Seeing below the surface: Mapping freshwater vegetation with remote sensing
This post discusses the publication: "Multi-Scale Spectral Separability of Submerged Aquatic Vegetation Species in a Freshwater Ecosystem."
Why study freshwaters?
To date, aquatic remote sensing has largely focused on clear, shallow coastal waters where it’s relatively easy to see the bottom, but comparatively few researchers work where the water is harder to see through, like in lakes and rivers. So, I wanted to fill in some of the existing knowledge gaps in remote sensing of freshwaters to provide a foundation for future work. I decided to use a multi-scale approach to detect and identify submerged targets, specifically submerged aquatic vegetation (SAV). If remote sensing is to be successful in mapping SAV, I first had to determine if different classes of SAV have enough spectral diversity to be discriminable based on reflectance alone (i.e., are they spectrally separable?). If so, I then wanted to investigate how that theoretical spectral diversity translates to a practical application. I additionally opted to look at how the choice of sensor, the sampling season, the coarseness of SAV class definitions, and the condition of the leaves might affect separability results as each of these could guide future project design.
Up close: How separable are SAV leaves?
First, to see if SAV would be a suitable subject for mapping with optical remote sensing, we collected plant samples from a study site in the Long Sault Parkway recreational area, in southeastern Ontario, Canada. Each sampling location was recorded with a Global Navigation Satellite System (GNSS) receiver to create a set of ground truth points that could be used later as input data for mapping SAV in airborne imagery. The plant samples were brought to a darkroom and their leaves were measured using a hand-held spectrometer that produced very high spectral resolution (i.e., 1776 bands between 375 nm to 2150 nm) spectra of each leaf (or leaf segment for long, ribbon-like leaves). All those bands were then ranked according to how well they could be used to discriminate between different species of SAV; the most effective bands were kept, and the rest were discarded. The maximum separability amongst classes was determined using only that culled band set. This process was repeated multiple times to examine four variables: using measurements taken from cleaned samples to see if the biofilm that covers aquatic plants affects spectral separability; using spectra that had been resampled to emulate other sensors to assess the effect of a sensor’s spectral resolution on separability; using samples collected later in the season to see if growth stage matters in discriminating between classes; and, by varying how coarsely the SAV classes are defined (e.g., species, genus, kingdom) to see how the spectral variability within a target itself may affect separability.
Our study site at the Long Sault Parkway. We focused our research on a small shallow bay just west of Philpott’s Island (yellow box). A flooded road and structural remnants are clearly visible. Figure from Rowan et al. (2021) used according to CC BY 4.0.
The results showed that SAV is, indeed, a good subject for optical remote sensing with hyperspectral data. Even at the finest class definition, the species, SAV samples could be confidently separated amongst one another. This separability only increased as the classes became more coarsely defined. The biofilm that covers SAV leaves had very little effect on the separability, so future researchers need not worry if their targets aren’t perfectly clean in situ. The samples collected later in the season were however much less separable than those from the peak of the growing season, meaning that future projects looking at similar targets may want to prioritize collecting their data before flowering and senescence begins. Lastly, resampling the signals to emulate other sensors had a profound effect on SAV separability; spectral separability was strongly dependent on the number of bands of the sensor. Spectra resampled to imitate another hyperspectral sensor, the Compact Airborne Spectrographic Imager (CASI), maintained their high separability while resampling to multispectral satellite sensors substantially decreased the separability. Based on these results, a practical mapping application would have the best chance of success using a hyperspectral sensor deployed in the peak of the growing season.
From afar: Can freshwater SAV be mapped in situ?
Freshwaters contain more dissolved and suspended material in the water column, things like minerals, nutrient, phytoplankton, or sediments, so discerning between the signal contributions of the water column (and its constituents) and the bottom is more complex than in clear coastal settings, which are in turn more complex than terrestrial applications. I therefore needed to compensate for the effect of the water column if I was going to have any chance at accurately mapping the vegetation. Deep Inamdar, a fellow student at ARSL and a co-author on the published article describing this work, produced a MATLAB program based on Lyzenga’s Depth-Invariant Index (DII) (Lyzenga 1978, 1981) to perform a water column compensation without needing to manually select band combinations (Inamdar et al. 2021, submitted). This program produces a transformed image of DII bands (instead of wavelengths of reflectance) selected to retain most useful information while minimizing redundancy. This method was applied to a hyperspectral image of the Long Sault Parkway site collected using a CASI mounted on a small research aircraft and produced an image of 124 DII bands. That image was then made into a Directly-Georeferenced Hyperspectral Point Cloud to help managed the effects of spatial uncertainty in the data and imagery. If you haven’t yet, check out Deep’s blog post about rasters and the DHPC, I promise it’s worth a read. You can find it here:
The three major stages in mapping SAV from aerial imagery: the original CASI image of the study site showing various bottom covers and flooded infrastructure; a raster image presenting the DII transformed image after being made into a DHPC – each coloured dot represents a data point from the CASI imagery; the results of detecting Potamogeton sp. in the DII image – red dots are ground truth points indicating Potamogeton presence, points correctly identified are circled in green.
As the CASI image has a spatial resolution of roughly 1 m, I focused only on detecting bottom cover classes that would be present in extent of a square meter or larger: plants with ribbon-like leaves (Sagittaria graminea and Vallisneria americana combined), metaphyton, Chara sp., Potamogeton richardsonii, Potamogeton sp., paved asphalt, and unvegetated silt/rock. Apart from Metaphyton, these classes were well detected (recall values between 79% and 100%; overall recall of 88%). Metaphyton’s poor detectability (recall of 0%) might be due to a lack of input data, by the clouds of filamentous algae being too small, or by the fact that those clouds may not have been dense enough to totally obscure materials below them (thus confounding the signal).
What does it all mean?
Combining the separability and practical mapping results shows that there is huge potential to map vegetation in freshwaters using remote sensing, to shed light on what is beneath the surface of lakes and rivers. Accurate, timely mapping and monitoring will contribute to maintaining clean and clear waterways by supporting conservation and restoration programs. However, these results also show that if we want to effectively leverage remote sensing for freshwater research, we need investment and innovation to make sensors with higher spectral and spatial resolutions widely available.
• Inamdar, D., Kalacska, M., Arroyo-Mora, J.P., & Leblanc, G. (2021). The Directly-Georeferenced Hyperspectral Point Cloud: Preserving the Integrity of Hyperspectral Imaging Data. Frontiers in Remote Sensing, 2. doi: 10.3389/frsen.2021.675323
• Inamdar, D., Rowan, G., Kalacska, M., Arroyo-Mora, J.P. (2021, submitted). Water Column Compensation Workflow for Hyperspectral Imaging Data.
•Lyzenga, D. R. (1978). Passive Remote Sensing Techniques for Mapping Water Depth and Bottom Features. Appl. Opt. 17 (3), 379–383. doi:10.1364/AO.17.000379
•Lyzenga, D. R. (1981). Remote Sensing of Bottom Reflectance and Water Attenuation Parameters in Shallow Water Using Aircraft and Landsat Data. Int. J. Remote Sensing 2 (1), 71–82. doi:10.1080/01431168108948342
• Rowan Gillian S. L., Kalacska Margaret, Inamdar Deep, Arroyo-Mora J. Pablo, Soffer Raymond. (2021) Multi-Scale Spectral Separability of Submerged Aquatic Vegetation Species in a Freshwater Ecosystem. Front. Environ. Sci., 9. doi: 10.3389/fenvs.2021.760372