A composed supervised/unsupervised approach to improve change detection from remote sensing L. Castellana * , A. D’Addabbo, G. Pasquariello Istituto di Studi sui Sistemi Intelligenti per l’Automazione I.S.S.I.A. – C.N.R., Via Amendola, 166/5, 70126 Bari, Italy Received 22 April 2005; received in revised form 11 July 2006 Available online 5 October 2006 Communicated by H. Sako Abstract In this paper a new approach to performing change detection analyses based on a combination of supervised and unsupervised tech- niques is presented. Two remotely sensed, independently classified images are compared. The change estimation is performed according to the Post Classification Comparison (PCC) method if the posterior probability values are sufficiently high; otherwise a land cover tran- sition matrix, automatically obtained from data, is used. The proposed technique is compared with the traditional PCC approach. It is shown that the new approach correctly detects the ‘‘true change’’ without overestimating the ‘‘false’’ one, while PCC points out ‘‘true change’’ pixels together with a large number of ‘‘false changes’’. Ó 2006 Elsevier B.V. All rights reserved. Keywords: Neural networks; Change detection; Remote sensing 1. Introduction Land cover change detection, carried out by using remote sensed images, is a challenging problem having a key function in many practical application areas such as deforestation assessment, damage assessment, disaster monitoring and urban expansion (Hame et al., 1998; Gopal and Woodcock, 1996; Ridd and Liu, 1998). Generally speaking, change detection involves the anal- ysis of two co-registered multi-spectral images acquired in the same geographical area at two different times; it can be performed both by an unsupervised and a supervised approach (Singh, 1989). The former is based on direct com- parison of the reflectance values of input images and can be performed by using different techniques such as Change Vector Analysis (CVA), Image Rationing and Vegetation Index Differencing (Fung, 1990). In this case, the most important problem is to distinguish between differences due to change in land cover classes and differences due to variations in solar illumination, atmospheric conditions or phenological development. In order to distinguish signi- ficant differences from unimportant ones (Radke et al., 2005), the difference image provided has to be suitably ana- lyzed using techniques such as significance and hypothesis tests or predictive models. In remote sensing, the analysis of the difference image is classically based on histogram thresholding techniques according to empirical strategies or trial-and-error procedures (Townshend and Justice, 1995). To overcome this limitation, an automatic thres- holding technique, based on the Expectation-Maximization (EM) algorithm (Dempster et al., 1977), has been proposed in literature (Bruzzone and Prieto, 2000, 2002a). Although the unsupervised change detection approach does not require any a priori information on the analyzed scene, in which case its application is favoured in many practical applications, it presents some critical limitations. Differences due to atmospheric and light variations at 0167-8655/$ - see front matter Ó 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2006.08.010 * Corresponding author. Fax: +39 080 5929460. E-mail addresses: castellana@ba.issia.cnr.it (L.Castellana), daddabbo@ ba.issia.cnr.it (A. D’Addabbo), pasquariello@ba.issia.cnr.it (G. Pasqua- riello). www.elsevier.com/locate/patrec Pattern Recognition Letters 28 (2007) 405–413