Temporal Mixture Analysis of Hypertemporal Antarctic Sea Ice Data in the Sense of Machine Learning

Abstract

Hypertemporal data, or time series acquired at high temporal frequencies, are often used to determine seasonal characteristics of environmental phenomena such as sea ice concentration. However, it is difficult to analyze long-term hypertemporal remote sensing data over extensive areas without prior information. Most pixels of hypertemporal data are highly mixed and contain several distinct temporal signals that represent seasonal characteristics of substances. This study performed temporal mixture analysis, which is algebraically similar to spectral mixture analysis, but occurs in the time domain instead of the spectral domain. Temporal mixture analysis was used to investigate the temporal characteristics of Antarctic sea ice. Two general steps were used to address mixing problems: 1) finding temporally unique signatures of pure components, which are referred to as temporal endmembers, and 2) unmixing each pixel in the time series data as a linear combination of the endmember fractional abundances. Because endmember selection is critical to the success of both spectral and temporal mixture analysis, it is important to select proper endmembers from large quantities of hypertemporal data. A machine learning algorithm was introduced to successfully identify endmembers without prior knowledge. A fully linear mixing model was then implemented in an attempt to produce more robust and physically meaningful abundance estimates. Experiments that quantitatively and qualitatively evaluated the proposed approaches were conducted. A temporal mixture analysis of high-temporal-dimensional data provides a unique summary of long-term Antarctic sea ice and noise-whitened reconstruction images via inverse processing. Further, comparisons of regional sea-ice fractions provide a better understanding of the overall Antarctic sea ice changes.