**4. Discussion**

#### *4.1. General Comments*

Mangrove forests typically grow in zones that are marshy and inaccessible [11]. Therefore, collecting GPS points as training data through field surveys is di fficult [14]. Nowadays, new developments in remote sensing techniques have grea<sup>t</sup> potential to overcome the problem of acquiring field data in inaccessible areas of mangrove ecosystems [58]. Between 1970 and 2018, approximately 435 studies mapping the area of mangroves were conducted, and after the year 2000 the majority used Landsat images [14]. While Landsat imagery has the advantages of free availability, a large archive and extensive coverage, its relatively coarse spatial resolution of 30 m can be a major limitation. The potential of di fferent datasets from Landsat, ALOS AVNIR-2, Worldview-2 and LIDAR to map a detailed land cover of mangrove ecosystems was recently evaluated [59]. The results clearly demonstrated the importance of a higher spatial resolution for mapping specific mangrove features, such as individual tree crowns and species communities.

With the present study, we contribute to this research with an e fficient mapping of mangrove features using multi-resolution datasets. We add to existing knowledge gained in a previous study [37], which focused on comparing four classification algorithms based on Landsat images for predicting six land cover classes in the mangrove ecosystem: mangrove forest, mud flat, other land cover, tidal zone, water and settlement. The results of this earlier research demonstrated that using Landsat data enables to potentially distinguish between di fferent mangrove forest stands and can be useful for detecting their changes over time. However, since mangrove forests usually consist of small patches, Landsat images are not suitable for extracting more details and are mainly only appropriate for detecting changes in mangrove forest canopies. This is in accordance with [14,59], in that only high-resolution images can be used to map more detailed land cover classes. By increasing the number of spectral bands and the spatial resolution, it is possible to discriminate between small objects and to detect small objects, such as individual trees and mudflats with aerial roots. Several studies have shown the potential of Worldview-2 data for detailed land cover mapping, including mangrove forest ecosystems [16,39,59–61]. However, the main reason for the limited use of such imagery is its high cost–in particular, for developing countries. Thus, in our study, an upscaling approach was applied that reduces costs while still enabling the generation of a more detailed map of land cover classes.

## *4.2. Modelling Approach*

In the last decade, several studies have been carried out combining satellite data of di fferent spatial resolutions to improve land cover maps in the forestry sector. Some investigations have considered the combination of Landsat data with datasets of higher spatial resolutions such as IKONOS [62], GeoEye-1 data [63] or Worldview-2 [29].

Comparison of the two confusion matrices clearly demonstrated that the accuracy and kappa of the upscale approach were lower than the accuracy and kappa of the map that had the same extent as the one based on the Worldview-2 imagery. The use of a large amount of reference data to predict the subset of Sentinel-2 data helped to reduce misclassification.

The confusion matrix of the upscaling approach (Table 4, Figure 8) indicates that the overall accuracy and kappa decreased with increasing map scale. There was a high incidence of misclassification of individual trees and tidal zone when Sentinel-2 data were used. Several possible reasons for this error exist. First, it might be due to the amount of reference data because the Worldview-2 data only cover about 27% of the Sentinel-2 image. On the other hand, in the Worldview-2 image, the area of these two classes is less than that of the other classes. It is well known that the number of reference samples from the Worldview-2 image a ffects classification accuracy. In a recent study, it was demonstrated that the large amount of reference data obtained from the Worldview-2 image was the main driving factor for the accuracy of the classification of two pine tree species by Landsat data [29]. Future work could include the collection of more training samples in order to further improve the distinction of these land cover classes. Second, the error could be a result of the similarity of the spectral profile of individual trees and open canopy mangrove forest. The use of fewer reference samples decreases the spectral separability of classes and potentially decreases the accuracy. Third, the decrease in accuracy could be related to the level pre-processing and viewing geometry of Sentinel-2 imagery.

Nevertheless, the present study demonstrates that areas with di fferent canopy densities and mudflat areas (occurrence of aerial root systems) can be accurately classified using the upscaling approach with Sentinel-2 images. Overall, high accuracies were obtained for mapping closed canopy mangrove (75% user's accuracy, 94% producer's accuracy) and aerial roots (72%, 66%). Moreover, the combined use of Worldview-2 and Sentinel-2 images further increases map accuracies–in particular when the overall accuracy is not very high, and the user's accuracy is low in problematic classes.

#### *4.3. Importantance of Mapping of Detailed Information on Mangrove Forests*

Detailed maps of mangrove ecosystems are a prerequisite for successful protection and management. Since mangroves occur in areas with a high salt concentration in the soil, they have developed aerial roots for physiological functions and cover a large area within the Hara Protected Area [64]. This specialized root system reduces the power of sea waves and guarantees sustainable establishment of mangrove communities, as well as providing a protected place for aquatic animals [42]. In order to plan the development of mangrove forests, both naturally or artificially, the selection of potential suitable land is relevant. The land areas on the map that show the mangrove forests and mudflat with aerial roots are preferred to other areas that are not covered by vegetation. Moreover, the occurrence of mangrove is an indication that the land provides optimal conditions for the development of mangrove forests in terms of soil parameters such as salinity and pH. Mapping the details of mangrove ecosystems is an e ffective way to visualize, evaluate and better understand mangrove ecosystem development. Changes over a long period, as well as the recognition of unexpected changes due to natural or dramatic anthropogenic impacts, can be assessed at an early stage [65,66]. Moreover, assessing changes in the aerial root area can indicate the status of these forests because these roots are destroyed by an increase in water level or sediments.
