Unsupervised Change Detection Using Multi-Temporal SAR Data : A Case Study of Arctic Sea Ice

University essay from KTH/Geodesi och geoinformatik

Abstract:

The extent of Arctic sea ice has decreased over the years and the importance of sea ice monitoring is expected to increase. Remote sensing change detection compares images acquired over the same geographic area at different times in order to identify changes that might have occurred in the area of interest. Change detection methods have been developed for cryospheric topics. The Kittler-Illingworth thresholding algorithm has proven to be an effective change detection tool, but has not been used for sea ice. Here it is applied to Arctic sea ice data. The objective is to investigate the unsupervised detection of changes in Arctic sea ice using multi-temporal SAR images. The well-known Kittler-Illingworth algorithm is tested using two density function models, i.e., the generalized Gaussian and the log-normal model.

The difference image is obtained using the modified ratio operator. The histogram of the change image, which approximates its probability distribution, is considered to be a combination of two classes, i.e., the changed and unchanged classes. Histogram fitting techniques are used to estimate the unknown density functions and the prior probabilities. The optimum threshold is selected using a criterion function directly related to classification error.

In this thesis three datasets were used covering parts of the Beaufort Sea from the years 1992, 2002, 2007 and 2009. The SAR and ASAR C-band data came from satellites ERS and ENVISAT respectively. All three were interpreted visually. For all three datasets, the generalized Gaussian detected a lot of change, whereas the log-normal detected less.

Only one small subset of a dataset was validated against reference data. The log-normal distribution then obtained 0% false alarm rate through all trials. The generalized Gaussian obtained false alarm rates around 4% for most of the trials. The generalized Gaussian achieved detection accuracies around 95%, whereas the log-normal achieved detection accuracies around 70%. The overall accuracies for the generalized Gaussian were about 95% in most trials. The log-normal achieved overall accuracies at around 85%. The KHAT for the generalized Gaussian was in the range of 0.66-0.93. The KHAT for log-normal was in the range of 0.68-0.77. Using one additional speckle filter iteration increased the accuracy for the log-normal distribution. Generally, the detection of positive change has been accomplished with higher level of accuracy compared with negative change detection.

A visual inspection shows that the generalized Gaussian distribution probably over-estimates the change. The log-normal distribution consistently detects less change than the generalized Gaussian.

Lack of validation data made validation of the results difficult. The performed validation might not be reliable since the available validation data was only SAR imagery and differentiating change and no-change is difficult in the area. Further due to the lack of reference data it could not be decided, with certainty, which distribution performed the best. 

  AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)