Deep Learning and SAR Applications

Classifying sea ice depth from Sentinel-1 data using CNN

While using neural networks for SAR data classification is not new, the application of deep learning for land cover classification seems to have significantly increased since the introduction of fully convolutional neural networks in 2015. Many studies have explored the feasibility of classifying public land cover using CNNs, such as roads, buildings, floods, urban areas, and crops. Deep learning has also been used for some interesting atypical land cover (or water cover) applications, such as identifying oil spills and classifying sea ice of different thicknesses. Generally, the use of deep learning tends to outperform traditional methods, although it may not be more efficient in terms of time and computational costs. As with all supervised learning techniques, performance heavily depends on the quality of training sample data.

Urban change in UAVSAR data

High-resolution, high-frequency monitoring SAR data is particularly well-suited for change detection applications, as it can see through clouds and capture changes in amplitude and phase coherence. I was surprised to find that a significant amount of research has utilized deep neural networks for change detection in SAR data. Unlike the object recognition dominated by CNNs, various neural network methods have been used to address the problem of identifying surface changes in SAR data. Typically, these methods either rely on classifying land cover in multiple sets of time series and then comparing the classified results, or classifying the radiometric or phase differences between multiple sets of time data.

So far, there seems to be little research on using deep learning methods to analyze interferograms or support InSAR processing. However, there is evidence that early studies have used CNNs for phase unwrapping tasks and integrated deep learning methods into the InSAR processing chain. There seems to be broader room for applying deep learning methods to phase unwrapping.

Generating high-quality visible images from SAR images using GAN

Some studies have applied CNNs and Generative Adversarial Networks (GANs) to applications in SAR data augmentation and data fusion. This is a highly valuable research area that also benefits object recognition, where CNNs can be used to estimate and reduce speckle noise in SAR data. The result is a “clean” SAR image produced through a single feedforward process. GANs can also be used to improve the resolution of SAR data.Additionally, some literature provides examples of image conversion from Sentinel-1 resolution to TerraSAR-X resolution. While these methods seem to not retain feature structures well, they present interesting application points for super-resolution and style transfer in SAR. At the same time, GANs can be used to make high-resolution SAR data appear more like optical images through speckle reduction and coloring, which aids in the visual interpretation of SAR images. New datasets are now available to help advance this work and other applications that require SAR and optical data fusion.

When reviewing the literature on these applications, a common challenge is evident: the lack of high-quality SAR training data, especially at high resolutions. As with all supervised learning methods, the performance of models and results heavily depends on the input training data. The most commonly used high-quality training dataset is the MSTAR dataset, but it contains only a limited number of military features. Other researchers have begun to create their own annotated training datasets, but the cost of establishing one’s own dataset is very high due to the limited selection of high-resolution source data, and few have open licenses (e.g., UAVSAR). Some methods can augment training data through simulation, but reasonable datasets are still needed as a starting point.