 |
| Titel |
Monitoring the Landscape Freeze/Thaw State by Radar Remote Sensing: a Novel Sensor Fusion Approach Based on Hidden Markov Models |
| VerfasserIn |
Simon Zwieback, Annett Bartsch, Thomas Melzer, Wolfgang Wagner |
| Konferenz |
EGU General Assembly 2011
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 13 (2011) |
| Datensatznummer |
250049150
|
|
|
|
|
|
| Zusammenfassung |
| The freezing and thawing of the landscape is closely interconnected with bio- and
geophysical processes: e.g. the hydrologic cycle, trace gas exchange and the metabolic
activity of plants and microbes. Consequently, observations of the dynamics of the
Freeze/Thaw (T/W) state can be beneficial for various disciplines. Due to the scarcity of
in-situ monitoring sites, remote sensing methods have been applied to observe the F/T state at
large scales; in particular, microwave remote sensing has been shown to be an apt
tool.
The sensitivity of the radar backscatter Ïă0, as measured by scatterometers, to various
influence factors depends strongly on the frequency. We thus propose a novel sensor fusion
approach based on scatterometry observations in C band (by ASCAT on MetOp)
and Ku band (by SeaWinds on QuikScat). As the dynamics of the F/T state vary
considerably over different climatic zones and biomes and because of the irregular
temporal sampling, probabilistic time series models are a promising tool due to their
flexibility.
The model is an adaptation of the well-known Hidden Markov model (HMM) – the F/T
state is assumed to be a Markov chain, whose value is not directly observable. It is a discrete
variable, taking on the values ’frozen’, ’non-frozen’ and ’thawing’. The observables are
modelled as random variables, whose probability distributions depend on the current
state. The simple structure assures that inference can be done efficiently, e.g. the
calculation of the probability of the state on a given day by the Forward-Backward
algorithm. Our approach does not use training data; the parameters are estimated for
each time series in an unsupervised fashion. This is achieved by maximizing the
marginal likelihood in the framework of the Expectation Maximization algorithm.
The choice of the parameterization and initial values as well as the conditional
independence assumptions have a considerable impact on the results – it is where domain
knowledge based on backscatter models and empirical data analysis comes into
play.
We study the algorithm by testing it in the region of 120 - 130E and 50 - 75N. This
area in Russia and northern China covers a variety of biomes such as Tundra, Taiga and
Steppe. The time series of the probability of the state are compared with in-situ snow and
temperature data as well as global climate models. In general, the accuracy exceeds 90%, but
the algorithm can fail in agriculturally used land (fields, pastures) and bare rock outcrops in
mountainous regions. The benefits of using two different data sources become evident in
certain cases, where correct classification based on a single frequency would be almost
impossible.
The proposed algorithm yields satisfactory results in general, particularly in the Tundra
and Taiga. Overall, the study confirms the importance of exploiting two different data sources
in a systematic fashion. The key advantage of probabilistic models is their flexibility, so that
the approach is applicable for large scale applications. The combination of a powerful
mathematical framework with domain knowledge – which is, for example, used in the choice
of initial values and constraints of parameters – is very promising. This approach for sensor
fusion is very flexible: different sensors could easily be incorporated and there are many
applications with the aim of detecting a dynamic, discrete state, such as the mapping of burnt
or inundated areas. |
|
|
|
|
|
|