Markov Chain Monte Carlo Sampling on Polar Sea Ice Classification
The Remund-Long (RL) Multi-Sensor Sea Ice Classification algorithm� combines both radiometer and scatterometer data using Principle Component Analysis and reduces the dimensionality and noise level of the data. The algorithm uses an iterative Maximum a Posteriori (MAP) method based on a multi-varian...
Main Author: | |
---|---|
Format: | Text |
Language: | unknown |
Published: |
DigitalCommons@USU
2004
|
Subjects: | |
Online Access: | https://digitalcommons.usu.edu/spacegrant/2004/2004/37 https://digitalcommons.usu.edu/cgi/viewcontent.cgi?article=1228&context=spacegrant |
Summary: | The Remund-Long (RL) Multi-Sensor Sea Ice Classification algorithm� combines both radiometer and scatterometer data using Principle Component Analysis and reduces the dimensionality and noise level of the data. The algorithm uses an iterative Maximum a Posteriori (MAP) method based on a multi-variant Gaussian model with a temporal prior. As a result, the algorithm successfully classifies Winter Antarctic region into five different ice types. However, due to the nature of this pixel wise classification algorithm, the final classification is more likely to be corrupted by slat-pepper-shaped artifacts. Such artifacts are introduced by the Scatterometer Image Reconstruction (SIR) algorithm which utilizes multi-swath raw scatterometer data to generate high resolution images. In order to resolve such problem in RL algorithm, posterior distribution function with spatial prior is embedded into the classification process. A Markov Chain Monte Carlo (MCMC) sampling method is one way to sample such posterior distribution of the state space in which each element of the space has the size of an entire image. This report gives a brief introduction to the concept of Metropolis-Hastings Markov Chain Monte Carlo (MH MCMC) algorithm, discusses its implementation on polar sea ice classification, and compares the result with the RL algorithm. |
---|