Part 1 - Background

Problem:

Identify seagrass beds and discriminate them from bare sand and macroalgae (seaweed) beds, using Planetscope or Rapideye imagery.

Justification:

Seagrass beds have been declining over time in the region. Can we quantify their current extent and monitor them in the future?

Study site:

Two areas on the coast of Pernambuco (Figure 1).

Figure 1 - Location of the study sites.

Part 2 - Analysis

Planet Imagery:

We obtained images from the PlanetScope constellation (B,G,R,NIR - 3.5m pixel size) and from the RapidEye constellation (B,G,R,RE,NIR - 5m pixel size). According to the dates available in the Planet Explorer database and the dates of low tide provided by collaborator Karina, we obtained the following images(Figures 2 to 4):

Figure 2. PlanetScope image of Area 1 from 2019-02-21.

Figure 3. RapidEye image of Area 1 from 2019-02-20.

Figure 4. PlanetScope image of Area 2 from 2019-02-21.

Field data

We also obtained a set of field coordinates that identify the type of underwater cover for the locations. The data was not quality-controlled, with the same class type is typed in different ways. To solve that, we reclassify the original levels into consistent names (Figure 5).

## OGR data source with driver: GPKG 
## Source: "/mnt/Dados/GoogleDrive/Work/Research/Karine/coordenadas_sad69.gpkg", layer: "coordenadas_sad69"
## with 125 features
## It has 6 fields
## [1] "alga"       "alga_rocha" "areia"      "croa"       "H.wrightii"
## [6] "rocha"      "sedimento"

Figure 5. Location of the ground truth points.

Wethen extracted the pixel responses from the field coordinates, and made scatterplots for all band pair combinations, them to test if we can obtain spectral differentiation. Classes that are likely to be separable in the images should for distinct cluster on at least one of the scatterplots. We did a single pixel extraction of the points, as the targets of interest are expected to be small.

The first extraction was for Area 1, from the PlanetScope image (Figure 6).

Figure 6. Separability plot for the classes of interest on Area 1 - PlanetScope.

We then repeated the process for the RapidEye image of Area 1 (Figure 7), and the Planetscope Image of Area 2 (Figure 8)

Figure 7. Separability plot for the classes of interest on Area 1 - RapidEye.

Figure 8. Separability plot for the classes of interest on Area 2 - PlanetScope.

Part 3 - Conclusions

The separability analysis shows almost complete overlap between the classes on all images and band combinations, meaning the algorithms would not be able to differentiate them. This occurs due to the combined effect of sediment load and water column depth, which scatters or and asbosrbs most of the incoming sunlight, resulting on a very weak return signal.