Mapping Tropical Wetland Dynamics using Radar Remote Sensing

Abstract

Mapping the spatial and temporal dynamics of tropical wetland environments is important for a wide range of applications, including greenhouse gas emission and storage, biodiversity, protection of ecosystem services, public health as well as food and water security. Yet inventories of some of the world’s largest and most important wetland systems are limited or inaccurate. Moreover, many inventories are static in nature, based on maps of mean or maximum extent and therefore do not account for the inherent variability in these complex ecosystems. Satellite Earth Observation has the potential to provide timely maps of inundation over large areas, but most existing approaches for mapping inundation are relatively coarse spatial resolution and/or they only detect open water and therefore significantly underestimate inundation extent. Satellite radar imagery is sensitive to open water and inundated vegetation and is not a↵ected by cloud cover enabling uninterrupted observations of the Earth’s surface allowing for intra and inter annual inundation dynamics to be investigated. Previous attempts to map inundation within herbaceous wetland systems using Sentinel 1 C-Band radar have sought to exploit a double bounce backscatter interaction as a means of detection. Yet there are no reported attempts to quantify this mechanism using field observations in a natural wetland setting. Field observations of vegetation characteristics and the presence of inundation in the Rupununi Wetlands in Guyana was compared against coincidental Sentinel 1 C-Band Synthetic Aperture Radar backscatter. Whilst areas of inundated vegetation were shown to
exhibit a double bounce like interaction with the C-Band Sentinel 1 radar pulse there was considerable overlap in backscatter response between areas of inundated and non-inundated vegetation. This evidence suggests that inundation cannot be mapped using simple thresholding and, therefore, rather a more sophisticated approach is required, such as machine learning methods. Previous wetland mapping attempts using radar remote sensing have shown promising
results, but they are localised in scale, giving no sense of their scalability to allow for wetland mapping on regional or continental scales. In response to this need, a novel inundation mapping routine was developed termed RadWet, which
utilises high time series variability in backscatter response as a means of automatically generating training data describing areas of inundated vegetation, supplied to a machine learning classifier to detect areas of inundated vegetation, as well as open surface water. RadWet was developed and tested in herbaceous dominated wetland environments utilising serial Sentinel-1 radar imagery acquired over the Barotseland Floodplain in Western Zambia, and the Rupununi Wetlands in Guyana over the period 2017–2022 every 12 days. Good agreement was found at both sites using random stratified accuracy assessment data (n = 28,223) with a median overall accuracy of 89% in Barotseland and 80% in the Upper Rupununi, outperforming existing approaches. The results revealed fine-scale hydrological processes driving inundation patterns as well as temporal patterns in seasonal flood pulse timing and magnitude, from which it has been possible to analyse inter and intra-annual variations in inundation extent. Inundated vegetation dominated wet season wetland extent, accounting for a mean 80% of total inundation, highlighting its importance as a target for future wetland mapping attempts. Despite it’s success in herbaceous dominated wetlands, RadWat was not able to detect inundation within forested environments, where dense vegetation foliage caused volume scattering saturation within the C-Band radar data. As such Rad-Wet was further developed and modified to map areas of inundated forests across the Amazon River basin using L-Band synthetic aperture radar imagery, which can recieve a signal from underneath the forest canopy due its relatively long wavelength. Applied over serial ALOS-2 PALSAR-2 L-band imagery, RadWet has been able to produce, for the first time, inundation maps across the whole Amazon River Basin with a 42 day revisit period, at a spatial resolution of 50 m. Timeseries estimates of inundation extent from RadWet-L was significantly correlated with NASA-GFZ GRACE-FO water thickness (R = 0.96, p < 0.01), USDA GREALM lake hight (Pearson R between 0.63 – 0.91, p < 0.01), and in-situ river stage measurements (R between 0.78 – 0.94, p < 0.01). Additionally, an evaluation of validation points against the input ALOS-2 PALSAR-2 data revealing spatial and temporal consistency in the approach (F1 score = 0.97). Serial classifications of ALOS-2 PALSAR-2 data by RadWet-L can provide unique insights into the spatio-temporal inundation dynamics within the Amazon Basin. Applying Rad-Wet in new areas will allow for comprehensive wetland inventories to be produced exposing inter and intra annual dynamics in response to global scale climate events such as El Nino, La Nina and the Indian Ocean Dipole, as well response to climate shocks such as extended periods of drought or excess precipitation.

Date of Award	2025
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Andy Hardy (Supervisor) & Pete Bunting (Supervisor)