Deep Image Prior Amplitude SAR Image Anonymization

An extensive evaluation of the Deep Image Prior (DIP) technique for image inpainting on Synthetic Aperture Radar (SAR) images.

Figure 1. Example of Vertical-Vertical (VV) amplitude SAR image anonymization with the proposed method. The urban agglomerate highlighted in red in the picture has been removed and substituted with content similar to the surrounding area. (a) target image to anonymize; (b) final anonymized image; (c) target image close-up; (d) final image close-up.

Synthetic Aperture Radar (SAR) images are a powerful tool for studying the Earth’s surface. They are radar signals generated by an imaging system mounted on a platform such as an aircraft or satellite. As the platform moves, the system emits sequentially high-power electromagnetic waves through its antenna. The waves are then reflected by the Earth’s surface, re-captured by the antenna, and finally processed to create detailed images of the terrain below.

SAR images are employed in a wide variety of applications. Indeed, as the waves hit different objects, their phase and amplitude are modified according to the objects’ characteristics (e.g., permittivity, roughness, geometry, etc.). The collected signal provides highly detailed information about the shape and elevation of the Earth’s surface. SAR images are also used for monitoring natural disasters, such as earthquakes, floods, and landslides, as well as for detecting changes in land-use patterns, such as urbanization and deforestation. Due to their nature of providing detailed imagery regardless of daylight and weather conditions, SAR images are also a precious asset in military applications. As a matter of fact, SAR images can be used to detect sensible military targets like aircraft, airports, ships, tanks or other vehicles.

In recent years, SAR images have become more widely available than ever before. Many portals now offer SAR images for free, making them accessible to researchers, students, and the general public alike. This has led to a surge in research using SAR imagery and increased deployment of this technology in commercial products and services.

The research community is concerned about the potential misuse of SAR images, which play a critical role in many applications, particularly military ones. The widespread availability of SAR images means that almost anyone can extract sensitive information from them, such as the location of troops or civilians during a conflict. As a result, some providers have begun to conceal features and information in their geospatial services, such as hiding traffic from maps that could reveal the location of refugees.

All these elements make it clear that techniques that anonymize or conceal target areas from remote sensing raster data are necessary. With this goal in mind, in this paper, we tackle the problem of removing sensitive objects from amplitude SAR images while preserving most of their content from a semantic point of view, i.e., most of their land-cover content.

We accomplish this task by relying on Deep Learning (DL) techniques for analyzing and processing SAR images. As a matter of fact, in recent years the SAR research community has gained a particular interest in DL tools like Convolutional Neural Networks (CNNs) due to their ability to learn complex features directly from data without manual processes involved. Recently, DL has been exploited for SAR image de-speckling, land-cover classification, oil spills detection, change detection of land use patterns, and so forth.

Also, the generation of SAR images using CNNs has found some compelling applications. Guo et al. relied on Generative Adversarial Networks (GANs) for generating fully synthetic SAR images of military vehicles starting from simple observation parameters like platform azimuth and target depression angles. Baier et al. generated both SAR and RGB satellite images by providing as inputs land-cover maps and digital elevation models to a conditional GAN. Moreover, there have been several other contributions exploiting RGB data to synthesize SAR images and vice versa, as well as methods exploiting SAR images to improve the quality of Electro-Optical (EO) images.

All these DL-based solutions typically need a training phase for the CNN to “learn” how to properly process the data for the task at hand. Training involves the collection of a corpus of data to optimize the parameters of the networks. This data collection must contain a considerable amount of samples to be representative of the data distribution in the real world, and to allow for the creation of training, validation, and test splits to avoid overfitting. Depending on the dimensions of the datasets, the complexity of the architecture, and the training process involved, the optimization of these methods can prove extremely computationally expensive and may require considerable manual effort.

This work was performed by Edoardo Cannas and Sara Mandelli for the Air Force Research Laboratory. For more information, download the Technical Support Package (free white paper) below. ADT-09234

This Brief includes a Technical Support Package (TSP).
Document cover
Deep Image Prior Amplitude SAR Image Anonymization

(reference ADT-09234) is currently available for download from the TSP library.

Don't have an account? Sign up here.