CN-122020092-A - Grassland snow depth inversion method based on multi-source remote sensing and machine learning

CN122020092ACN 122020092 ACN122020092 ACN 122020092ACN-122020092-A

Abstract

The invention discloses a grassland snow depth inversion method based on multi-source remote sensing and machine learning, which comprises the steps of firstly obtaining multi-source input data and extracting characteristic variables for snow depth inversion, secondly carrying out correlation analysis and screening to obtain effective input characteristics, thirdly constructing a training sample set and training a machine learning integrated model, fourthly carrying out inversion calculation on the snow depth of a target grassland area and carrying out accuracy verification on inversion results, fifthly carrying out snow depth risk grade division on the target grassland area, sixthly carrying out prediction on the snow depth of a future period of the target grassland area.

Inventors

BAI HAIHUA
CHUN LIANG
Hong Mingfa
LU BAOFENG
GUO XIAOMENG
SUN JIALI
TIAN SHICHAO

Assignees

中国农业科学院草原研究所

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (7)

1. The grassland snow depth inversion method based on multi-source remote sensing and machine learning is characterized by comprising the following steps of: Firstly, acquiring multi-source input data, preprocessing the multi-source input data, and extracting characteristic variables for snow depth inversion; Step two, carrying out correlation analysis and screening on the extracted characteristic variables based on the Pearson correlation coefficient, and reserving the characteristic obviously related to the snow depth target variable as an effective input characteristic; Thirdly, constructing a training sample set by using the screened effective input features and corresponding snow depth sample investigation data, constructing a machine learning integrated model for training, and establishing a nonlinear mapping relation from the multi-source features to the snow depth; Step four, utilizing a trained machine learning integrated model to carry out inversion calculation on the snow depth of the target grassland area, and utilizing an independent verification data set to carry out accuracy verification on an inversion result; Fifthly, based on the snow depth space distribution data obtained through inversion, combining a preset snow depth threshold value, and carrying out snow depth risk level division on a target grassland area; and step six, inputting weather forecast data of a future period into a trained machine learning integrated model, and predicting the snow depth of the future period of the target grassland area.
2. The method for inverting the snow depth of a grassland based on multi-source remote sensing and machine learning according to claim 1, wherein the multi-source input data in the first step comprises satellite remote sensing image data, analysis meteorological data, terrain data, ground actual measurement grass height data and grassland snow depth pattern investigation data, wherein the satellite remote sensing image comprises Sentinel-2 or MODIS data, and the analysis meteorological data is ERA5 data.
3. The method for inverting the snow depth in the grassland based on multi-source remote sensing and machine learning as set forth in claim 1, wherein the characteristic variables in the first step include normalized snow index, normalized vegetation index, accumulated precipitation in snow season, geographic position and topography parameters, and grass height, wherein the vectors of the geographic position and topography parameters include longitude, latitude and elevation, and the grass height is obtained through ground actual measurement and is used for quantifying and characterizing the blocking effect of vegetation on wind and snow flow and the trapping effect of snow.
4. The grassland snow depth inversion method based on multi-source remote sensing and machine learning as set forth in claim 1, wherein the screening process in the second step is specifically to calculate the pearson correlation coefficient r between each characteristic variable and the snow depth target variable, and then screen the characteristic variables satisfying |r| >0.2 as effective input characteristics, wherein the calculation formula of r is represented by the following formula Wherein n is the number of samples, X i and Y i are the characteristic value and the snow depth value of the ith sample respectively, And The sample mean values of the characteristic variable and the target variable are respectively.
5. The method for inverting the snow depth of the grassland based on multi-source remote sensing and machine learning according to claim 1, wherein in the third step, the machine learning integrated model structure is represented by SD=ML [ H grass ,P accum ,I NDSI ,I NDVI ,T geo ], wherein SD is the depth of the snow depth of the grassland obtained by inversion, ML represents a machine learning model option, one or more of extreme gradient lifting tree, random forest, support vector regression and K nearest neighbor, H grass is the height of the grassland, P accum is the accumulated precipitation in snow season, I NDSI is a normalized snow index, I NDVI is a normalized vegetation index, and T geo is a geographic position and terrain parameter vector.
6. The grassland snow depth inversion method based on multi-source remote sensing and machine learning according to claim 1, wherein in the third step, training is performed on the model by using a constructed training sample set, and model parameter tuning and performance evaluation are performed by adopting a cross-validation method during training.
7. The grassland snow depth inversion method based on multi-source remote sensing and machine learning according to claim 1 is characterized in that the index of accuracy verification in the fourth step comprises root mean square error, average absolute error and decision coefficient, and the result of the snow depth risk level division in the fifth step is used for generating grassland snow disaster monitoring and early warning information.

Description

Grassland snow depth inversion method based on multi-source remote sensing and machine learning Technical Field The invention relates to the technical field of equipment, in particular to a grassland snow depth inversion method based on multi-source remote sensing and machine learning. Background The depth of snow is an important physical parameter for describing snow, and is a direct index for describing the coverage condition and thickness variation of snow. The accumulation and ablation process of the snow is affected by multiple factors such as climate conditions, topography, surface coverage types and the like, and is a complex nonlinear process, while the traditional snow depth inversion model is often constructed into an empirical linear regression model, so that the complex nonlinear relationship under the combined action of multiple factors in the snow forming process is difficult to describe. The machine learning model has strong nonlinear fitting capability and variable selection capability, can fully mine potential relation between multi-source remote sensing data and environmental factors, has been widely studied and applied in snow depth estimation, but most of research areas are selected in mountain areas or forests, and lack snow depth inversion research aiming at grasslands. The existing snow depth inversion model based on mountain areas or forests is difficult to be directly applied to grasslands, on one hand, grassland vegetation cover has a wind resistance and snow resistance effect, the characteristics of wind and snow flowing can be changed, snow is redistributed to a certain extent, on the other hand, the height and density of the grasslands can interfere with optical and microwave signals, and the model inversion precision is not ideal, so that the grassland snow depth inversion method based on multi-source remote sensing and machine learning is provided to solve the problems in the prior art. Disclosure of Invention Aiming at the problems, the invention aims to provide a grassland snow depth inversion method based on multi-source remote sensing and machine learning, which effectively characterizes the blocking and interception effects of a grassland under-pad on wind blowing snow by introducing the height of a grassland as a physical constraint parameter, adopts a characteristic optimization-integrated modeling strategy, enhances the adaptability of a model to shallow, medium and deep snow areas, and improves inversion precision and robustness. The invention aims at realizing the purpose by adopting the following technical scheme that the grassland snow depth inversion method based on multi-source remote sensing and machine learning comprises the following steps: Firstly, acquiring multi-source input data, preprocessing the multi-source input data, and extracting characteristic variables for snow depth inversion; Step two, carrying out correlation analysis and screening on the extracted characteristic variables based on the Pearson correlation coefficient, and reserving the characteristic obviously related to the snow depth target variable as an effective input characteristic; Thirdly, constructing a training sample set by using the screened effective input features and corresponding snow depth sample investigation data, constructing a machine learning integrated model for training, and establishing a nonlinear mapping relation from the multi-source features to the snow depth; Step four, utilizing a trained machine learning integrated model to carry out inversion calculation on the snow depth of the target grassland area, and utilizing an independent verification data set to carry out accuracy verification on an inversion result; Fifthly, based on the snow depth space distribution data obtained through inversion, combining a preset snow depth threshold value, and carrying out snow depth risk level division on a target grassland area; and step six, inputting weather forecast data of a future period into a trained machine learning integrated model, and predicting the snow depth of the future period of the target grassland area. The method is further improved in that the multi-source input data in the first step comprise satellite remote sensing image data, analysis meteorological data, topographic data, ground actual measurement grass height data and grassland snow depth pattern investigation data, wherein the satellite remote sensing image comprises Sentinel-2 or MODIS data, and the analysis meteorological data is ERA5 data. The method is further improved in that in the first step, the characteristic variables comprise normalized snow index, normalized vegetation index, accumulated precipitation in snow season, geographic position and topography parameters and grass height, wherein vectors of the geographic position and the topography parameters comprise longitude, latitude and elevation, and the grass height is obtained through ground actual measurement and is used for quantitativel