A long-term time series of wetland inundation maps is essential for eco-hydrological analysis and modelling of these vital ecosystems. Although the Landsat archive offers extensive optical/infrared global datasets, mapping wetland inundation using surface water algorithms remains challenging due to the complex spectral responses from soil, algae, and vegetation cover. In this study, we developed a method to rapidly generate binary inundation maps using Google Earth Engine by transforming Landsat imagery into a "mirror image" collection composed of bands related to water and vegetation indices. The entire time series over the Macquarie Marshes, a Ramsar-listed wetland in New South Wales, Australia, was classified using a random forest algorithm trained on representative samples covering diverse inundation scenarios.
This method allows for the rapid generation of a continuous long-term inundation time series, in contrast to event-based approaches, such as change detection using pre- and post-flood imagery or thresholding of single-date images during flood events, which are limited to isolated events and require specific temporal conditions. The resulting maps showed strong agreement with existing inundation datasets and demonstrated improved performance in capturing inundation patterns compared to traditional surface water mapping techniques. Additionally, we produced inundation probability maps and compared them with vegetation-type maps, revealing a high level of consistency. The proposed method offers a robust, scalable solution for generating wetland inundation maps when tailored to local conditions in other wetland regions.