth using SAR satellite remote sensing data: For different land cover types, combining the time-series snow depth prediction model based on the ADC-GRU network and the SAR satellite remote sensing snow depth inversion results based on the FF-DNN model, the time-series snow depth prediction model of meteorological station data is applied to the spatiotemporal snow depth prediction model of SAR remote sensing data to obtain high temporal and spatial resolution spatiotemporal snow cover prediction results; the specific implementation consists of the following two parts (a) and (b): (a) Using the FF-DNN model constructed in step three and SAR satellite remote sensing images, snow depth in the study area was inverted to obtain the initial T0 and final T0 values ​​for the study area. k Sentinel-1 SAR data, ERA5 meteorological data, SRTMDEM 90M geographic information data, and forest cover and canopy height data were used as input data for the FF-DNN model. The FF-DNN model was then used to obtain the initial T0 and final T0 values ​​for the study area. k Snow depth value; Among them, the initial snow depth value T0 is used as the input data for the snow depth spatiotemporal prediction module, and the final snow depth value T0 is used as the input data for the snow depth spatiotemporal prediction module. k Snow depth values ​​are used as output verification data for the snow depth spatiotemporal prediction module. When the study area is farmland: the features selected by the FF-DNN model include radar local incidence angle (LIA) and Beta obtained based on Sentinel-1 data. VV Gamma VH and Beta VH / Beta VV Temperature, snow density, and snow albedo at 2m depth were obtained using ERA5 meteorological data; elevation and slope were obtained using SRTMDEM 90M data. When the study area is a forest type: the features selected by the FF-DNN model include the radar local incidence angle LIA, ellipsoidal incidence angle IAFE, and backscattering coefficient Gamma for different polarization modes obtained from Sentinel-1 data. VH Beta VV and Sigma VH / Sigma VV Beta VH / Beta VV Sigma VH -Sigma VV Beta VH +Beta VV , (Sigma VH -Sigma VV ) / (Sigma VH +Sigma VV ); 2m temperature, snow density, and snow albedo obtained using ERA5 meteorological data; elevation, slope, and aspect obtained using SRTMDEM90M data; forest canopy height data obtained using national forest canopy height data; and forest cover obtained using MOD44B VCF data; (b) Implement spatiotemporal prediction of SAR satellite remote sensing snow depth based on ADC-GRU network and FF-DNN model. The specific implementation process is as follows: (1) Input data preparation: The data mainly includes snow depth, meteorological data, geographic information data and time data of the initial T0 of SAR satellite remote sensing images obtained based on the FF-DNN model; the meteorological data are the 2m temperature, 2m dew point temperature and atmospheric pressure obtained by ERA5 data processing in step one; the geographic information data are the elevation, slope and aspect data obtained by SRTMDEM 90M processing in step one; the time data are the date, year, month and day data of manually labeled information. (2) For different land cover types, the ADC-GRU model proposed in step 2 is used to predict the snow depth in the study area in time series. The initial T0 snow depth value is sent to the trained ADC-GRU model to predict the snow depth value of the next day T1. At the same time, this prediction result is returned as the input information for the next prediction process. (3) Iterate through step (2) to obtain the daily snow depth prediction results for the study area from the initial T0 to the final Tk.

2. The spatiotemporal prediction method for SAR satellite remote sensing snow depth based on ADC-GRU network and FF-DNN model according to claim 1, characterized in that, In step one, during data preprocessing: Sentinel-1 SAR data preprocessing selected Sentinel-1 IW GRD-level data for preprocessing. The preprocessing process mainly included: orbit correction, thermal noise removal, radiometric calibration, speckle filtering, and terrain correction. After preprocessing, the data were cropped and stitched according to the selected study area. Meteorological data preprocessing involved cropping and stitching according to the selected study area and resampling. Geographic information data preprocessing involved cropping and stitching according to the selected study area and resampling. Forest feature data preprocessing involved cropping and stitching according to the selected study area and resampling.

3. The spatiotemporal prediction method for SAR satellite remote sensing snow depth based on ADC-GRU network and FF-DNN model according to claim 1, characterized in that, Step 3 uses Sentinel-1 SAR data and auxiliary data as input data to construct the FF-DNN model: the FF-DNN model constructed for farmland areas has 13 input features, and the number of neurons in the hidden layer is 32, 64, 64 and 32 respectively; the FF-DNN model constructed for forest areas has 21 input features, and the number of neurons in the hidden layer is 32, 128, 64 and 64 respectively.

Images