Fusion of multitemporal contextual information by neural networks for multisensor remote sensing image classification

Farid Melgani, Sebastiano B. Serpico, Gianni Vernazza
2003 Integrated Computer-Aided Engineering  
The contextual analysis of a multitemporal sequence of images of a given site represents a way to improve the accuracy with respect to the non-contextual single-time classification. The proposed contextual multitemporal classification scheme consists of two stages of multilayer perceptron (MLP) neural networks for each single-time image of the multitemporal sequence. The first stage is a one-hidden layer MLP whose role is to estimate the single-time posterior probability of each class, given
more » ... feature vector. These probability estimates represent spectral information; in addition, they are utilized to generate a non-contextual classification map. The neighboring class labels of a given pixel in the non-contextual classification map are exploited to extract spatial information, while temporal information is deduced from the non-contextual maps produced by the remaining single-time images. Spatial and temporal contextual information, together with spectral information, serves as input for the second stage network where the fusion takes place. As the network configuration can influence the classification performances, three MLP-based configurations are investigated. Experimental results on a multitemporal data set consisting of two multisensor (Landsat TM and ERS-1 SAR) images are presented and the performances of the proposed method are compared with those of both a classifier based on Markov random fields and a statistical contextual classifier. * Corresponding author. spatial aspects in a classification scheme, in particular, have not been intensively dealt with in the remote sensing literature. The correlation characterizing a set of images acquired at different dates makes the temporal dimension worth to integrate in a classification scheme. An early work in that direction was presented in [21] . The authors generalized the Bayes optimal strategy to the case of multiple observations by assuming a classconditional independence of feature vectors derived from different temporal data sets and by allowing class changes over time. The approach proposed in [12] is based on a stochastic system representing land-cover types through a non-stationary Gaussian process, as input, and the temporal spectral behavior, as output. The probability of transition from a class at one date to a
doi:10.3233/ica-2003-10108 fatcat:wnxp3jv4njcnlom6ftyrptgsnm