The MODIS EVI data cube is analysed to characterise the time series trends for each pixel. This process represents the time series in a form that allows it to be classified more easily. The initial characterisation of the time series data involves three steps:
(1) Piece-wise Linear Segmentation
The time series segmentation algorithm approximates a time series using a series of piece-wise linear segments. Each segment is characteristic of a sequence of environmental behaviour that typifies the associated land cover class.
(2) Sub-sequence Discretization
Each segment is then detailed by three values describing its starting value, slope and length to form a vector input to the classification process.
(3) Sequence Labelling
The sequences are then clustered into groups of like behaviour, and each group then assigned a behaviour label. A time series is thus transformed from a large collection of varying values to a smaller set of behaviour types, joined as a kind of “DNA” string representing the pixel through time.
(4) Markov modelling
The DNA string provides an initial class which is then modelled as a dynamic Markov Chain to estimate the likelihood that a particular land cover class would occur after any other land cover class.
(5) Bayesian classification
The final class is then assigned using a Bayesian classifier.