Snow depth prediction method, storage medium and electronic equipment

Abstract

In order to solve the problem that the snow depth predicted by a sentinel No.1 data set is difficult to meet high spatial resolution and high temporal resolution at the same time, the invention provides a snow depth prediction method, a storage medium and electronic equipment, and the method comprises the steps: obtaining sentinel No.1 data, auxiliary data and station data about the snow depth, constructing a geographic time weighted snow depth parameter based on the site data; carrying out resampling and space-time matching on the data to obtain a first data set; inputting the first data set into an integrated deep learning model to obtain a preliminary snow depth; obtaining time sequence climate data and geographical time data of the research area, performing resampling, and converting the preliminary snow depth data into time sequence data; the resampled time sequence climate data and geographical time data and the time sequence data of the initial snow depth form a second data set; and inputting the second data set into the ResNet-BiLSTM time sequence model to obtain snow depth data. According to the invention, snow depth prediction data with high temporal-spatial resolution can be generated.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and particularly to a snow depth prediction method, a storage medium, and an electronic device. Background Art

[0002] Snow depth is one of the important parameters describing snow cover characteristics, and its change is of great significance for global hydrological cycle, climate change research, and ecological environment protection. Snow cover is an important water resource reserve in alpine, polar, and cold regions, and its melting process directly affects river runoff, water resource management, and agricultural irrigation. At the same time, snow significantly affects the Earth's hydrological cycle and energy balance due to its heat insulation performance and high reflectivity, and plays a crucial role in regional and global climate systems.

[0003] Field observations and remote sensing techniques are common methods for obtaining snow depth. Although manual measurements or automatic sensors can provide high-quality time-series snow depth data, due to the uneven distribution of observation stations, it is impossible to characterize the overall snow cover distribution and changes in the entire region. With the progress of remote sensing technology, using remote sensing data to monitor snow depth has gradually become the mainstream, and common means include optical remote sensing, passive microwave remote sensing, and active microwave remote sensing. Optical remote sensing obtains snow cover information by detecting the optical properties of snow cover (such as reflectivity), and then estimates snow depth. Its advantage is relatively high spatial resolution, which can clearly capture local snow cover changes. However, optical remote sensing is highly sensitive to weather conditions, and cloud cover will seriously interfere with data acquisition, making it difficult to stably apply in cloudy or snowfall-frequent regions. Passive microwave remote sensing uses the radiation characteristics of snow cover to invert snow depth through microwave brightness temperature data. The greatest advantage of this method is its all-weather and all-time observation ability, which can work under cloud cover and night conditions. In addition, passive microwave remote sensing has a wide coverage range and is suitable for large-scale snow depth monitoring. However, its spatial resolution is relatively low (usually 10 kilometers or lower), making it difficult to meet the requirements for high-resolution snow depth data in complex terrain regions. Active microwave remote sensing (such as synthetic aperture radar, SAR) detects the scattering characteristics of snow cover by emitting microwave signals and receiving echo signals. Compared with passive microwave, active microwave remote sensing can provide higher spatial resolution and is particularly suitable for capturing snow cover changes in complex terrain regions. Sentinel-1 is a new generation of C-band synthetic aperture radar (SAR) satellite launched by the European Space Agency (ESA), which has all-weather, all-time, high-resolution, and large-scale observation capabilities, providing strong data support for snow depth inversion. Its high spatial resolution (about 20 meters) makes it possible to monitor snow cover changes in regions with complex terrain and variable climate. However, the time resolution of Sentinel-1 data is relatively limited (usually 6 to 12 days), making it difficult to generate a dense time-series snow depth dataset and difficult to meet the requirements for high time resolution while satisfying high spatial resolution. Summary of the Invention

[0004] The object of the present invention is to propose a snow depth prediction method to solve the problem that it is difficult for the predicted snow depth of Sentinel-1 datasets to simultaneously meet high spatial resolution and high temporal resolution. The method includes the following steps:

[0005] S1. Obtain Sentinel-1 data, auxiliary data, and station data related to snow depth, and construct a geotemporal weighted snow depth parameter based on the station data; resample and perform spatio-temporal matching on the Sentinel-1 data, auxiliary data, and geotemporal weighted snow depth parameter to obtain a first dataset;

[0006] S2. Construct an integrated deep learning model through multiple deep learning models, and input the first dataset into the integrated deep learning model to obtain a preliminary snow depth;

[0007] S3. Obtain the temporal climate data and geotemporal data of the study area, resample the temporal climate data and geotemporal data according to the preliminary snow depth, and convert the preliminary snow depth data into temporal data; the resampled temporal climate data and geotemporal data, and the temporal data of the preliminary snow depth constitute a second dataset;

[0008] S4. Input the second dataset into a ResNet-BiLSTM temporal model to predict and obtain a snow depth dataset.

[0009] Further, obtaining the Sentinel-1 data related to snow depth specifically includes:

[0010] Download the original Sentinel-1 data and perform preprocessing, including: removing thermal noise, orbit correction, noise filtering, radiometric calibration, Doppler terrain correction, and finally performing decibelization processing, outputting the backscattering coefficients in two polarization modes, and calculating the cross-polarization backscattering coefficient.

[0011] Further, the auxiliary data includes: MODIS surface vegetation cover data, MODIS land cover type data, ERA5-Land reanalysis dataset providing snow property information, and elevation and slope information extracted from ASTER GDEM V3 data.

[0012] Further, the calculation formula of the geotemporal weighted snow depth parameter is as follows:

[0013]

[0014] where STsnow represents the geotemporal weighted snow depth parameter, m represents the number of days, k j represents the time weight of the j-th day, n represents the number of neighbor stations of the station to be measured, snow-depth ij represents the snow depth observation value of the i-th neighbor station of the station to be measured on the j-th day, w iRepresents the spatial weight of the i-th neighbor site of the site to be measured, d i Represents the distance between the site to be measured and the i-th neighbor site, and σ is a dynamic parameter.

[0015] Furthermore, the integrated deep learning model is stacked by DNN, CNN, LSTM, and Transformer through the Stacking method.

[0016] Furthermore, the time-series climate data includes: precipitation, temperature, dew point temperature, and vapor pressure deficit; the geographical time data includes: longitude, latitude, date, elevation, and slope.

[0017] The present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above snow depth prediction method is implemented.

[0018] The present invention also proposes an electronic device, including a processor and a memory. The processor is interconnected with the memory. Among them, the memory is used to store a computer program, and the computer program includes computer-readable instructions. The processor is configured to call the computer-readable instructions to execute the above snow depth prediction method.

[0019] The beneficial effects brought by the technical solution provided by the present invention are:

[0020] The present invention first constructs a geographical time-weighted snow depth parameter using site data. This parameter not only captures the cumulative or melting trend of snow depth over time but also considers the influence of geographical location on the snow depth distribution. Using the high spatial resolution information of Sentinel-1, auxiliary data, and the geographical time-weighted snow depth parameter, an integrated deep learning model is used to generate high spatial resolution snow depth data. The high spatial resolution snow depth data is combined with time-series climate data and geographical time data, and a ResNet-BiLSTM time-series model is used to generate high spatio-temporal resolution snow depth data. The present invention can obtain snow depth prediction data with high spatial resolution and high time resolution. Description of the Drawings

[0021] Figure 1 Is a flowchart of the snow depth prediction method in an embodiment of the present invention;

[0022] Figure 2 Is a block diagram of an electronic device in an exemplary embodiment of Embodiment 1 of the present invention;

[0023] Figure 3 Is a high spatio-temporal resolution snow depth image predicted by the method of the embodiment of the present invention. Detailed Embodiments

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0025] The flowchart of the snow depth prediction method according to the embodiment of the present invention is as Figure 1 , and specifically includes the following steps:

[0026] S1. Obtain Sentinel-1 data, auxiliary data, and station data related to snow depth, and construct a spatio-temporal weighted snow depth parameter based on the station data; resample and perform spatio-temporal matching on the Sentinel-1 data, auxiliary data, and spatio-temporal weighted snow depth parameter related to snow depth to obtain a first data set.

[0027] Specifically, first download the Sentinel-1 data of the study area from the official website of the European Space Agency, and use the Python API of the ESA Sentinel application platform (SNAP) to preprocess the data, including operations such as thermal noise removal, orbit correction, noise filtering, radiometric calibration, and Doppler terrain correction. Finally, generate the backscattering coefficient after decibelization processing, and calculate the cross-polarized backscattering coefficient. At the same time, download the required auxiliary data, including the tree coverage rate (PTC) and non-vegetation coverage rate (PNV) provided by the MODIS surface vegetation coverage data (MOD44B), the land cover type (LC) provided by the MODIS land cover type data (MCD12Q1), the snow surface temperature, snow density, snow albedo and other snow property information provided by the ERA5-Land reanalysis dataset, and the elevation and slope information extracted from the ASTER GDEM V3 data. The station data is sourced from the snow depth station dataset of the Global Historical Climatology Network (GHCN), and a spatio-temporal weighted snow depth parameter STsnow is constructed based on the snow depth station data.

[0028] The construction of STsnow is based on the spatio-temporal continuity characteristics of snow depth changes. The change of snow depth does not occur instantaneously, but is affected by the cumulative impact of climatic conditions in the previous few days or even longer. The process of snow melting or accumulation is affected by a combination of factors such as temperature, precipitation, and solar radiation, and these factors have obvious continuity in the time series. In addition, the snow depth conditions of surrounding stations can provide important environmental background information for a certain station, because stations with similar geographical locations may face similar climatic conditions and topographic features, and thus have similar snow depth change trends. This method not only captures the cumulative or melting trend of snow depth over time, but also considers the influence of geographical location on snow depth distribution, providing a more accurate snow depth measurement and prediction method. The calculation method of the spatio-temporal weighted snow depth parameter (STsnow) is as follows:

[0029]

[0030] Among them, STsnow represents the geotemporal weighted snow depth parameter, m represents the number of days of the selected snow depth data, and k j represents the time weight of the j-th day, and the weight becomes smaller as time goes by. n represents the number of neighboring stations of the station to be measured, and snow-depth ij represents the snow depth observation value of the i-th neighboring station of the station to be measured on the j-th day, and w i represents the spatial weight of the i-th neighboring station of the station to be measured.

[0031] w i Adopts a non-linear weight calculation method based on the Gaussian function, and the calculation formula is as follows:

[0032]

[0033] d i represents the distance between the station to be measured and the i-th neighboring station, and σ is a dynamic parameter, designed to be half of the average distance.

[0034] MOD44B provides the tree cover rate (PTC) and the non-vegetation cover rate (PNV), and its spatial resolution is 250 meters; MCD12Q1 provides land cover type data (LC), and its spatial resolution is 500 meters; ERA5-Land provides relevant information on snow properties in the area, including snow surface temperature, snow density, and snow albedo, and its resolution is 0.1 ° ; After processing, GDEMV3 extracts elevation and slope data, and its resolution is 30 meters; all data are resampled to a resolution of 20 meters to match Sentinel-1 data, and spatio-temporal matching is performed according to the station information to obtain the first data set.

[0035] S2. Construct an integrated deep learning model through multiple deep learning models, input the first data set into the integrated deep learning model to obtain the preliminary snow depth, and the preliminary snow depth data has a high spatial resolution.

[0036] Ensemble learning is a machine learning technique that improves the overall prediction performance by combining multiple models. Common ensemble methods include Bagging, Boosting, and Stacking. Bagging performs multiple random samplings with replacement on the data, trains multiple independent base learners, and then fuses their prediction results through methods such as averaging or voting. Typical algorithms include Random Forest. Boosting iteratively trains multiple base learners, focusing on the misclassified samples in the previous round in each iteration and gradually reducing the overall prediction error. Representative algorithms include AdaBoost and XGBoost. Stacking is a method that takes the prediction results of multiple basic learners as input and then trains a meta-learner (or called a combined model) for the final prediction. With the rapid development of deep learning technology, the idea of ensemble learning has also been introduced into deep learning, forming the "ensemble deep learning" method. Ensemble deep learning further improves the performance and generalization ability of the model by combining the prediction results of multiple deep learning models. Using the Stacking method to stack DNN, CNN, LSTM, and Transformer deep learning methods according to the data characteristics can give full play to the characteristics of each model while combining the advantages of all models, resulting in better prediction results compared to a single model.

[0037] S3. Obtain the temporal climate data and geotemporal data of the study area, resample the temporal climate data and geotemporal data according to the preliminary snow depth, and convert the preliminary snow depth data into temporal data; the resampled temporal climate data and geotemporal data, as well as the temporal data of the preliminary snow depth, constitute the second dataset.

[0038] Based on the high-spatial-resolution snow depth data predicted in step 2, obtain the temporal climate data of the study area, including precipitation, temperature, dew point temperature, and vapor pressure deficit, and resample it to 20 meters to match the snow depth grid data. At the same time, extract the geotemporal information corresponding to each pixel point, including longitude, latitude, date, elevation, and slope. The time interval of the high-spatial-resolution snow depth data predicted in step 2 is 12 days, and there are real values on the 1st and 13th days of the snow depth sequence. The missing values in the middle of the sequence are filled with -1.

[0039] S4. Input the second dataset into the ResNet-BiLSTM temporal model to predict the snow depth dataset.

[0040] Among them, ResNet solves the problem of gradient vanishing in deep networks through residual connections. Its multi-level residual blocks can efficiently fuse local details and global spatial features, enhancing the ability to model non-linear relationships in complex geographical data. BiLSTM captures the temporal dependencies before and after the sequence simultaneously through a bidirectional recursive structure. Its input gate, forget gate, and output gate mechanisms can precisely control the long-term and short-term information flow, being able to trace the historical trend of snow accumulation and predict the potential relevance of the ablation process. The combination of the two achieves a deep synergy of spatial-temporal features: the multi-scale spatial features extracted by ResNet provide inputs with high information density for temporal modeling, while BiLSTM analyzes the causal and feedback mechanisms in the time dimension through bidirectional dynamic weight allocation. This architecture takes into account spatial heterogeneity and temporal evolution laws in snow depth prediction. The high-resolution features of ResNet can preserve terrain details, and the bidirectional dependency modeling of BiLSTM can accurately depict the dynamic process of snow cover, ultimately generating a high-precision snow depth dataset with both spatio-temporal continuity and physical consistency, providing reliable support for remote sensing quantitative inversion.

[0041] In an exemplary embodiment, there is provided a computer-readable storage medium storing a computer program which, when executed by a processor, implements the above-described snow depth prediction method.

[0042] Please refer to Figure 2 , in an exemplary embodiment, there is further provided an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0043] Among them, a computer program is stored on the memory. The computer program includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-described snow depth prediction method.

[0044] To verify the effectiveness of the method of the present invention, Sentinel-1 images and auxiliary data of the corresponding study area were downloaded from a relevant data platform, and a series of data preprocessing operations were performed, including data calibration, registration, and noise removal, etc., to ensure the accuracy and availability of the data. After the preprocessing was completed, the obtained data was processed according to the method of the present invention to obtain high spatio-temporal resolution snow depth. The high spatio-temporal resolution snow depth image predicted by the method of the embodiment of the present invention refers to Figure 3 , Figure 3 shows the high spatio-temporal resolution snow depth prediction images for a total of 8 days from February 1, 2022 to February 8, 2022. The scale on the right indicates the snow depth in centimeters (cm). It can be seen from the figure that as time goes by, the changing trend of snow accumulation gradually appears.

[0045] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A snow depth prediction method, characterized in that: The following steps are involved: S1. Acquire Sentinel-1 data, auxiliary data, and station data on snow depth, and construct geo-time weighted snow depth parameters based on the station data; resample and perform spatiotemporal matching on the Sentinel-1 data, auxiliary data, and geo-time weighted snow depth parameters to obtain a first data set; S2. construct an integrated deep learning model through multiple deep learning models, input the first data set into the integrated deep learning model, and obtain a preliminary snow depth; S3, obtaining time series climate data and geographic time data of the study area, resampling the time series climate data and geographic time data according to the preliminary snow depth, and converting the preliminary snow depth data into time series data; the resampled time series climate data and geographic time data, and the time series data of the preliminary snow depth constitute a second data set; S4. Input the second data set into the ResNet-BiLSTM time series model to predict the snow depth data set.

2. A snow depth prediction method according to claim 1, characterized in that: Get Sentinel-1 data on snow depth, specifically: The original Sentinel-1 data was downloaded and preprocessed, including thermal noise removal, orbit correction, noise filtering, radiation calibration, Doppler terrain correction, and finally decibel processing to output the backscatter coefficients under two polarization modes and calculate the cross-polarization backscatter coefficient.

3. A snow depth prediction method according to claim 1, characterized in that: Auxiliary data include: MODIS surface vegetation cover data, MODIS land cover type data, ERA5-Land reanalysis dataset providing snow attribute information, and elevation and slope information extracted from ASTER GDEMV3 data.

4. A snow depth prediction method according to claim 1, characterized in that: The calculation formula of the geographic time-weighted snow depth parameter is as follows: Among them, STsnow represents the geographic time weighted snow depth parameter, m represents the number of days, k j represents the time weight of the jth day, n represents the number of neighboring sites of the site to be tested, snow-depth ij represents the snow depth observation value of the i-th neighboring station of the tested station on the j-th day, w i represents the spatial weight of the i-th neighboring site of the site to be tested, d i represents the distance between the site to be tested and its ith neighbor site, and σ is a dynamic parameter.

5. A snow depth prediction method according to claim 1, characterized in that: The integrated deep learning model is formed by stacking DNN, CNN, LSTM and Transformer through the Stacking method.

6. A snow depth prediction method according to claim 1, characterized in that: Time series climate data include: precipitation, air temperature, dew point temperature and evaporation pressure difference; geographic time data include: longitude and latitude, date, elevation and slope.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1 to 6.

Images