Comparing different classification algorithms for monitoring mangrove cover changes in southern Iran

Abstract

Mangrove forests in Iran are highly productive and complex ecosystems since they represent the interface between land and sea. They are a unique environment for supporting biodiversity, and they provide direct and indirect benefits to humans. Investigating changes in mangrove forests is essential for ecologists and forest managers to improve the assessment and conservation of natural ecosystems. The goals of the present study include: (I) to evaluate and compare four supervised classification algorithms based on Landsat time series imagery to detect mangrove cover in southern Iran, (II) to detect changes in mangrove cover between 1985, 1998, and 2017; and (III) to compare the four different predictions resulting from the applied classification algorithms. An accuracy assessment was conducted using k-fold cross-validation and independent validation, and differences between the classification techniques were analyzed. Although all four algorithms produced high overall accuracy (ranging from 81% to 93%) and Kappa values (from 0.81 to 0.92), visual comparisons of the predictions revealed that Random Forest (RF) performed best. The results of the change analysis showed that mangrove cover areas decreased by approximately 4% from 1985 to 1998 and then increased by approximately 8.9% from 1998 to 2017. A change detection map shows a decrease in mangrove cover in near coastal regions, such as the Tabl and Gavarzin areas, and an increase in mangrove cover at a distance from the Qeshm coastline that involves open spaces between the trees. Rising water levels and human development are important factors in the decline of mangroves. The findings of this research are useful for the management, restoration and conservation planning of mangrove forest in southern Iran. Keywords: Mangrove, Random forest classifier, Support vector machine, Discriminant analysis, Land cover change, Statistical analysis