Gradual land cover change detection based on multitemporal fraction images Daniel C. Zanotta a,b,n , Victor Haertel c a National Institute for Space Research, Av. dos Astronautas, 1758, S ~ ao Jose´ dos Campos, SP, Brazil b Federal Institute for Education, Science and Technology at Rio Grande do Sul, Rua Eng. Alfredo Huch, 475, Rio Grande, RS, Brazil c Federal University at Rio Grande do Sul, Av. Bento Gonc - alves, 9500, Center for Remote Sensing, Porto Alegre, RS, Brazil article info Article history: Received 27 May 2011 Received in revised form 1 November 2011 Accepted 3 February 2012 Available online 14 February 2012 Keywords: Remote sensing Change detection Linear mixture model Spatial context abstract This study proposes a new approach to change detection in remote sensing multi-temporal image data. Rather than allocating pixels to one of two disjoint classes (change, no-change) which is the approach most commonly found in the literature, we propose in this study to define change in terms of degrees of membership to the class change. The methodology aims to model images depicting the natural environment more realistically, taking into account that changes tend to occur in a continuum rather than being sharply distinguished. To this end, a sub-pixel approach is implemented to help detect degrees of change in every pixel. Three experiments employing the proposed approach using synthetic and real image data are reported and their results discussed. & 2012 Elsevier Ltd. All rights reserved. 1. Introduction Change detection in sets of image data covering the same scene but acquired at different times is of major interest in areas that include remote sensing, medical diagnosis, urban planning, and video surveillance. In remote sensing, change detection techniques based on multitemporal multispectral image data have been extensively applied to monitor agricultural fields, forests, and urban areas, among many other applications. Broadly speaking, the different approaches to change detection can be grouped into two categories: supervised and unsupervised meth- ods [1–6]. In the supervised approach, the change map is basically obtained by comparing classified images produced from multi- spectral image data acquired at two different dates. The drawback related to this application is the necessity of ground truth data for both dates whereas unsupervised approach is based on the analysis of the multispectral image data itself, requiring no additional information. The supervised approach has some advan- tages over the unsupervised approach, such as the understanding of the changes in terms of the actual land-cover classes. In addition, it does not require image radiometric normalization to take account of different conditions at the two acquisition dates [7,8]. The multitemporal ground truth data required by the supervised approach gives rise to a problem however, since its acquisition is usually expensive and time consuming, rendering the supervised approach impractical in many real-world applications. The unsupervised approaches to change detection are gener- ally based on difference images. These are produced by subtract- ing, pixel by pixel, images acquired at two different times [9]. The differences can be computed either from the original features or from features extracted from the original data, such as principal components or vegetation indices. In either case, an image of differences is produced and a threshold is applied to distinguish pixels where change has occurred from pixels that remain unchanged. Many approaches proposed in the literature attempt to model the distributions for changed and unchanged classes in order to estimate an adequate value for a threshold separating both classes [2,10,11]. The change detection process can be performed using a single spectral band, which captures the changes of interest in the scene, or using a set of p spectral bands. One widely used technique to analyze the differences when more than one spectral band is used is Change Vector Analysis (CVA) [1,12,13]. In the CVA method applied to multispectral image data, differences are calculated at every spectral band and the total change is represented by a vector X in the p-dimensional space [12]: X ¼ X p j ¼ 1 ðY 2, j Y 1, j Þ, ð1Þ where Y 1,j and Y 2,j are the feature vectors in spectral band j associated with corresponding pixels in two images acquired at Contents lists available at SciVerse ScienceDirect journal homepage: www.elsevier.com/locate/pr Pattern Recognition 0031-3203/$ - see front matter & 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.patcog.2012.02.004 n Corresponding author at: Federal Institute for Education, Science and Technology at Rio Grande do Sul, Rua Eng. Alfredo Huch, 475, Rio Grande, RS, Brazil. Tel./fax: þ55 53 32018718. E-mail addresses: daniel.zanotta@ifrs.riogrande.edu.br (D.C. Zanotta), victor.haertel@ufrgs.br (V. Haertel). Pattern Recognition 45 (2012) 2927–2937