**1. Introduction**

Mangrove forests are considered one of the most important ecosystems on the earth. They occur in the inter-tidal zones along coasts in most tropical and semi-tropical areas [1,2]. Despite the large ecological benefits of mangrove forests, such as carbon sequestration, protection of land from erosion, purification of coastal water quality, and maintenance of ecological balance and biodiversity, mangroves have been destroyed worldwide as a result of climate change and human activities [3–6].

Qeshm Island, located o ff the southern coast of Iran in the Persian Gulf, is dominated by the cosmopolitan mangrove species *Avicennia marina*. Many studies have focused on the ecological and physiological characteristics of *A. marina* [7,8]. *Avicennia* species grow in oxygen-poor sediments that cannot supply the underground roots with su fficient oxygen. Consequently, their root system also includes vertically growing aerial roots (pneumatophores). These aerial roots also anchor the plants

during the frequent inundation with seawater in the soft substrate of tidal systems, and they play a significant role in sustaining mangroves [9]. Sea-level rise, a main consequence of climate change, will have a significant influence on future growing conditions [10]. Recent estimates of the extent of mangrove forests indicate that their total area has already decreased substantially, by 50% during the last half-century [11–13].

Identification of the aerial root system at a high spatial resolution would enable e fficient planning of reforestation in mangrove ecosystems, but this detailed information is currently missing. Image resolution is directly correlated with the ability to identify objects of the same type [14]. Despite the grea<sup>t</sup> value of Landsat images for numerous applications, the specifications are inappropriate for distinguishing mudflats with aerial roots from mudflats without aerial roots. This is also due to the spectral similarities of these classes and the influence of the soil in the tidal zone (dry and wet conditions). A more detailed mapping of the mangrove ecosystem, e.g., trees and aerial root systems, is required to improve assessments of their status and recommend appropriate protection measures.

In the last years, a range of low- to high-resolution aerial images [15–17], hyperspectral images [18], Synthetic Aperture Radar (SAR) data [19], and Light Detection and Ranging (LiDAR) data [20] has been used to map the extent and distribution of mangrove cover classes. In the past decade, data have become available from Very High-Resolution (VHR) satellites, such as Worldview-2 and Pléiades-1, leading to improved mapping of mangrove cover classes [21,22]. However, the main limiting factor is the high cost of data acquisition. Consequently, alternatives have been investigated, in particular combining satellite data of di fferent spatial resolutions [23]. Only recently, studies focusing on the use of freely available VHR data have been completed [24–28]. For example, in the forestry sector, a combination of commercially available Worldview-2 (WV-2) images and Landsat time-series data has been used to map tree species [29]. Di fferent classification techniques, such as traditional statistical regression [30], machine learning [31], artificial neural networks [32], and tree-based methods [33,34], have successfully been used with a large geographic extent and high level of detail.

Machine learning techniques such as Random Forest (RF), artificial neural networks (ANN) and Support Vector Machine (SVM) have gained exceptional attention to classify Land cover/Land use and identify mangrove forests because they perform better than traditional techniques [33,34]. These techniques use algorithms to learn the relationship between a response and its predictors and have been categorized into two sub-types: supervised and unsupervised techniques, respectively [35]. A main advantage is that they are all nonparametric classification techniques that require no assumptions about the distribution of the data and thus no prior knowledge about the characteristics of feature data is needed either [31]. Many studies in the field of Land cover/Land use classification have been carried out using di fferent machine learning algorithms as well as comparing them among each other [35]. In the last decade, RF has recently become preferred for mapping land cover classes in several realms [36,37]. RF is a nonparametric technique based on a set of decision trees. Unlike parametric techniques, RF can be used to predict land cover classes even based on a small sample size and therefore reduces both cost and time [38]. Moreover, embedded feature selection in the model generation process makes it possible to obtain high mapping accuracy. Several studies have demonstrated that RF, in combination with satellite data (Landsat) [37] and a high spatial resolution [16], can be used to successfully map mangrove cover classes. Moreover, the latest advances in remote sensing data and techniques, i.e., increasing availability of datasets in combination with higher temporal, spatial and spectral resolutions (e.g., ESA Copernicus Program Sentinel-1/-2), enable improved characterization of mangrove ecosystems. They make it possible to derive leaf area index, height and biomass, map the mangrove forest extent, and monitor mangrove status over time [39]. Several studies have been carried out to explore satellite data of di fferent spatial resolutions for improving land cover maps, i.e., in forestry that have combined data sets from Landsat and AVHRR [40] or Landsat and MODIS [41]. However, to the best of our knowledge, no study exist that combine Worldview-2 and Sentinel-2 images to classify mangrove ecosystems in greater detail which is a prerequisite for managing this ecosystem. Therefore, freely available Sentinel-2 data, in combination with commercially available

high-spatial-resolution imagery, has grea<sup>t</sup> potential for mapping wall-to-wall mangrove cover at a high level of detail, i.e., distinguishing between land cover classes with similar spectral properties.

In the present study, we investigated whether the combination of Sentinel-2 and Worldview-2 imagery can be used to accurately map the most relevant land cover classes for mangrove ecosystem management. We developed a three-step approach: (i) extraction of reflectance values from high-resolution Worldview-2 imagery, (ii) segmentation based on spectral and spatial features, and (iii) wall-to-wall mapping of the eight land cover classes based on Sentinel-2 imagery.

The study aims at developing a cost-e ffective, accurate method that can be applied widely and in a standardized manner, particularly when field surveys are restricted.

#### **2. Materials and Methods**

In order to produce a wall-to-wall map of mangrove cover classes for Qeshm Island, a two-step method was applied: (i) Reference data were generated at a 0.5-m spatial resolution using an object-based method performed on Worldview-2 images. The Worldview-2 data were dispersed across the entire study area and covered 27% of the total land cover. (ii) Reference data based on Worldview-2 images were used for the upscaling.
