Supervised Segmentation of Remote-Sensing Multitemporal Images Based on the Tree-Structured Markov Random Field Model Luca Cicala, Giovanni Poggi and Giuseppe Scarpa DIET, Universit` a Federico II di Napoli, via Claudio, 21 – 80125 Napoli, Italy email: {cicala,poggi,giscarpa}@unina.it Abstract— We deal with the supervised segmentation of multi- temporal remote-sensing images following a statistical Bayesian approach. To take into account prior information on the class of images, like the correlation between neighboring pixels, as well as the available knowledge about the structure of the current image, we model the image as a tree-structured Markov random field. The data collected at two different dates are jointly processed as a single multi-component image, with the classes defined a priori based on ground truth information and grouped in changed and unchanged macro-classes. Experimental results in terms of classification accuracy prove the effectiveness of the proposed technique with respect to to non-contextual methods, as well as to a disjoint approach. In addition, the classification tree allows for a direct interpretation of the result. I. I NTRODUCTION Most application based on remote-sensing imagery require some form of segmentation as an intermediate step for classi- fication and higher level processing. In the case of multitem- poral images, in particular, it is very important to segment the image in regions where changes have or have-not occurred, which is of interest for many land protection applications. Change detection is characterized by several peculiar factors that render ineffective some of the multitemporal image anal- ysis techniques typically used in other application domains. The main difficulties affecting change detection in remote- sensing images arise from: the lack of a priori information about the shapes of changed areas; the absence of a refer- ence background; differences in light conditions, atmospheric conditions, sensor calibration, and ground moisture at the two acquisition dates considered; problems of alignment of multitemporal images (registration noise) [1]. The task of change detection becomes more manageable in a supervised setting when some prior information about the classes of interest is available, but even in this case it remains extremely challenging. In order to exploit all prior knowledge about the images of interest, we resort here to a statistical Bayesian approach, and model the unknown image as a Markov Random Field [2]. With the MRF model, one assumes that each pixel depends statistically on the rest of the image only through a selected group of neighbors. This greatly simplifies the problem of assigning a meaningful prior of the image since only local characteristics need be specified. In particular, we will use the recently proposed tree- structured Markov random field (TS-MRF) model [3], [4] in which the image regions or classes are organized according to a binary structure. This hierarchical organization, in fact, presents a number of advantages • different models can be used to describe different spa- tial/spectral structures; • region parameters are estimated locally, based on the statistics of the regions of interest, leading to a simple adaptive model; • segmentation become much faster because only binary splits are carried out; • the tree structure itself represents a valuable information about the semantic of the image. In essence, the TS-MRF model tries to describe the hidden structure of the data by a nested set of binary MRFs, each corresponding to a node in a binary tree. This description, and not only the ensuing segmentation/classificatio map, is an important result of the modeling process. In the context of multitemporal image analysis, a care- ful choice of the segmentation tree structure, based on the available information, will provide us with an easy and sim- ple interpretation of the segmentation results, in terms of changed/unchanged regions, or any other grouping of interest. For example, although the two available images are jointly processed, the resulting map will easily provide, if correctly observed, both the marginal segmentations: a simple “pruning” corresponds to the ground classification at the first time, while a few “merging” give the classification at the second time. Although the most interesting feature of the proposed technique is the interpretability of the results, we will show that also in terms in terms of classification accuracy it com- pares favourably against well-known non-contextual reference techniques, namely, minimum distance (MD) and maximum likelihood (ML). Section II presents the TS-MRF model and his properties, while Section III deals with the experimental results and their discussed; finally, Section IV draws conclusions. II. THE STATISTICAL SEGMENTATION MODEL A random field X defined on a lattice S is said to be a MRF with respect to a given neighborhood system if the Markovian property holds for each site s. Quite often, in order to limit complexity, only the 4 or 8 closest pixels (system η 1 and η 2 , respectively) are included in a pixel’s neighborhood. 1569 0-7803-8742-2/04/$20.00 (C) 2004 IEEE