A snow depth fusion method
Through the application of multi-source data processing and random forest model, the scalability and accuracy of snow-covered data fusion in the existing technology are solved, and efficient and automated snow-covered depth estimation is achieved, which is suitable for different geographical environments and data updates. .
Patent Information
- Application Number
- CN202210258103.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-16
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-16
AI Technical Summary
The prior art has problems such as low scalability, difficulty in accurately constructing nonlinear models, and requires manual participation in the fusion of snow-covered depth data, resulting in low accuracy and high computational complexity.
Multi-source data is used to obtain the initial data and impact data of snow accumulation depth, and the target initial data set is formed through reprojection transformation and space-time matching, and the random forest model is used for training and verification to achieve automated space-time continuous snow accumulation depth estimation.
It improves the accuracy and efficiency of snow depth estimation, reduces calculation overhead, and realizes automated processing without manual participation, and can adapt to different geographical environments and data updates.
Smart Images

Figure CN114818857B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of aviation and aerospace, and in particular to a snow depth fusion method. Background Art
[0002] As an important element of the earth's surface, snow plays an important role in the global terrestrial ecosystem, climate change, water cycle and energy cycle. Snow stores abundant freshwater resources, and snowmelt becomes an important source of water for surface runoff and groundwater in arid and semi-arid regions. Abnormal changes in snow cover can also trigger natural disasters such as avalanches, snowmelt floods, and windblown snow, which have an important impact on the sustainable development of regional ecology, economy and society. Snow depth (snow depth) data provides spatial distribution information and material energy information of snow cover. It is one of the important parameters to characterize the characteristics of snow cover, and is also an important parameter for studying the climate effect of snow cover, water balance in watersheds, simulation of snowmelt runoff, and monitoring and evaluating the level of snow disasters.
[0003] Prior art text (key points)
[0004] At present, most studies are to evaluate the accuracy of multiple snow depth data, and then through error statistical analysis, grasp the performance of each data under different geographical environmental factors, select one or several snow depth data for weighted average, and finally obtain spatiotemporal continuous snow depth data. The specific process is as follows:
[0005] (1) The currently available snow depth data is obtained through crawler technology, and then the collected snow depth data is preprocessed to obtain a snow depth dataset that is consistent in time and space;
[0006] (2) Obtain the pixel value of each snow depth data based on the latitude and longitude of the ground observation station;
[0007] (3) Evaluate the observation accuracy of each data in different geographical locations
[0008] (4) A weighted average fusion model of snow depth data under different geographical environments was established, and then calculations were performed at different spatial locations within the study area based on this model. Summary of the invention
[0009] The purpose of the present invention is to solve the above problems existing in the prior art.
[0010] To achieve the above-mentioned objectives, the present invention provides a snow depth fusion method, comprising: acquiring multi-source data of a first area, the multi-source data comprising: initial data of snow depth and influencing data affecting snow depth; extracting target initial data and target influencing data of a target area of the first area based on the multi-source data; acquiring spatial geographic location data of the target area, combining the target initial data, the target influencing data and the spatial geographic location data of the target area into a target initial data set; acquiring a fusion model, and inputting the target initial data set into the fusion model to obtain a spatiotemporally continuous target snow depth estimation value, wherein the fusion model is obtained through training and accuracy verification of the target initial data set.
[0011] The method of the present invention can not only be expanded as the snow depth data increases, but also can accurately and effectively fuse the existing snow depth data, becoming a method that reduces overhead and conveniently and efficiently improves the accuracy of snow depth estimation. In addition, no human participation is required in the fusion estimation process, and fully automated estimation is realized. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 It is a flow chart of the snow depth fusion technology based on weighted average in the prior art;
[0013] Figure 2 It is a flow chart of the snow depth fusion method according to an embodiment of the present invention;
[0014] Figure 3 This is the evaluation result of the snow depth fusion method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.
[0016] The existing technology methods are as follows Figure 1 As shown in the figure, most studies evaluate the accuracy of multiple snow depth data, and then perform error statistical analysis to understand the performance of each data under different geographical environmental factors, select one or several snow depth data for weighted averaging, and finally obtain spatiotemporal continuous snow depth data. The specific process is as follows:
[0017] (1) The currently available snow depth data is obtained through crawler technology, and then the collected snow depth data is preprocessed to obtain a snow depth dataset that is consistent in time and space;
[0018] (2) Obtain the pixel value of each snow depth data based on the latitude and longitude of the ground observation station;
[0019] (3) Evaluate the observation accuracy of each data in different geographical locations
[0020] (4) A weighted average fusion model of snow depth data under different geographical environments was established, and then calculations were performed at different spatial locations within the study area based on this model.
[0021] However, the existing technology has certain technical defects:
[0022] (1) The snow depth fusion technology based on weighted average is based on the evaluation results of snow depth, and a decision tree is established based on this evaluation result. Then, for each decision tree, a selection criterion is established or a fusion weight value corresponding to different snow depth data is assigned. By making a decision classification based on the geographical environment characteristics of the study area, a spatiotemporal continuous snow depth data is finally obtained. The result of this method depends on the precision of the binary tree construction. However, in practical applications, it is difficult to fully consider the influencing factors, and the influencing factors are also interrelated, so it is difficult to accurately establish a decision tree.
[0023] (2) The scalability of this method is not high. If there is new snow depth data or model, it is necessary to re-evaluate the accuracy of the added snow depth data and build a new decision tree.
[0024] (3) This method only performs a simple weighted average on each decision tree, and often cannot accurately construct a relationship model between the input snow depth data and the target output snow depth, such as a nonlinear model.
[0025] Figure 2 Flow chart of a snow depth fusion method according to an embodiment of the present invention. According to the present invention, a snow depth fusion method is implemented, including:
[0026] S100: Acquire multi-source data of a first area, where the multi-source data includes: initial data of snow depth and influencing data that affects the snow depth.
[0027] In this step, the initial data of snow depth in the first area is obtained. The first area can be anywhere in the world, that is, the mainstream snow depth data currently available at home and abroad is obtained, mainly including WESTDC, MERRA2, ERA-Interm and CMC data. The data that interferes with the snow depth can be land cover type data, snow type, forest coverage data and surface roughness data.
[0028] In some embodiments of the present invention, the snow depth can be obtained fully automatically through a Python program, so that the information can be automatically obtained without human intervention, which is convenient and fast, and improves efficiency.
[0029] Furthermore, before the target initial data, target impact data and spatial geographic location data of the target area are merged into the target initial data set, the target initial data needs to be reprojected to obtain transformed initial data, and the target impact data needs to be spatially and temporally matched with the transformed initial data.
[0030] Specifically, the collected initial data of snow depth were reprojected and transformed into WGS84 projection, and the nearest neighbor method was used to resample the collected initial data of snow data to 0.25°×0.25°. In addition, the collected influencing data such as terrain + environmental factors were matched with the existing snow depth data in time and space using the nearest difference method, with a spatial resolution of 0.25°×0.25° and a projection coordinate system of WGS84.
[0031] It is understandable that most of the snow depth data and impact data to be fused are raster data, and the initial resolutions of the collected snow depth data and impact data are inconsistent. In order to facilitate subsequent data fusion, the resolution of all data is unified to 0.25°×0.25° to ensure the uniformity of the raster and prevent the data from being inaccurately matched during the subsequent fusion process.
[0032] Furthermore, after reprojecting and resampling all acquired initial data and influencing data, the default values of the target initial data and the target influencing data of the target area need to be eliminated.
[0033] When obtaining snow depth data from all over the world, the corresponding observation values may not be obtained due to certain reasons. Therefore, a unique value will be used to identify it in the data record. This part of the data is collectively referred to as the default value.
[0034] Therefore, in the actual implementation of the method of the embodiment of the present invention, it is necessary to eliminate this part of data to ensure the correctness of subsequent calculations.
[0035] S200: extracting target initial data and target impact data of a target area of a first area based on multi-source data;
[0036] Specifically, as can be seen from the above, the initial data and impact data of the publicly available snow depth in the world, that is, all the data of the first area, have been obtained through the automatic acquisition method. The snow depth fusion method as an embodiment of the present invention needs to calculate the snow depth data of the target area, so the target initial data and target impact data about the snow depth of the target area can be extracted from the multi-source data.
[0037] The target initial data for the extracted snow depth are WESTDC, MERRA2, ERA-Interm and CMC data, and the target influencing data are land cover type, snow type data, DEM elevation data and surface roughness data.
[0038] S300: Acquire spatial geographic location data of a target area, and combine target initial data, target impact data, and spatial geographic location data of the target area into a target initial data set.
[0039] In this step, the target area is the area where the snow depth is to be calculated, and the corresponding location information, that is, the spatial geographic location data, is obtained, and then combined with the target initial data and target impact data corresponding to the location, the target initial data set of the data at the location can be obtained.
[0040] The specific steps of combining the target initial data, target impact data and spatial geographic location data into the target initial data set are as follows:
[0041] First, the first pixel value and the second pixel value of the target initial data and the target influencing data of the target area are respectively obtained, that is, the pixel value of the target initial data, i.e., the first pixel value, and the pixel value of the target influencing data, i.e., the second pixel value, are respectively obtained.
[0042] In other words, the initial data of snow depth mentioned above come from all over the world, and the initial data obtained at different locations are different. Therefore, the spatial geographic location data is the corresponding location information of the obtained initial data of snow depth.
[0043] For example, all the initial data can come from weather stations around the world, and the longitude and latitude of the spatial location of the weather station is the spatial geographic location data. After spatiotemporal matching, the pixel value of the target initial data of the snow depth at the corresponding location is extracted. As mentioned above, the initial data of snow depth can be WESTDC, MERRA2, ERA-Interm and CMC data. At the same time, the terrain and environmental data of the location corresponding to the spatial geographic location data can also be obtained through spatiotemporal matching, that is, the target impact data mentioned above, which can specifically include: land cover type, snow type data, DEM elevation data and pixel values of surface roughness data.
[0044] Furthermore, the acquired spatial geographic location data and the pixel values of the target initial data and the target impact data of the snow depth matched with the target area corresponding to the spatial geographic location data are combined into a data set.
[0045] S400: Obtain a fusion model, and input the target initial data set into the fusion model to obtain a spatiotemporally continuous target snow depth estimation value, wherein the fusion model is obtained through training and accuracy verification of the target initial data set.
[0046] The first pixel value, the second pixel value and the spatial geographic location data are fused into the target initial data set, thereby forming a data set that is input into the model as raw data to obtain the spatiotemporally continuous snow depth estimation value of the target area.
[0047] Furthermore, in the method of the embodiment of the present invention, the fusion model used is a random forest model, which includes multiple random tree models. The data of the target initial data set are respectively input into the multiple random tree models to obtain multiple snow depth estimation values.
[0048] Those skilled in the art will appreciate that the random forest model is composed of multiple random tree models. Therefore, multiple snow depth estimates can be obtained by inputting the data in the above target initial data set into multiple random tree models respectively.
[0049] The random forest model used is a combination classification algorithm of ensemble learning, which has strong anti-overfitting ability. In addition, the random forest model uses multiple decision trees to fit regression models in different geographical environments, which can adaptively solve the problem of inconsistent regression models in different geographical environments. If you need to add new input parameters, you only need to organize the parameters into the corresponding input format and import them into the trained random forest model. It will be automatically updated without manual intervention and selection.
[0050] At the same time, in order to reduce the error, the average of multiple snow depth estimates is taken as the target snow depth estimate. This ensures that a relatively accurate snow depth estimate, i.e., the target snow depth estimate, can be obtained by using the random tree force model. The formula for calculating the snow depth estimate by the random tree model is as follows:
[0051] SD=f(SD WESTDC ,SD MERRA2 ,SD ERA-Interim ,SD CMC ,lat,lon,date,DEM,ff,Rou,SC)
[0052] Among them, SD is the estimated value of snow depth, f() is the corresponding function expression of the random tree model, SD WESTDC is the snow depth value of WESTDC data, SD MERRA2 is the snow depth value of MERRA2 data, SD ERA-Interm is the snow depth value of ERA-Interim data, SD CMCis the snow depth value of CMC data, lat is latitude, lon is longitude, date is the time corresponding to the estimated snow depth, DEM is the elevation value corresponding to the estimated snow depth, ff is forest coverage, Rou is surface roughness, and SC is the corresponding snow type.
[0053] In the snow depth fusion method of the embodiment of the present invention, the random tree model used may have errors when used directly, so the target initial data set is input into the random forest model to make the final result more perfect. At the same time, it can ensure that the final model is more robust and accurate. Therefore, the random forest model needs to be trained and verified.
[0054] The accuracy of the model can be verified by dividing the data of the acquired target initial data set into two parts, one as a training and construction data set, and the other as a verification data set.
[0055] Specifically, the data of the target initial data set can be divided by the "ten-fold" cross-validation method. For example, 70% of the data can be randomly selected from the target initial data set as the training data set, and 30% of the data can be randomly selected from the target initial data set as the verification data set.
[0056] For example, the training data set is input into the random forest model to update the parameters of the random forest model and obtain the learned random forest model. The learning method is to calculate the training data set according to the above formula, so that the random forest model can be learned and automatically updated.
[0057] In one embodiment of the present invention, when establishing a random tree model of a random forest model, the number of random trees can be optimized according to a grid search method to obtain a curve showing the change in the prediction accuracy of the random forest model and the number of random trees. Secondly, it can be found that the prediction error of the random forest model decreases as the number of constructed random trees increases, and tends to be stable when the number of trees is 50, indicating that constructing 50 trees is sufficient to achieve a higher prediction accuracy, which reduces the amount of calculation to a certain extent compared with a model with 200 trees, so the number of trees is set to 50.
[0058] Furthermore, by inputting the validation dataset into the learned random forest model, the accuracy of the snow depth estimated by this method can be quantified by adopting quantitative accuracy evaluation indicators, including correlation coefficient (R) and root mean square error (RMSE).
[0059] The calculation of the correlation coefficient is achieved by the following formula:
[0060]
[0061] Among them, SD d,i and SD obs,i Represent the snow depth estimate and ground observation value / cm respectively, N represents the number of samples, where the unit of the number of samples is pieces. Therefore, the correlation coefficient can reflect the correlation between variables and is used to measure the closeness between data and observations. That is, the larger the correlation coefficient, the smaller the error of the trained random forest model.
[0062] The calculation of the root mean square error is achieved through the following formula:
[0063]
[0064] Among them, SD d,i and SD obs,i Represent the snow depth estimation value and ground observation value / cm respectively, i represents the i-th sample data, N represents the data volume, and the unit of data volume is piece. Therefore, by calculating the root mean square error, the square root of the ratio of the sum of squares of the deviations between the snow depth value of the data to be evaluated and the observed value to the number of observations can be obtained, which is used to measure the degree of deviation between the evaluation data and the observed value.
[0065] Thus, a random forest model with high accuracy after training and validation was obtained, so that the model can be applied to the calculation of snow depth in the target area to obtain spatiotemporally continuous target snow depth estimates.
[0066] like Figure 3 As shown, it can be seen that Figure 3 In the several accompanying drawings, Num represents the number of samples, R represents the sample coefficient, RMSE represents the correlation coefficient, the horizontal axis of the coordinate system represents the ground observation value, and the vertical axis represents the estimated value.
[0067] Figure 3 Figure (a) shows the calculation results after random forest fusion. It can be seen that all data are scattered near the 1:1 line, indicating that the error is small. Figure 3 As can be seen from Figures (b) to (f), all the points are scattered irregularly, and the errors are relatively large.
[0068] The present invention is not limited to this. The input parameters of the present invention, such as the initial data and influencing data of snow depth, can be added as the types of the disclosed initial data of snow depth increase. In addition, as the research goes deeper, more auxiliary factor parameters can be introduced into the types of influencing data.
[0069] With the rapid development of space-ground stereoscopic observation technology, more and more satellite sensors are launched, remote sensing inversion algorithms and model simulation technologies are constantly improving and perfecting, which will also lead to the emergence of more snow depth data. The data publishing platform will receive massive spatiotemporal data produced by different algorithms, models or sensors from different scientific research institutions. However, it is also accompanied by many problems, such as large amount of data, wide sources and uneven quality.
[0070] The present invention integrates the currently available snow depth data, integrates the advantages of observation data from different sources, realizes the complementary advantages of various data sources, and thus improves data quality. It is a method for improving the utilization rate of existing data and efficiently improving the current snow depth estimation accuracy, making up for the lack of high-precision snow depth data in climate models and disaster monitoring, and providing a reliable basic data source for water resources assessment and management, disaster monitoring and forecasting, and climate research.
[0071] In summary, if Figure 2 As shown, the snow depth fusion method of the embodiment of the present invention first pre-processes the collected multi-source data. The pre-processing mainly includes projection transformation, spatial resampling, image cropping, spatiotemporal matching and default value elimination, which are converted into a data set of uniform spatiotemporal scale to form a sample data set, and then the sample data set is randomly divided into a training data set and a test data set, i.e., a first data set and a second data set. The training data set, time, geospatial information and other auxiliary information are used to train and learn the random forest model, and finally the constructed random forest fusion model is used to estimate the snow depth.
[0072] In other words, the snow depth fusion method of the embodiment of the present invention uses the existing random forest machine learning algorithm as the basic model, adopts the current mainstream snow depth data (WESTDC, MERRA2, ERA-Interterm and CMC data), combined with the daily snow depth observed by China's meteorological stations, and introduces NASA's SRTM digital elevation model (DEM) data, land cover type data, snow type data and forest coverage data, performs data preprocessing and time-space matching processing, constructs a sample data set, estimates the daily snow depth by establishing a regression relationship model, and fuses the satellite inverted snow depth, model simulated snow depth, Canadian reanalysis snow depth data and China's meteorological station interpolated snow depth to achieve a high-precision long-term series snow depth estimation model.
[0073] The snow depth fusion method of the embodiment of the present invention is based on the random forest model. By learning and training the existing snow depth data, the method can automatically establish snow depth fusion decision trees in different geographical environments. With the help of the random forest machine learning model, the currently existing mainstream snow depth data is effectively integrated together, which becomes a method that reduces overhead and is convenient and efficient, and improves data quality by inheriting the advantages of each data.
[0074] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A snow depth fusion method, characterized in that: include: Acquire multi-source data of the first area, the multi-source data comprising: initial data of snow depth and influencing data affecting the snow depth; Based on the multi-source data, extracting target initial data and target influence data of the target area of the first area; performing reprojection transformation on the target initial data to obtain transformed initial data, and performing spatiotemporal matching between the target influence data and the transformed initial data; Acquiring the spatial geographic location data of the target area, and combining the target initial data, the target impact data and the spatial geographic location data of the target area into a target initial data set, including: respectively acquiring the first pixel value and the second pixel value of the target initial data and the target impact data of the target area; combining the first pixel value, the second pixel value and the spatial geographic location data into the target initial data set; Obtain a fusion model, input the target initial data set into the fusion model to obtain a spatiotemporally continuous target snow depth estimation value, The fusion model is obtained through the training and accuracy verification of the target initial data set, and the fusion model includes: a random forest model, which includes multiple random tree models, and the data of the target initial data set are respectively input into the multiple random tree models to obtain multiple snow depth estimation values; the average of the multiple snow depth estimation values is taken as the target snow depth estimation value; The multiple random tree models calculate the snow depth estimation value by the following formula: , Wherein, SD is the estimated value of snow depth, f() is the corresponding function expression of the random tree model, SD WESTDC is the snow depth value of WESTDC data, SD MERRA2 is the snow depth value of MERRA2 data, SD ERA-Interm is the snow depth value of ERA-Interim data, SD CMC is the snow depth value of CMC data, lat is latitude, lon is longitude, date is the time corresponding to the estimated snow depth, DEM is the elevation value corresponding to the estimated snow depth, ff is forest coverage, Rou is surface roughness, and SC is the corresponding snow type.
2. The method according to claim 1, characterized in that After obtaining the multi-source data of the first region, the method further includes: The default values of the target initial data and the target impact data of the target area are eliminated.
