CN-122021362-A - Soil moisture inversion model training method, soil moisture inversion method and device

CN122021362ACN 122021362 ACN122021362 ACN 122021362ACN-122021362-A

Abstract

The invention provides a soil moisture inversion model training method, a soil moisture inversion method and a soil moisture inversion device, which belong to the technical field of remote sensing information processing, wherein the method comprises the steps of determining a target correction factor by using a consistency analysis result of a synthetic aperture radar remote sensing image of a series of satellites, and correcting a target inversion physical model to obtain a corrected inversion physical model; and simulating and generating a theoretical simulation data sample set by using a correction inversion physical model, performing offset correction on the theoretical backscattering coefficient by using an actual measurement backscattering coefficient determined based on the synthetic aperture radar remote sensing image, and finally constructing a correction simulation training sample set by using the corrected backscattering coefficient and each physical variable parameter value in the theoretical simulation data sample so as to pretrain the soil moisture inversion model. According to the invention, a semi-empirical deep learning fusion model combining a multi-star collaborative SAR inversion system and physical pilot and data learning is constructed, so that the accuracy and stability of soil moisture inversion are remarkably improved.

Inventors

WANG CHUNMEI
YANG JIAN
CHENG TIANHAI
ZHAN YULIN
QIU XINXIN
ZHANG LILI

Assignees

中国科学院空天信息创新研究院

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (13)

1. The soil moisture inversion model training method is characterized by comprising the following steps of: Performing consistency analysis on the synthetic aperture radar remote sensing images of the region to be inverted, which are acquired by the series satellites in the same sampling period, and determining a target correction factor and a target inversion physical model according to a consistency analysis result; Correcting the target inversion physical model based on the target correction factor to obtain a corrected inversion physical model; Simulating to generate a theoretical simulation data sample set comprising a plurality of theoretical simulation data samples according to reasonable value ranges of physical variable parameters in the corrected inversion physical model, wherein each theoretical simulation data sample comprises a theoretical backscattering coefficient corresponding to each physical variable parameter value; Performing offset correction on the theoretical backscatter coefficient by using an actual measurement backscatter coefficient to obtain a corrected backscatter coefficient, wherein the actual measurement backscatter coefficient is determined based on the synthetic aperture radar remote sensing image; Determining a corrected simulated training sample set comprising a plurality of corrected simulated training samples based on the corrected backscatter coefficients and the physical variable parameter values; and pre-training the soil moisture inversion model by using the correction simulation training sample set.
2. The method of training a soil moisture inversion model according to claim 1, wherein said performing offset correction on said theoretical backscatter coefficients using measured backscatter coefficients to obtain corrected backscatter coefficients comprises: And fitting the actually measured backscattering coefficient and the theoretical backscattering coefficient based on a preset deviation correction model to obtain the corrected backscattering coefficient.
3. The method of claim 1, wherein determining a corrected simulated training sample set comprising a plurality of corrected simulated training samples based on the corrected backscatter coefficients and the physical variable parameter values comprises: determining a soil volume moisture content tag value based on the soil volume moisture content value in the physical variable parameter values; Determining a simulated training vector sample of the corrected simulated training samples based on the corrected backscatter coefficient and other physical variable parameter values of the physical variable parameter values other than the soil volume moisture content value; the corrected simulated training samples are determined based on the simulated training vector samples and the soil volume moisture content tag value.
4. The method for training a soil moisture inversion model according to claim 1, wherein the performing a consistency analysis on the synthetic aperture radar remote sensing images of the region to be inverted acquired by the series of satellites in the same sampling period comprises: Performing fine registration on the synthetic aperture radar remote sensing image by using the space reference optical image of the region to be inverted to obtain a space co-located remote sensing image; calculating the gradient and the slope direction of each pixel in the space co-located remote sensing image by using a high-resolution digital elevation model, and generating an original gradient grid image; Determining an effective gradient grid map according to pixels of an earth surface bare area or a sparse vegetation area in the original gradient grid map; constructing a multi-star pixel pairing database based on the effective gradient raster pattern; Consistency analysis is carried out on the multi-star pixel pairing database to obtain a scattered point contrast graph and a gradient layering error graph; and determining the consistency analysis result based on the scattered point comparison graph and the gradient layering error graph.
5. The method for training a soil moisture inversion model according to claim 1, wherein the region to be inverted belongs to a hilly region, the target correction factors comprise gradient, slope direction and satellite azimuth angle, and the target inversion physical model is an Oh semi-empirical model; the correcting the target inversion physical model based on the target correction factor to obtain a corrected inversion physical model comprises the following steps: Determining a local angle of incidence using the satellite azimuth, the slope, and the slope direction; and replacing radar incidence angles in the Oh semi-empirical model by the local incidence angles to obtain the modified inversion physical model.
6. The method of claim 5, wherein the physical variable parameters include surface roughness, soil volume moisture content, radar wavelength, and the local angle of incidence; simulating to generate a theoretical simulation data sample set comprising a plurality of theoretical simulation data samples according to the reasonable value range of each physical variable parameter in the modified inversion physical model, wherein the theoretical simulation data sample set comprises: determining the combination of the values of the physical variable parameters based on the reasonable value range of the physical variable parameters; Substituting the physical variable parameter value combinations into the correction inversion physical model for each physical variable parameter value combination to obtain the theoretical backscattering coefficient corresponding to the physical variable parameter value combination, and determining the theoretical simulation data sample based on the physical variable parameter value combinations and the theoretical backscattering coefficients corresponding to the physical variable parameter value combinations.
7. The method of training a soil moisture inversion model of claim 1, further comprising, after said pre-training the soil moisture inversion model using said set of corrected simulated training samples: And (3) performing fine adjustment on the pre-trained soil moisture inversion model by using the actually measured training sample set to obtain a trained soil moisture inversion model.
8. A method of soil moisture inversion comprising: Constructing an input feature vector based on the back scattering remote sensing image of the region to be inverted and the corresponding physical variable parameters; Inputting the input feature vector into a soil moisture inversion model to obtain the predicted soil volume water content output by the soil moisture inversion model; The method comprises the steps of acquiring a series of satellite images, determining a back scattering remote sensing image based on a synthetic aperture radar remote sensing image of the region to be inverted acquired by a series of satellites in the same sampling period, and training the soil moisture inversion model based on the soil moisture inversion model training method according to any one of claims 1-7.
9. Soil moisture inversion model trainer, characterized in that includes: The consistency analysis module is used for carrying out consistency analysis on the synthetic aperture radar remote sensing images of the region to be inverted, which are acquired by the series satellites in the same sampling period, and determining a target correction factor and a target inversion physical model according to a consistency analysis result; the physical model correction module is used for correcting the target inversion physical model based on the target correction factor to obtain a corrected inversion physical model; The simulation data acquisition module is used for simulating and generating a theoretical simulation data sample set comprising a plurality of theoretical simulation data samples according to the reasonable value range of each physical variable parameter in the correction inversion physical model, wherein each theoretical simulation data sample comprises each physical variable parameter value and a theoretical backscattering coefficient corresponding to each physical variable parameter value; The deviation correction module is used for carrying out deviation correction on the theoretical backscattering coefficient by utilizing the actually measured backscattering coefficient to obtain a corrected backscattering coefficient, wherein the actually measured backscattering coefficient is determined based on the synthetic aperture radar remote sensing image; a training sample acquisition module for determining a corrected simulated training sample set comprising a plurality of corrected simulated training samples based on the corrected backscatter coefficients and the physical variable parameter values; and the pre-training module is used for pre-training the soil moisture inversion model by utilizing the correction simulation training sample set.
10. A soil moisture inversion apparatus, comprising: the vector construction module is used for constructing an input feature vector based on the back scattering remote sensing image of the region to be inverted and the corresponding physical variable parameters; the model processing module is used for inputting the input feature vector into a soil moisture inversion model to obtain the predicted soil volume water content output by the soil moisture inversion model; The backward scattering remote sensing image is determined based on the synthetic aperture radar remote sensing image of the region to be inverted, which is acquired by a series of satellites in the same sampling period; the soil moisture inversion model is trained based on the soil moisture inversion model training device according to claim 9.
11. An electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor when executing the computer program implements the soil moisture inversion model training method of any one of claims 1 to 7 or the soil moisture inversion method of claim 8.
12. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the soil moisture inversion model training method of any one of claims 1 to 7 or the soil moisture inversion method of claim 8.
13. A computer program product comprising a computer program, characterized in that the computer program when executed by a processor implements the soil moisture inversion model training method of any one of claims 1 to 7 or the soil moisture inversion method of claim 8.

Description

Soil moisture inversion model training method, soil moisture inversion method and device Technical Field The invention relates to the technical field of remote sensing information processing, in particular to a soil moisture inversion model training method, a soil moisture inversion method and a soil moisture inversion device. Background With the rapid development of satellite remote sensing technology, synthetic aperture radar (SYNTHETIC APERTURE RADAR, SAR) has become an important means for monitoring earth surface soil moisture due to its all-weather and all-day observation capability. In SAR-based soil moisture inversion studies, the dominant methods include semi-empirical methods based on physical scattering models and data-driven methods based on machine learning. The data driving method based on machine learning is based on a Random Forest (RF), a support vector machine (Support Vector Machine, SVM), a deep neural network (Deep Neural Network, DNN) and other machine learning methods, and the nonlinear relation between the backscattering coefficient, the vegetation index, the topography parameters and the like and the soil moisture is automatically learned and modeled from a large number of observation samples. However, for a semi-empirical method based on a physical scattering model, the semi-empirical physical model generally assumes that the earth surface is a horizontal or approximately flat surface, and systematic deviation is easy to be caused due to influence of terrain factors such as gradient, slope direction and the like on radar incidence geometry, and meanwhile, for a data driving method based on machine learning, the model has poor generalization capability in a scene outside a sample due to lack of physical constraint because the model is usually trained completely by sample data, and particularly in regions with complex terrain and large change of observation conditions, the problem of unstable model output result and weak interpretability exists. Disclosure of Invention The invention provides a soil moisture inversion model training method, a soil moisture inversion method and a soil moisture inversion device, which are used for solving the defects that a semi-empirical physical model is easy to cause systematic deviation due to influence of factors such as non-considered terrain and the like, and a data driving method based on machine learning is unstable in output and poor in generalization capability in a terrain complex region due to lack of physical constraint in the prior art. The invention provides a soil moisture inversion model training method, which comprises the following steps: Performing consistency analysis on the synthetic aperture radar remote sensing images of the region to be inverted, which are acquired by the series satellites in the same sampling period, and determining a target correction factor and a target inversion physical model according to a consistency analysis result; Correcting the target inversion physical model based on the target correction factor to obtain a corrected inversion physical model; Simulating to generate a theoretical simulation data sample set comprising a plurality of theoretical simulation data samples according to reasonable value ranges of physical variable parameters in the corrected inversion physical model, wherein each theoretical simulation data sample comprises a theoretical backscattering coefficient corresponding to each physical variable parameter value; Performing offset correction on the theoretical backscatter coefficient by using an actual measurement backscatter coefficient to obtain a corrected backscatter coefficient, wherein the actual measurement backscatter coefficient is determined based on the synthetic aperture radar remote sensing image; Determining a corrected simulated training sample set comprising a plurality of corrected simulated training samples based on the corrected backscatter coefficients and the physical variable parameter values; and pre-training the soil moisture inversion model by using the correction simulation training sample set. According to the training method of the soil moisture inversion model, the actual measurement backscattering coefficient is utilized to carry out offset correction on the theoretical backscattering coefficient to obtain the corrected backscattering coefficient, and the method comprises the steps of fitting the actual measurement backscattering coefficient and the theoretical backscattering coefficient based on a preset offset correction model to obtain the corrected backscattering coefficient. The method for training the soil moisture inversion model comprises the steps of determining a correction simulation training sample set comprising a plurality of correction simulation training samples based on the correction backscattering coefficient and each physical variable parameter value, determining a soil volume moisture content label value based on a soil volume moistur