Abstract
Deep learning is an increasingly popular solution for object detection and object classification in satellite-based remote sensing images. Unfortunately, its widespread use is encumbered by its need for vast amounts of training data. Transfer learning is one solution to the problem of scarce training data, in which some or all of the features learned for solving one problem are used to solve a different problem. Existing literature suggests that transfer learning is feasible and produces good results for classification in optical remote sensing images; however, the same cannot be said for synthetic aperture radar (SAR). The primary challenge for SAR relates to the SAR's imaging mechanism, itself. The magnitude and direction of the backscattered signal depends on the electric permittivity of the imaged scene, its surface roughness, the nuances of any physical structures, as well as the parameters of the SAR, including its wavelength and polarization. Consequently, variations in the configuration of any objects between a series of imaged scenes, relative to the SAR sensor, can produce very different signals between SAR images. Further, while deep learning has concurrently emerged as a powerful tool to accomplish the task of object detection, most research in the satellite remote sensing community has focused on its application to optical imagery. A second possible solution to the problem of scarce training data is domain adaptation, in which the goal is to find a domain-invariant feature space, rather than applying the same feature space to the two different asks, as in transfer learning.
The are three main goals of this dissertation. The first is to investigate the transferability of features learned from other imaging modalities for object classification in SAR images. We investigate the efficacy of transfer learning from ImageNet and simulated SAR images. In particular, we investigate the efficacy of transfer learning from the ImageNet domain, as well as a domain based on computer-aided design (CAD)-synthetic SAR images. Our goal is to enhance the understanding of how transfer learning may or may not be beneficial for classification tasks on SAR data, as well as what, if any, caveats there might be to its application. Our results indicate that there are strong dependencies on CNN architecture generalization capability as well as the incidence angles of training samples, and that transfer learning from ImageNet generally outperforms transfer learning from a CAD-synthetic domain.
Our second goal is to develop a deep learning-based single-stage model for object detection in satellite-based SAR imagery, with a specific focus on the influence of image incidence angle, satellite acquisition parameters, and backbone CNN architecture on detection performance. Specifically, we ask whether or not it is possible to leverage pre-trained CNN architectures with this object detection model, and how performance varies by incidence angle of the acquired images. With this model, we achieve good results with an inference time on large images of roughly 0.3 seconds; however, we show there is a distinct dependence of object detector performance on image acquisition parameters such as incidence angle and orbit direction. Our results also strongly suggest that there is a definite, but poorly understood, shortcoming to using classical CNNs to detect and classify objects in SAR images.
Finally, we present an application of domain adaptation for satellite-based SAR image classification using CAD-synthetic SAR data. We use an existing state-of-the-art CNN architecture, modified specifically for our task and applied modified versions of state-of-the-art domain adaptation algorithms. Ultimately, we demonstrate that domain adaptation indeed improves the performance on our SAR classification model, and our results support the hypothesis that domain adaptation is a promising technique for SAR DL applications. We demonstrate that domain adaptation indeed yields better performance on our SAR data set, as compared to transfer learning, and that domain adaptation is a promising technique for SAR object classification in situations where large quantities of real SAR data are not available.