Land Cover Classification in Mangrove Ecosystems Based on VHR Satellite Data and Machine Learning—An Upscaling Approach

Abstract

Mangrove forests grow in the inter-tidal areas along coastlines, rivers, and tidal lands. They are highly productive ecosystems and provide numerous ecological and economic goods and services for humans. In order to develop programs for applying guided conservation and enhancing ecosystem management, accurate and regularly updated maps on their distribution, extent, and species composition are needed. Recent advances in remote sensing techniques have made it possible to gather the required information about mangrove ecosystems. Since costs are a limiting factor in generating land cover maps, the latest remote sensing techniques are advantageous. In this study, we investigated the potential of combining Sentinel-2 and Worldview-2 data to classify eight land cover classes in a mangrove ecosystem in Iran with an area of 768 km2. The upscaling approach comprises (i) extraction of reflectance values from Worldview-2 images, (ii) segmentation based on spectral and spatial features, and (iii) wall-to-wall prediction of the land cover based on Sentinel-2 images. We used an upscaling approach to minimize the costs of commercial satellite images for collecting reference data and to focus on freely available satellite data for mapping land cover classes of mangrove ecosystems. The approach resulted in a 65.5% overall accuracy and a kappa coefficient of 0.63, and it produced the highest accuracies for deep water and closed mangrove canopy cover. Mapping accuracies improved with this approach, resulting in medium overall accuracy even though the user’s accuracy of some classes, such as tidal zone and shallow water, was low. Conservation and sustainable management in these ecosystems can be improved in the future.