Characterizing Remote Sensing Data Compression Distortion for Improved Automated Exploitation Performance
Date of Award
Ph.D. in Engineering
Department of Engineering
Advisor: Eric Balster
All remote sensing systems and sensors contend with unwanted signals, broadly categorized as distortions, that contribute error to the ideally sensed image. Great effort is dedicated by system designers and end-users toward characterizing and mitigating distortions, which facilitates the ability to confidently exploit the sensed data. Distortions to the ideal image can be categorized into two classes: observed and constructed. Observed distortion is the class of distortions that contains all unwanted errors within an image due to uncontrolled processes that impede an ideal reconstruction of the scene such as, but not limited to, illumination changes, perspective changes, seasonal changes, atmospheric and meteorological effects, occlusions, blur, lens distortion, and various noise sources such as shot noise, thermal noise, and fixed-pattern noise. Because observed distortions occur before or during the digitization of the scene, only estimates of the undistorted ideal image can be made. In contrast to observed distortion, constructed distortion is a class of distortion that is intentionally introduced into the scene, possibly as a byproduct of other processing.If the data needs to be transmitted from a sensing platform to another location, or if the data needs to be more efficiently stored, compression is often used. If lossy algorithms are selected for compression, the data incurs distortion. However, since the undistorted image is known and the distortion is added to the image, compression distortion is categorized as constructed distortion. Despite being able to estimate the significance of compression distortion relative to the undistorted image, compression distortion is often uncharacterized and uncontrolled due to limited knowledge on its relationship to image exploitation algorithms. This gap in knowledge is the focus of this research.The presented work is developed with the intent of bridging the fields of lossy image compression with automated image exploitation, which are often developed under differing assumptions, motivated by real-world operational context. This work derives estimates of compression distortion for basic change detection measures and isolates the distortion to the individual images such that remote sensing systems do not need both images to estimate distortion's significance for change detection. Implications and limitations of these properties are investigated as they relate to advanced change detection algorithms. A system is developed to demonstrate how the receiver can leverage these distortion estimates for synthetic aperture radar (SAR) noncoherent change detection (NCCD), SAR coherent change detection (CCD), visible imagery change detection, and hyperspectral image (HSI) target detection.For SAR NCCD, the proposed system provides a 33% reduction in false alarms at a 0.1 probability of detection and 40:1 JPEG2000 compression relative to a traditional 40:1 JPEG2000 compression system. It is also observed to maintain near-distortionless false alarm rates across a wide range of compression ratios. At a 0.1 probability of detection for CCD and 15:1 JPEG2000 compression, the system provides a 37% reduction in false alarms. The CCD system is also demonstrated to maintain low false alarm rates across a range of compression ratios. Visible imagery is demonstrated to benefit from the developed metrics and system with 50% to 80% of compression-related false alarms suppressed in images with minimal observed distortion; however, images with significant illumination differences benefit marginally from the proposed system since the larger observed distortions overwhelm the constructed distortions. Lastly, an HSI compression system for monitoring on-target and background detection separability without knowing scene truth is demonstrated. This system is tested using ARCHER and SpecTIR data with JP3D compression and a derived signal-to-clutter ratio loss factor (SCRLF) metric to measure on-target and background separation. Results show correlated results between estimated SCRLF and sensed image SCRLF in both scene-based and physical model reflectance conversion.
Remote Sensing, SAR, compression, exploitation, remote sensing, change detection, signature detection
Copyright 2018, author
McGuinness, Christopher, "Characterizing Remote Sensing Data Compression Distortion for Improved Automated Exploitation Performance" (2018). Graduate Theses and Dissertations. 6649.