Search

CN-122017833-A - Synthetic aperture radar interferometry elevation inversion method based on convolution network

CN122017833ACN 122017833 ACN122017833 ACN 122017833ACN-122017833-A

Abstract

The invention discloses an elevation inversion method, in particular to an elevation inversion method for synthetic aperture radar interference measurement based on a convolution network, which comprises the following steps that 1, an interference phase diagram and an intensity diagram are received as input data; and 2, processing input data by using a convolutional neural network model of an improved U-Net architecture, training the model, and 3, inputting data to be tested into the trained model to obtain an inference result. According to the invention, the interference phase diagram and the SAR intensity diagram are utilized, and the contribution of different modes in different areas is dynamically weighted through the self-adaptive feature fusion module, so that the model depends on an intensity structure in a smooth area, focuses on phase change in a fringe dense area, and obviously enhances the adaptability to complex terrain.

Inventors

  • WANG HAIBO
  • LIANG XU
  • HONG XIN

Assignees

  • 上海银帆信息科技有限公司

Dates

Publication Date
20260512
Application Date
20260129

Claims (7)

  1. 1. The synthetic aperture radar interferometry elevation inversion method based on the convolution network is characterized by comprising the following steps of: Step 1, receiving an interference phase diagram and an intensity diagram as input data; Step 2, processing input data by using a convolutional neural network model of an improved U-Net architecture, and training the model; And step3, inputting the data to be tested into the trained model to obtain an inference result.
  2. 2. The synthetic aperture radar interferometry elevation inversion method based on the convolution network of claim 1, wherein in the step 1, an interference phase diagram represents a phase difference after two SAR images are interfered and comprises ground surface elevation information, and the intensity diagram represents an amplitude square of an SAR echo signal.
  3. 3. The synthetic aperture radar interferometry elevation inversion method based on the convolutional network of claim 1, wherein step 1 further comprises normalizing the interference phase map and the intensity map to obtain a normalized image.
  4. 4. The synthetic aperture radar interferometry elevation inversion method based on the convolutional network of claim 1, wherein in step 2, the model comprises an encoder, a bottleneck layer and a decoder, and the contribution weights of the phase characteristics and the intensity characteristics are dynamically adjusted through the adaptive characteristic fusion module, so that effective fusion of the phase and the intensity information is realized.
  5. 5. The synthetic aperture radar interferometry elevation inversion method based on the convolutional network of claim 1, wherein in step 2, the convolutional neural network model of the improved U-Net architecture outputs a single-channel digital elevation model, and the training process is optimized through a multi-objective composite loss function to improve elevation precision, terrain structure fidelity, physical rationality and noise robustness.
  6. 6. The synthetic aperture radar interferometry elevation inversion method of claim 5, wherein said multi-objective composite loss function comprises at least one of, but is not limited to, L 1 absolute error loss, gradient consistency loss, structural similarity loss, frequency domain consistency loss, physical constraint loss, and uncertainty perception loss.
  7. 7. The synthetic aperture radar interferometry elevation inversion method according to claim 1, wherein in said step 3, inverse normalization processing is performed on said output digital elevation model in an inference stage, including but not limited to median filtering, slope constraint smoothing and void filling.

Description

Synthetic aperture radar interferometry elevation inversion method based on convolution network Technical Field The invention relates to the technical field of new generation information, in particular to a synthetic aperture radar interferometry elevation inversion method based on a convolution network. Background An interferometric synthetic aperture radar (InSAR, interferometric Synthetic Aperture Radar) technology is used as an advanced remote sensing means to generate three-dimensional elevation information of the earth surface by using the difference in phase difference between SAR images acquired from two or more different perspectives. However, accurate extraction of digital elevation models (DEMs, digital Elevation Model) from raw interferometric phase maps has been challenged by atmospheric conditions, noise disturbances, and complex topographical features. Traditional InSAR elevation inversion methods rely mainly on interference phase unwrapping and physical model-based elevation mapping algorithms. Although these methods can provide accurate elevation information under ideal conditions, they are often limited in practical application by firstly, the 2 pi ambiguity problem (i.e. "wrapping") in the interferometric phase data, which needs to be solved by complex phase unwrapping algorithms, which tend to introduce errors, especially in the presence of high noise or discontinuous terrain, and secondly, conventional methods often require fine corrections for atmospheric delays, track errors, etc., which not only increase the complexity of the process flow, but also limit its degree of automation and processing efficiency, and finally, conventional algorithms have difficulty in ensuring consistent high precision and high resolution output in the face of large-scale, diverse terrain. In recent years, with the development of deep learning technology, especially the successful application of Convolutional Neural Network (CNN) in the field of image recognition and processing, a new idea is provided for solving the above challenges. The CNN has strong feature extraction capability, can automatically learn the nonlinear relation between input data and target variables, and does not need to explicitly define a complex physical model. For InSAR elevation inversion, CNN is directly mapped to DEM from interference phase diagram and intensity diagram, so that not only can problems possibly occurring in the phase unwrapping process be effectively avoided, but also the processing speed and robustness can be improved, and better adaptability and generalization capability are particularly shown in a high-noise environment. Disclosure of Invention The invention aims to solve the problems in the prior art, and provides a synthetic aperture radar interferometry elevation inversion method based on a convolution network, so as to solve the existing technical problems. The technical scheme adopted for solving the technical problems is as follows: the invention provides a synthetic aperture radar interferometry elevation inversion method based on a convolution network, which comprises the following steps: Step 1, receiving an interference phase diagram and an intensity diagram as input data; Step 2, processing input data by using a convolutional neural network model of an improved U-Net architecture, and training the model; And step3, inputting the data to be tested into the trained model to obtain an inference result. Preferably, in the step 1, the interference phase map represents a phase difference after the two SAR images interfere and includes ground surface elevation information, and the intensity map represents an amplitude square of the SAR echo signal. Preferably, the step 1 further includes performing normalization processing on the interference phase map and the intensity map to obtain a normalized image. Preferably, in the step 2, the model includes an encoder, a bottleneck layer and a decoder, and the adaptive feature fusion module dynamically adjusts the contribution weights of the phase feature and the intensity feature to realize effective fusion of the phase and the intensity information. Preferably, in the step 2, the convolutional neural network model of the improved U-Net architecture outputs a single-channel digital elevation model, and optimizes the training process through a multi-objective composite loss function, so as to improve elevation precision, terrain structure fidelity, physical rationality and noise robustness. Preferably, the multi-objective composite loss function includes, but is not limited to, at least one of an L 1 absolute error loss, a gradient consistency loss, a structural similarity loss, a frequency domain consistency loss, a physical constraint loss, and an uncertainty-aware loss. Preferably, in the step 3, inverse normalization processing is performed on the output digital elevation model in an inference stage, including but not limited to median filtering, slope constraint s