New approaches to increase the accuracy and applicability of change detection in multi and hyperspectral remote sensing imagery
Student: Simon Adar
Advisors: Prof. Yoel Shkolnisky and Prof. Eyal Ben Dor
Monitoring planet earth is one of the most important challenges of mankind. It enables decision makers to act against global warming causes, pollution sources, activities such as illegal mining and more.
Change detection and multitemporal analyses aim to detect changes occurring over a specific geographical area using two or more images acquired at two or more different times. Spaceborne missions in the last decades enabled ongoing affordable monitoring of the earth with various earth observation sensors. These have shown the ability to detect different types of changes such as vegetation stress, pH values, water quality, urban growth and more.
With the enormous amount of data that is collected everyday with spaceborne instruments there is a need to have automatic procedures to produce monitoring products such as change maps from one time to another. In practice, up to the present, reliable monitoring products require ground truth collected data. This limits the adoption of this technology for specific areas of interest, leaving most of the earth surface unmonitored with the highest spatial resolution, although the data is collected and archived but not really used. In addition, govermental space agencies around the world are planning more spaceborne missions within the next few years with improved spectral and spatial resolution and increased temporal coverage.
For this reason, advances in reliable unsupervised change detection techniques are needed. These techniques do not require ongoing measurements of ground truth data, and thus can be applied for all types of acquired and archived images.
In this thesis we present and explore new methods related to unsupervised change detection. Our contribution can be divided into two main subjects. The first being a new thresholding approach for unsupervised binary change detection thresholding. This approach focuses on determining the threshold that discriminates between change and no-change pixels. The difference between pixels in the two images can be associated with real changes or noise. We propose a thresholding scheme that separates the threshold into two parts: 1) a threshold that accounts for the errors related to sensor stability, atmospheric conditions and data-processing variations, and 2) a threshold associated with georectification errors. We demonstrate our method using both multispectral Landsat images, airborne imaging spectroscopy HyMap images and examine small spectral changes of different soil samples in laboratory conditions as well. The second being a new method of predicting spectral emissivity in the long-wave infrared (LWIR) spectral region using the visible (VIS) spectral region. The proposed method is suitable for two main scenarios of missing data - sensor malfunctions and narrow FOV. We demonstrate the usefulness and limitations of this prediction scheme using the airborne hyperspectral scanner (AHS) sensor, which consists of both VIS and LWIR spectral regions.
The Sokolov mining area located in the western part of the Czech Republic is used as a case study for the presented methods.