Mapping Wetland Types in Semiarid Floodplains: A Statistical Learning Approach

Abstract

Detailed vegetation maps are needed for wetland conservation and restoration as different vegetation communities have distinct water requirements. It is a continuous challenge to map the distribution of different wetland types on a regional scale, and a trade-off between the categorical details and availability of resources to ensure broad applications is often necessary for operational mapping. Here, we evaluated the capacity and performance of statistical learning in discriminating wetland types using Landsat time series and geomorphological variables computed from Light Detection and Ranging (LiDAR) and Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM). Our study showed that there was a discrimination limit of statistical learning in wetland mapping. The approach was clearly inadequate in distinguishing certain wetland types. In semiarid Australia, our results suggested that the appropriate level for floodplain wetland mapping included four classes: tree-dominated woodlands, shrublands, vegetated swamps, and non-flood-dependent terrestrial communities. Our results also demonstrated that the geomorphological metrics significantly improved the accuracy of wetland classification. Furthermore, geomorphological metrics derived from the freely available coarser resolution SRTM DEM were as beneficial for wetland mapping as those extracted from finer scale commercially-based LiDAR DEM. The finding enables the widespread applications of our approach, as both data sources are freely available globally.

Coastal Wetland Mapping Using Ensemble Learning Algorithms: A Comparative Study of Bagging, Boosting and Stacking Techniques

“…The original map discriminates wetlands into 13 major groups based on dominant plant species. Since large-scale wetland classification with satellite imagery has practical limitations to the number of classes [45], we aggregated the 13 groups into five broad wetland types based on the vegetation structure and the salinity; including three freshwater types (i.e., forested wetland, shrubland wetland, and grassy wetland) and two saline wetlands (i.e., mangrove and saltmarsh). In addition, we randomly sampled 2888 points from other land covers using the land use/land cover map of NSW (NSW Landuse 2017, downloaded on 12 August 2019 from https://data.nsw.gov.au/data/dataset/nsw-landuse-2017) as "background" points.…”

“…Many performance metrics or measures, such as Log-Loss, kappa coefficient, producer's and user's accuracy, and AUC, are given in the literature with the aim of providing a better measure of mapping accuracy [82]. Among these metrics, overall and individual class accuracies derived from the error matrix are the simplest, therefore have been widely used to evaluate model performance (e.g., References [45,83]). The kappa coefficient, which measures the difference between the actual agreement in a confusion matrix and the chance agreement, provides a better measure for the accuracy of a classifier than the overall accuracy as it takes into account the whole confusion matrix rather than the diagonal elements alone [84].…”

“…Ensemble algorithms, such as RF [44] and Gradient Boosting Machine (GBM) [45], are now commonly applied in the field of remote sensing, and generally produce satisfactory results. The performance and potential of other EL methods, such as extreme gradient boosting (XGBoost) [46] and bagged trees [47], are rarely compared and tested for wetland mapping.…”

Exploring Wetland Dynamics in Large River Floodplain Systems with Unsupervised Machine Learning: A Case Study of the Dongting Lake, China

“…However, the results of all three employed methods (i.e., elbow, gap statistic, and average silhouette) suggested that our approach might not be consistent for finer (i.e., above three classes) clustering. Powell et al [42] suggested that discrimination limit for semi-arid floodplain using Landsat based NDVI metrics was four classes, beyond which the performance was significantly deteriorated.…”

Interannual Variability of Global Wetlands in Response to El Niño Southern Oscillations (ENSO) and Land-Use

“…Numerous studies have been conducted based on remote sensing both at regional and global scale to analyze the wetlands changes (Pham-Duc et al, 2017;Aires et al, 2018;Wohlfart et al, 2018;Li et al, 2019;Powell et al, 2019). Microwave datasets have an advantage of not being affected by the cloud cover with higher sensitivity to detect inundation changes (Schroeder et al, 2015).…”