Plenary Session 3
Tuesday, October 4
Automating benthic change detection at large scales using model inversion techniques
John Hedley, Numerical Optics, Ltd
Satellite-based remote sensing of shallow waters offers great potential for monitoring environments which are important for ecosystem services and conservation, such as seagrass beds and coral reefs. While many efforts are focussed on mapping of benthic types, a more fundamental objective is the ability to detect benthic change regardless of the types present. For habitats that range over hundreds of square kilometres, flagging areas where ‘something’ has changed can efficiently target follow-up field surveys and form the basis of understanding the drivers of change at large scale.
However, reliably detecting benthic change is problematic due to the many confounding factors that occur between the bottom of the water column and the satellite sensor: variable water column constituents, tide state, sea surface state, floating algae, clouds, cloud shadows, atmospheric constituents and solar-view geometry – all these factors act to reduce confidence in any apparent change identified between a pair of images. Unless these uncertainties can be managed, automating change detection at large scales may be too unreliable to be of practical use.
In this talk I will describe progress toward a system for robust automated benthic change detection that aims to address the problem of the many uncertainties that can occur. Using this system, bathymetry and seagrass canopy density (Thalassia testudinum leaf area index, LAI) have been mapped over an area of 900 km2 of reef lagoon along the Mexican Yucatán coast, for each year from 2016 to 2020. The mapping was based on radiative transfer model inversion methods applied to each image of the entire archive of Sentinel-2 acquisitions in the area (approximately 800 images). By processing all images, all sources of variability were propagated to give a set of results for each pixel of the per-year output products. The spread of values at each pixel provided the basis for statistical testing of year-to-year change. Importantly, a statistical significance level can be assigned to any apparent change detected.
In this case study, the results in each year compared favourably to ICESat-2 data for bathymetry (mean absolute error < 1 m), and to field surveys for canopy density (mean absolute error 0.59 on a range of LAI from 0 to 5). Identified changes in seagrass canopies (p < 0.01) were localised and spatially plausible. Rejecting changes with statistical significance lower than 1% (p > 0.01) was effective at removing noise from the change maps. However, the identified changes, which often comprised a subtle darkening or brightening of the bottom reflectance, correlated very well in space and time with a delayed effect of documented incursions of pelagic floating Sargassum in 2018 and 2019. The working hypothesis is that these benthic changes are the subsequent decomposition of Sargassum into organic material in the sediments followed by later dispersion, rather than changes in seagrass canopies per se. Hence, the approach appears to be effective and robust at detecting benthic change, but careful interpretation of the basis of the change is required. While monitoring shallow water environments over large scales may be technically achievable, the ecological interpretation of detected changes remains challenging.