*2.2. Image Analysis*

All Landsat and Sentinel images were acquired from Earth Explorer and georeferenced to UTM WGS 1984. Data collection included satellite images and ancillary data with local expert knowledge. Prior to the classification of the Landsat images, we stacked the total number of bands in each satellite image by producing a composite. The Mask tool was used to extract the study area and exclude urban coastal regions, the inland region, and the open ocean [56]. Expert knowledge informed the land class designation. Thematic maps produced from the supervised classification were used for change detection in the analyses of thematic change dynamics and the time-series (Figure 4). All image analyses were performed in ArcGIS Version 10.6.1.

## *2.3. Image Classification*

We used a supervised classification method and the ML clustering algorithm with composite images of all Landsat bands. Supervised classification has been the most frequent method by which the remotely sensed data of mangrove areas have been classified, and the ML algorithm has been found to be a robust technique that is capable of repeated refinement and reclassification [14]. With supervised training, it is important that the training area be a homogeneous sample of the respective

class but at the same time includes the range of variability for the class [57]. Therefore, more than one training area per class was used. An accurate classification depends on the extent of overlap between class signatures. The ML classifier minimizes the total error in the classification if the estimate of the underlying probability distribution is correct [58]. Based on Bayes' theorem, the ML algorithm uses a discrimination function to assign pixels to the class with the highest likelihood [59]. Images were classified by using spectral signatures that were obtained from training samples. Training sample polygons represent distinct sample areas of various land cover types to be classified. Distinguishable classes represented by the training samples were examined from the spectral band characteristics [60]. By using the statistical tools in ArcGIS, we determined the samples to be distinguishable by their histograms and distinct scatter-plots [59]. Between 7 and 140 training polygons were generated for each feature class. We classified images into nine information (land) classes: cropping/grazing, oceanic, sand beach, open forest, mangrove forest, estuarine wetland, saltpan, bare mudflat, and saltmarsh grass. Wetland land-use classes include emergen<sup>t</sup> vegetation, riparian vegetation, and riverine and palustrine wetlands (e.g., vegetated swamps) (Table 3).


**Table 3.** Land classes used in the classification analysis.
