Unsupervised multi-scale change detection from SAR imagery for monitoring natural and anthropogenic disasters

Abstract

Radar remote sensing can play a critical role in operational monitoring of natural and anthropogenic disasters. Despite its all-weather capabilities, and its high performance in mapping, and monitoring of change, the application of radar remote sensing in operational monitoring activities has been limited. This has largely been due to: (1) the historically high costs associated with obtaining radar data; (2) slow data processing, and delivery procedures; and (3) the limited temporal sampling that was provided by spaceborne radar-based satellites. Recent advances in the capabilities of spaceborne Synthetic Aperture Radar (SAR) sensors have developed an environment that now allows for SAR to make significant contributions to disaster monitoring. New SAR processing strategies that can take full advantage of these new sensor capabilities are currently being developed. Hence, with this PhD dissertation, I aim to: (i) investigate unsupervised change detection techniques that can reliably extract signatures from time series of SAR images, and provide the necessary flexibility for application to a variety of natural, and anthropogenic hazard situations; (ii) investigate effective methods to reduce the effects of speckle and other noise on change detection performance; (iii) automate change detection algorithms using probabilistic Bayesian inferencing; and (iv) ensure that the developed technology is applicable to current, and future SAR sensors to maximize temporal sampling of a hazardous event. This is achieved by developing new algorithms that rely on image amplitude information only, the sole image parameter that is available for every single SAR acquisition. The motivation and implementation of the change detection concept are described in detail in Chapter 3. In the same chapter, I demonstrated the technique's performance using synthetic data as well as a real-data application to map wildfire progression. I applied Radiometric Terrain Correction (RTC) to the data to increase the sampling frequency, while the developed multiscaledriven approach reliably identified changes embedded in largely stationary background scenes. With this technique, I was able to identify the extent of burn scars with high accuracy. I further applied the application of the change detection technology to oil spill mapping. The analysis highlights that the approach described in Chapter 3 can be applied to this drastically different change detection problem with only little modification. While the core of the change detection technique remained unchanged, I made modifications to the pre-processing step to enable change detection from scenes of continuously varying background. I introduced the Lipschitz regularity (LR) transformation as a technique to normalize the typically dynamic ocean surface, facilitating high performance oil spill detection independent of environmental conditions during image acquisition. For instance, I showed that LR processing reduces the sensitivity of change detection performance to variations in surface winds, which is a known limitation in oil spill detection from SAR. Finally, I applied the change detection technique to aufeis flood mapping along the Sagavanirktok River. Due to the complex nature of aufeis flooded areas, I substituted the resolution-preserving speckle filter used in Chapter 3 with curvelet filters. In addition to validating the performance of the change detection results, I also provide evidence of the wealth of information that can be extracted about aufeis flooding events once a time series of change detection information was extracted from SAR imagery. A summary of the developed change detection techniques is conducted and suggested future work is presented in Chapter 6.