Compression of multitemporal remote sensing images through Bayesian segmentation M.Cagnazzo, G.Poggi, G.Scarpa, L.Verdoliva Dipartimento di Ingegneria Elettronica e della Telecomunicazioni Universit` a Federico II di Napoli, via Claudio, 21 – 80125 Napoli, Italy cagnazzo, poggi, giscarpa, verdoliv @unina.it Abstract— Multitemporal remote sensing images are useful tools for many applications in natural resource management. Compression of this kind of data is an issue of interest, yet, only a few paper address it specifically, while general-purpose compression algorithms are not well suited to the problem, as they do not exploit the strong correlation among images of a multitemporal set of data. Here we propose a coding architecture for multitemporal im- ages, which takes advantage of segmentation in order to compress data. Segmentation subdivides images into homogeneous regions, which can be efficiently and independently encoded. Moreover this architecture provides the user with a great flexibility in transmitting and retrieving only data of interest. I. I NTRODUCTION Multitemporal images allow one to follow the evolution in time of a given region of interest by means of change detection techniques [1-3], and therefore represent valuable tools for natural resource management. For many advanced applications the basic data-unit of interest becomes a set of multispec- tral or hyperspectral images acquired at different times. The transmission and archival of such a huge amount of data is a very demanding task, despite the constant improvement of storage and communication media, and some form of data compression is often desirable or necessary. General-purpose image compression techniques, such as JPEG2000, are not suitable in this case, as they neglect important information about the source, especially the strong functional dependence among the various spectral bands and temporal images. Despite the relevance of this problem, only a few papers address the compression of multitemporal images in the litera- ture [4,5], with approaches not really tailored to the task. More relevant to the problem is the literature on the compression of multispectral images, where techniques for the joint analysis and compression of images are currently under investigation [6,7]. Segmentation-based compression techniques, in particular, seems very promising for several reasons. First of all, seg- mentation can single out regions where significant changes have occurred; this information is embedded in the encoded stream and represents a very valuable information for many applications. Turning to compression efficiency, segmentation allows one to encode each region independently (object-based coding) and therefore to devote more encoding resources to regions of interest (e.g. those where changes have occurred), adapt encoding parameters to local features of the region (adaptive coding), or even select different encoding techniques for different regions (dynamic coding). Finally, for multispec- tral multitemporal images, the additional cost of encoding the segmentation map is shared by all bands, and hence it is typically negligible. Given all these considerations, we propose here a coding scheme for multitemporal images based on the following major steps: 1) the images collected at times , are jointly segmented; 2) based on the analysis of the segmentation maps, changed and unchanged regions are detected; 3) the segmentation maps are jointly encoded without loss of information; 4) the region textures are lossy coded independently, with rate allocation based on the results of change detection. Segmentation is carried out by means of a Bayesian technique, based on a TS-MRF (tree-structured Markov random field) image model, recently developed and assessed in [8,9]. Texture compression is carried out by means of shape-adaptive wavelet transform [10] followed by a shape-adaptive version of the SPIHT algorithm [11] to quantize transform coefficient. The paper has the following structure: in Section 2 we pro- vide more detail on the processing tools used for segmentation and coding, while in Section 3 we describe the experimental setting and discuss numerical results; finally Section 4 draws conclusions. II. SEGMENTATION AND CODING TOOLS Our compression algorithm comprises four major steps 1) supervised classification of the multitemporal image; 2) extraction of a simple change detection map; 3) lossless encoding of the map; 4) lossy coding of the image regions. that will be briefly described in the following. To make description more concrete and easy to understand, we refer already to the data that will be used in the exper- iments. We work on two Landsat TM images (only optical bands) of an agricultural area near the river Po in Italy taken in April and May 1994 and accurately registered. The images are 494x882 pixel but we take a smaller square region of 512x512 pixel (some white lines added at the margin) to speed up processing. Fig. 1 and Fig. 2 show band 3 of the selected area in April and in May.