Automatic change detection in remotely sensed hyperspectral imagery (Case study: wetlands and waterbodies)

Wetlands are one of the important types of ecosystems that play a fundamental role in the environment andprovide significant benefits due to the resources that they contain. Therefore, it is necessary to monitor thechanges in these ecosystems. The alterations in Earth’s ecosystems caused by the natural activities, such asdrought, as well as human activities and population growth has been affecting the wetlands and waterbodiesarea. Therefore, for achieving a better detection of these changes over time, it is important to generatedescriptive location maps based on the characteristics of wetlands. Hyperspectral images have shownpotential use in many applications due to their high spectral resolution, and consequently, their highinformative value. This study presents a hybrid procedure for automatic detection of changes in wetlandsusing a new approach which can provide more details about the changes with high accuracy. The hybridproposed method is based on incorporating chronochrome, Zscore analysis, Otsu algorithm, simplex viasplit augmented lagrangian (SISAL), Harsanyi–Farrand–Chang (HFC), Pearson correlation coefficient(PCC), and support vector machine (SVM) to detect changes using hyperspectral imagery. The proposedmethod in the first step, produce a training data for tuning SVM and kernel parameters. The second step,predicted change areas based on a chronochrome algorithm and binary change map obtained using SVMclassifier. The third step, the amplitude of changes is created by ZScore analysis and binary change mask.Finally, the multiple change map is produced based on the estimation of number and extraction ofendmembers and similarity measure. The proposed method evaluated and compared the performances withother common hyperspectral change detection methods using three realworld datasets of multitemporalhyperspectral imagery. The empirical results reveal the superiority of the proposed hybrid method inextracting the change map with an overall accuracy of nearly 96% and a kappa coefficient of 0.89 whileother hyperspectral change detection methods have the overall accuracy lower than 93% and kappacoefficient 0.80. In addition, this hybrid method can provide ‘multiple changes’ as well as the magnitude ofextracted changes.