A JOINT SPATIAL-TEMPORAL CLASSIFICATION AND FEATURE BOUNDARY 
UPDATING MODEL 
P. Caccetta 
* CSIRO Mathematical and Information Sciences, Private Bag 5, Wembley, Perth, Western Australia 6913, 
peter.caccetta@csiro.au 
KEY WORDS 
classification, hidden Markov model, segmentation. 
ABSTRACT 
Here we consider a hidden Markov model for jointly estimating class label images and region boundary positions using information from 
imaging sensors and region boundary starting estimates. The required final class label images and updated region boundaries are treated 
as latent variables, induced from a corresponding set of observational variables. This model is motivated by the desire to incorporate 
information captured and represented by boundary information from processes independent of a given classification task, and the wish to 
use the information both to improve classification accuracies and to update the boundary positions based on image observations. We use 
the incorporation of (possibly) incomplete forest boundary inventory data with landsat satellite observations as an example of model 
application. 
1. INTRODUCTION 
Classification and/or segmentation approaches are typically 
employed for estimating land use and cover type change 
information from digital imaging satellite and airborne sensors. 
The method used for performing the classification is generally 
called the classifier, while the recorded measurements are referred 
to as data. The state of the process recognised by the classifier is 
labelled as belonging to a particular class. After this stage, the 
labelled data are called the classification and the data are said to 
have been classified. Typically results will be prepared as maps or 
images of class labels. 
The process of classification requires that a) the number of 
possible classes are defined; and a choice of model is made for b) 
assessing the information in the available data; and c) deciding the 
class label after having assessed the information in the data. The 
accuracy of the resulting information is a key consideration in 
determining the suitability of an approach for a particular 
application. 
To improve classification accuracies, there has been a long 
tradition of augmenting a source of remotely sensed data with 
other data, as well as many alternate methods for analysis and 
classification proposed. For instance, a popular classifier is the 
maximum likelihood classifier (mlc) (Rao, 1966). Early examples 
of incorporating ancillary data (by using the data to specify class 
prior probabilities) into this classifier was provided by Strahler, 
1980. More generally, modem methods for combining multiple 
sources of data, possibly for the task of classification, are 
commonly referred to as data fusion methods. 
Here we concentrate on the problem of combining ancillary data, 
which exist as a set of closed boundaries, with mlc with the view 
to improving classification accuracies. Acknowledging that the 
ancillary data may be incomplete or incorrect, we wish to update 
the boundary positions to better reflect the class label positions 
observed from remotely sensed data while at the same time using 
the boundary information to influence the class labelling. We use 
as our motivation the desire to incorporate forest inventory 
boundary information with the classification of time series 
remotely sensed landsat TM data. In section 2.2 we specify a time 
series model that is compatible with that used for current 
Australian national mapping of forest presence/absence (Caccetta 
et al 2003, 2007), with some extra terms to incorporate the option 
for boundary updating. In section 3 we use a simple (contrived) 
example of classifying a single date of imagery for forest 
presence/absence while updating a boundary known to be wrong. 
We note that in its single date formulation, the model has 
conceptual similarities with that proposed by Wu and Albert, 
2007. In section 4 we experiment with the more difficult problem 
of using real inventory boundaries for improving a multi-class 
forest classification problem characterised by poor class spectral 
separation. 
2. MATERIALS AND METHODS 
2.1 Model 
Given a set Y = {Y b Y 2 , ... Y n } of n images representing n time 
steps , B = {B!,B 2 ,...,B n } boundary images each composed of Bj 
= {bi,b 2 ,...,b q } q region boundaries starting position (that is, each 
boundary image may be composed of multiple region 
boundaries), and training data sufficient to define L = 
{Li,L2,...,Ln} class label images, we wish to estimate the 'true’ 
class label images L’ = {L’ 1 ,L’ 2 ,...,L , n } and boundary positions 
B’ = {B’ 1 ,B’ 2 ,...,B’ n }. The boundary images are 2 class class- 
label images having labels “inside” and “outside” (the boundary). 
The model to be described is applied iteratively, successively 
updating the estimates for L’ and B’, and we will use the 
superscript ’ to identify those terms where information derived 
from the previous iteration is used.