CN-121980905-A - Double-frequency scatterometer wind field joint inversion method based on deep learning

CN121980905ACN 121980905 ACN121980905 ACN 121980905ACN-121980905-A

Abstract

A double-frequency scatterometer wind field joint inversion method based on deep learning solves the problems of insufficient precision and poor adaptability of the existing inversion technology, and belongs to the technical field of space remote sensing and marine environment monitoring. The inversion method comprises the steps of simulating satellite-borne multi-angle measurement data of a dual-band scatterometer to obtain simulation data, constructing physical enhancement features based on a sea surface scattering mechanism of the scatterometer, establishing a dual-branch deep learning wind field inversion model which comprises a full wind speed model and a sub wind speed model, extracting a trunk network by sharing features, realizing wind speed and wind direction inversion respectively by two output branches, training the full wind speed model and the sub wind speed model based on the simulation data, and carrying out wind field inversion based on the trained full wind speed model and sub wind speed model.

Inventors

TIAN XIAOQING
ZHANG CHI
ZHANG QINGJUN
WANG YANG
YU SONGBAI
LU XIANG
LI RUIFENG

Assignees

中国空间技术研究院

Dates

Publication Date: 20260505
Application Date: 20251212

Claims (10)

1. A double-frequency scatterometer wind field joint inversion method based on deep learning is characterized by comprising the following steps: Simulating satellite-borne multi-angle measurement data of the dual-band scatterometer to obtain simulation data; establishing a double-branch deep learning wind field inversion model, wherein the double-branch deep learning wind field inversion model comprises a full wind speed model and a split wind speed model, a shared characteristic extraction backbone network and two output branches respectively realize wind speed and wind direction inversion; Training a full wind speed model and a partial wind speed model based on simulation data; and carrying out wind field inversion based on the trained full wind speed model and the trained partial wind speed model.
2. The method for joint inversion of the wind field of the double-frequency scatterometer according to claim 1 is characterized in that satellite-borne multi-angle measurement data of the double-frequency scatterometer are simulated, wherein Ku frequency band VV and VH polarized backscattering coefficient data are generated by adopting NSCAT-4DS models, C frequency band VV backscattering coefficients are generated by adopting CMOD7 models, C frequency band VH channel backscattering coefficients are generated by adopting H14S models, and simulated data correspond to sea surface scattering characteristics under different wind direction conditions respectively.
3. The dual frequency scatterometer wind farm joint inversion method of claim 1, wherein the physical features comprise: the sine and cosine characteristics of the relative azimuth angle reflect the wind direction modulation effect; wind speed second and third order polynomial characteristics: expressing a nonlinear wind field response; Polarization difference characteristics, namely characterizing polarization sensitivity of Bragg scattering; Ku and C channel differential features, enhancing surface roughness and frequency dependent expression.
4. The dual-frequency scatterometer wind farm joint inversion method of claim 1, wherein in the dual-branch deep learning wind farm inversion model: a shared feature layer, namely a two-layer ReLU fully connected network and a Dropout is introduced; The self-adaptive interaction layer is used for enhancing the feature fusion capability through the leachable weight; A wind speed regression branch for outputting a non-negative continuous wind speed predicted value; and a wind direction vector branch circuit outputs normalized wind direction cosine/sine vectors, so that the problem of angle periodicity is avoided.
5. The dual-frequency scatterometer wind farm joint inversion method of claim 1, wherein training the full wind speed model comprises: Based on simulation data, inputting a full wind speed model after standardized processing, dividing the full data into a training set and a test set according to 8:2, and dividing a verification set from the training set; Training is carried out on a training set containing multi-scanning angle Ku and C double-frequency back scattering observation data, a random mixed noise enhancement strategy is adopted in the training stage, and stability of the model in a complex sea surface environment is improved.
6. The dual frequency scatterometer wind farm joint inversion method of claim 1, wherein training the split wind speed model comprises: based on simulation data, dividing a wind speed value into a plurality of intervals according to a wind field observation statistical rule of a scatterometer, dividing data of each wind speed interval into a training set and a testing set according to 8:2, dividing a verification set from the training set, independently training a set of models, and carrying out customized optimization on response characteristic differences of scattering coefficients of different wind speed sections; Training is carried out on a training set containing multi-scanning angle Ku and C double-frequency back scattering observation data, a random mixed noise enhancement strategy is adopted in the training stage, and stability of the model in a complex sea surface environment is improved.
7. The dual frequency scatterometer wind farm joint inversion method of claim 1, wherein the wind farm inversion comprises: the physical enhancement feature extraction and standardization of the test set are completed; screening submodels meeting the threshold value of the minimum sample number according to the wind speed section; During inversion, firstly executing a full wind speed model to perform wind speed inversion, and then calling a corresponding wind speed dividing model to complete inversion; and finally outputting a real dimensional wind speed and wind direction angle result.
8. The dual-frequency scatterometer wind farm joint inversion method of claim 1, further comprising an inversion performance evaluation, in particular comprising: s51, performing error-free inversion; S52, performing a noise robustness test; s53, analyzing the performance and physical interpretability of the wind-splitting section.
9. The method of claim 8, wherein in S52, a noise robustness test is performed by adding random errors to the test set.
10. The method according to claim 8, wherein in S53, the backward scattering coefficient sensitivity to wind direction is calculated for the test sample according to the wind speed interval for analyzing the performance and physical interpretability of the wind speed interval.

Description

Double-frequency scatterometer wind field joint inversion method based on deep learning Technical Field The invention relates to a double-frequency scatterometer wind field joint inversion method based on deep learning, in particular to a deep learning sea surface wind speed and wind direction joint inversion method based on Ku and C double-frequency scatterometer observation data, and belongs to the technical field of space remote sensing and marine environment monitoring. Background The sea surface wind field is a core physical quantity for representing sea-air interaction, and has important application in aspects of sea dynamics research, wind energy resource evaluation, disaster monitoring and the like. The satellite-borne scatterometer is one of key loads of ocean wind field remote sensing detection and mainly works in microwave frequency bands (such as Ku and C bands). The basic principle is that microwave pulses are transmitted to the sea surface and back scattering echo signals are received, and the inversion of the sea surface wind field is realized by establishing a Geophysical Model Function (GMF) between a back scattering coefficient and sea surface wind vectors (wind speed and wind direction). However, the conventional method has the following disadvantages: 1. the decoupling error of wind speed and wind direction is large, and strong correlation exists between the wind speed and the wind direction, so that the prior background field is usually required to be used for constraint, and accumulated deviation is easy to introduce. 2. The single-frequency observation information is insufficient, the multi-scale sea surface rough features are difficult to comprehensively describe, and the inversion performance is reduced under the conditions of low wind speed and complex sea. 3. The method depends on background fields and iterative optimization, and has the advantages of complex processing flow, large calculation amount and possibility of misjudgment in sudden sea and air events. In summary, the existing traditional inversion method is limited in the aspects of multi-frequency band fusion, complex sea state processing and the like, influences wind field inversion accuracy, and urgent needs to develop a novel intelligent inversion technology, fuse multi-frequency information and improve the utilization capability of scatterometer data. Disclosure of Invention The invention aims to solve the technical problems of overcoming the defects of the prior art and solving the problems of insufficient precision and poor adaptability of the prior inversion technology. The invention aims at realizing the following technical scheme: a double-frequency scatterometer wind field joint inversion method based on deep learning comprises the following steps: Simulating satellite-borne multi-angle measurement data of the dual-band scatterometer to obtain simulation data; establishing a double-branch deep learning wind field inversion model, wherein the double-branch deep learning wind field inversion model comprises a full wind speed model and a split wind speed model, a shared characteristic extraction backbone network and two output branches respectively realize wind speed and wind direction inversion; Training a full wind speed model and a partial wind speed model based on simulation data; and carrying out wind field inversion based on the trained full wind speed model and the trained partial wind speed model. In one embodiment of the invention, satellite-borne multi-angle measurement data simulation is carried out on a dual-band scatterometer, wherein Ku band VV and VH polarized backscattering coefficient data are generated by adopting NSCAT-4DS model, C band VV backscattering coefficient is generated by adopting CMOD7 model, C band VH channel backscattering coefficient is generated by adopting H14S model, and simulation data correspond to sea surface scattering characteristics under different wind directions respectively. In one embodiment of the present invention, the physical features include: the sine and cosine characteristics of the relative azimuth angle reflect the wind direction modulation effect; wind speed second and third order polynomial characteristics: expressing a nonlinear wind field response; Polarization difference characteristics, namely characterizing polarization sensitivity of Bragg scattering; Ku and C channel differential features, enhancing surface roughness and frequency dependent expression. In an embodiment of the present invention, in the inversion model of the wind field for double-branch deep learning: a shared feature layer, namely a two-layer ReLU fully connected network and a Dropout is introduced; The self-adaptive interaction layer is used for enhancing the feature fusion capability through the leachable weight; A wind speed regression branch for outputting a non-negative continuous wind speed predicted value; and a wind direction vector branch circuit outputs normalized wind direc