Application of deep learning in InSAR - Zhihui Yuan, Ph.D
InSAR is a combination of Interferometric Synthetic Aperture Radar (InSAR) and SAR imaging technology, which can accurately measure three-dimensional information and deformation information of the earth's surface. This technology not only has the advantages of all-day, all-weather SAR, but also can use the phase information generated by interference to obtain accurate three-dimensional information and deformation information of the surface. Therefore, it is widely used in topographic mapping and surface deformation monitoring, such as DEM acquisition, volcano, earthquake, landslide, glacier movement and urban land subsidence.
In recent years, Deep Learning (DL) has become a hot research direction in the academic world. It belongs to a new research direction in the field of Machine Learning (ML) and was rated as one of the future 10 Breakthrough technologies in 2013. The concept of DL originates from the research of artificial neural network. It adopts multi-layer structure network to form more abstract high-level features by combining low-level features, so as to discover distributed feature representation of data and effectively extract spatial and temporal information of data. DL acquires some key implicit information by learning the inherent laws and presentation hierarchies of sample data, which is of great help in interpreting data such as words, images, and sounds. The ultimate goal is for machines to be able to learn analytically, like humans, and to recognize data such as text, images and sound. DL has achieved far more in speech and image recognition than previous related technologies. In recent years, with the rapid development of remote sensing big data and computing resources, DL has gradually been widely applied in the field of InSAR.
The principle of InSAR is to first measure the interference phase between signals received by two antennas at different locations, then unwrap the resulting interferogram and convert the absolute phase into height to extract terrain information. However, the actual interferogram often has a large number of singularities, which are caused by interference distortion and noise in radar measurement. These singularities often cause unwrapping error, which leads to the low quality of DEM. To solve this problem, Ichikawa and Hirose applied complex valued neural networks (CV-NN) to recover singulas in the spectral domain, learning the ideal relationship between the adjacent pixel and the central pixel spectrum through a hidden layer OF CV-NN with the help of a complex Markov random field filter. It is worth noting that the center pixel of each training sample is assumed to be an ideal pixel, which indicates that no singular points are input into the network during the training process. Similarly, Oyama and Hirose used CV-NN to recover singularities in the spectral domain. Costante et al. proposed a full CNN encod-decoder architecture for DEM estimation from single-navigation interferometric SAR images. They demonstrate that the model can extract high-level features from the input radar images using an encoder and then reconstruct full-resolution DEM through the decoder. In addition, this network can effectively solve the layover phenomenon in single-view SAR images with context characteristics.
In addition to DEM reconstruction, DL can also be used to detect and predict surface deformation. Schwegmann et al. proposed an interferogram settlement deformation detection technique based on CNN. They used a nine-layer neural network to extract salient information from interferograms and displacement maps to distinguish deformation targets from deformation-like targets. In addition, Anantrasirichai et al. used pre-trained CNN to automatically detect volcanic surface deformation from InSAR images. They split each image into patches, relabeled them with binary labels -- "background" and "volcano" -- and fed them back to the network to predict the deformation of the volcano. The method was later refined to detect slow-moving volcanoes using interferograms of time series. In another study related to automatic detection of volcanic deformation, Valade et al. designed and trained a CNN to learn the de-correlation mask from scratch from the input winding interferogram, which was then used to detect volcanic surface deformation. Later, by constructing LSTM deep recurrent neural network, deep learning was introduced into the field of surface subsidence prediction, and high-precision and time-sensitive subsidence prediction was realized.
Another based on the motivation of geophysical InSAR data depth study of the case study is about the earthquake information, it's actually earlier than previously proposed based on the research of CNN, the researchers used a simple feedforward shallow layer neural network to seismic events, and the use of neural network in solving nonlinear problems in the ability to automatically inversion of source parameters. More recently, deep learning has also been used for tomographic processing. Some researches have proposed an expanded deep network with vector approximate message passing algorithm, and experiments have been carried out using simulation and real data, which shows the spectral estimation gain and achieves relatively competitive performance. In addition, real-valued deep neural network is applied to multi-input and multi-output SAR 3D imaging. Compared with other methods based on compressed sensing, it shows better superresolution.
Another very important application of DL in InSAR is Phase Unwrapping (PU), among which the research of Yu Hanwen and Zhou Lifan is the most representative. They propose three different approaches for applying DL to InSAR in 2020 and 2021, respectively. The first method is to apply deep convolutional neural network (DCNNs) - Phase gradient neural network (PGNet) to phase gradient estimation. The second method is to apply DL to the placement of branch tangents, which transforms the residual balance problem into semantic segmentation problem, namely the branch cutting network (BCNet) problem. The third method applies DL to Multibaseline phase unwrapping (MBPU) and constructs a cluster analysis neural network (CANet), as shown in FIG. 1.
Figure 1 CANet architecture.
In general, the application of deep learning methods in InSAR is still at a very early stage. Although the InSAR technology into the deep learning in different applications, but in addition to Yu Han Vivian and Hirose pioneering work, interference figure have not been give full play to their full potential, many applications are interference figure and deformation which are obtained by interference figure figure as similar to the RGB and gray image, so the complexity of the interference pattern has not been noticed. In addition, the lack of ground truth data for SAR spot removal problems, detection and image restoration problems related to deep learning provides impetus for the development of semi-supervised and unsupervised algorithms combining deep learning with InSAR. This direction will definitely be a research hotspot in the future.