*2.5. Change Detection*

Change detection procedures estimate that a change in the reflectance of the study area results from a corresponding change in surface cover or surface material [62]. We characterised changes by using a suite of analytics. The first two methods of change detection were post-classification change analysis, and visual interpretation of images and comparison with Google Earth images from similar dates. Using local expert knowledge, we applied the on-screen digitisation of Landsat 2017 true-colour images to show areas of mangroves, saltpan/saltmarsh grass and estuarine wetland that have been altered from 2004 to the oceanic information class in 2017. The third method of change detection was the use of thematic change dynamics by using a remote sensing software tool to portray the dynamics of land cover change that occurred at Rocky Dam Creek/Cape Palmerston National Park. This tool measures the transition dynamics of a land cover class to another class at a given extent. Firstly, an initial state image was input as a time 1 image (2004), and then a final state image was input as time 2 (2017). Only changed areas were taken to visualize the overall dynamics. The output was a transition probability matrix signifying the "from–to" change that exemplified the past and present state of di fferent land cover classes. The fourth and final method of change detection was a raster-based trend analysis where a time-series stack of thematic maps that was constructed into a 3D array and indexed via row and column was built to ge<sup>t</sup> a time-series vector. The occurrence of each class in each pixel across the time-series (i.e., the maximum spatial extent for each class) was used to create the output map. For example, the maximum extent map for open forest would show pixels that contained at least one occurrence of that class across the time-series array. We fitted a linear regression model to the data array with a slope value related to areal cover change per year in the time-series. The trend analysis measured the net change between pixels through the time series data, integrating six raster datasets with the same spatial extent, the output being a spatial map of slope and trend analysis. Information classes were combined to give a broad statistical appraisal of the region's LULC change dynamics. The map showing slope of the regression line is displayed as ranked data that are representations of the data's spatial attributes; this was created with the Jenks optimization method, a data clustering method that determines the best arrangemen<sup>t</sup> of values into di fferent classes. The method seeks to reduce the variance within classes and maximize the variance between classes. If the values increment in time, they have a positive slope (red area) and, in the case of a decreasing regression line, a negative slope (green area). Temporal relationships were evaluated among the years by using the Pearson's correlation coe fficient r value. Values represent the direction and magnitude of land cover change through the time-series. Post classification change analysis, visual interpretation, and thematic change were quantified by using Erdas Imagine Version 16.5.1 and the Spatial Analyst Toolset in ArcGIS Version 10.6.1. ArcGIS and Python routines were used for the time series analysis.
