A Deep Dive into Mapping Caribbean Benthic Habitats with eCognition

One of the first “jobs” I had was volunteering with the local branch of The Nature Conservancy (TNC) in Montpelier, Vermont. So it has been great to see Trimble’s relationship with TNC grow over the years.

With the help of the TAMA Group in Germany, Dr. Steve Schill, the Lead Scientist in the Caribbean Division at TNC, has taken on the challenge of mapping tropical benthic habitats, in particular coral reefs and seagrass, at scale and at a high enough resolution to support local actions.

The importance of these fragile habits is immense. Not only do they “provide essential habitat for one-quarter of all known marine species, but they also provide billions of dollars of economic value and direct benefits to at least 500 million people who live in close proximity." Yet, many mapping efforts have focused on the use of 30 m and 60 m resolution data which do not allow for the identification of “narrow fringing or linear habitats” and of course are in general limited to the size of features that can be detected. Hence, benthic cover is underestimated.

Comparison of multiple scale benthic habitat data around the island of Carriacou, Grenada: (a) PlanetScope Dove Classic imagery (4 m); (b) Global-scale Landsat-derived (30 m) five-class coral reef map; (c) National-scale PlanetScope-derived (4 m) thirteen-class benthic habitat map.

In their recent publication entitled “Regional High-Resolution Benthic Habitat Data from Planet Dove Imagery for Conservation Decision-Making and Marine Planning”, Schill et al. created a “first-of-its-kind high-resolution” benthic habitat map with 4 m optical imagery. Using Trimble eCognition software, the team created an automated workflow to map 13 classes across a study area larger than 200,000 square kilometers spanning the Insular Caribbean. PlanetScope (PS) Dove Classic satellite imagery was used and in total 38,642 scenes were analyzed. The Planet sensor was selected due to its advantageous coverage and revisit times, “obtaining a cloud-free observation with calm, clear water conditions and minimal sunglint and turbidity is critical for achieving accuracy in benthic habitat classification” and therefore as the authors state “the foundation for successful image feature extraction is based on the selection of optimal imagery."

Ultimately, a 4-band (RGB & NIR), multi-year image mosaic was generated for the expansive study area – a 300 GB cloud-free, seamless data set – and used as input within eCognition where an object-based analysis was applied.
Final shallow area (<30 m depth) seamless, non-cloudy, four-band image composite of 38,642 PS Dove scenes.

The rule set developed for this project was designed in blocks, “with each block responsible for a certain step in the analysis." The initial step was used to create a land-sea-mask using a combination of vector-based segmentation and classification. In a second step, depth information was used to simply classify the study area into depth increments up to 20 m. A third step considered the deeper areas beyond the range of the bathymetry model – here, RGB spectral characteristics were used.

Subsequently, the reed habitat classes were mapped predominantly based on a combination of RGB and depth data. The authors then took full advantage of the power behind eCognition, applying a “spatial approximation routine” that combined the strengths of image object hierarchy and data fusion capabilities. Their approach used what they call a “parental guidance routine” which was applied “within the polygon outline, the predominant spectral response was calculated, and an object was formed around it." This object then became the seed for new objects, so long as their spectral response remained within a defined range of values. In addition, objects outside the author’s rules yet inside the geomorphic zone polygons were merged via a best-fitting technique, “allowing outside objects to intrude inside the polygon, refining the original object boundary."

Mapping of the reef crest and back and fore reefs: (a) PS imagery; (b) Polygon boundary of the geomorphic zone used in the parental guidance technique to refine the reef type boundaries; (c) Classification of reef crest (red), fore reef (brown), and back reef (orange) using the parental guidance approach; (d) Comparison to the polygon boundary differences indicating the improvement of reef type mapping using the parental guidance technique.

Smaller reefs were detected based on a depth classification – a kernel as applied to “scan” the model and identify elevated humps of a specified elevation threshold.

The detection of seagrass beds vs. sand utilized the spectral characteristics of the entire lagoon zone. An object-specific range was calculated and used as a threshold to differentiate between bright sand and darker seagrass.

After all objects were classified a final step applied a minimum mapping unit to remove objects under a project-determined, class-specific, size threshold.

Given the expanse of the project area and the volume of data, the authors took advantage of the batch and parallel processing capabilities within eCognition Server. The automated routine they developed was not only used for processing the data, but also for an automated project creation via the integration of a customized import routine. This ensured the automatic setup of 604 eCognition projects with all necessary input data and layer alias naming to support the various rule set options. In addition, an eCognition Server grid was set up to support 3 parallel engines which significantly reduced processing time – in the end it only took 2 days and 2 hours to analyze the entire project area.

(a) An example area of the final PS image-derived benthic habitat classification for the Turks and Caicos Islands; (b) Zoomed-in area on the west side of Grace Bay in the northwest corner of Caicos Island showing the detail of the PS imagery; (c) Subsequent benthic habitat classification of the same area.

The area of each benthic habitat class was calculated with the different Exclusive Economic Zone (EEZ) jurisdictions. The authors generated a color-coded table that nicely shows the percentage of each classified benthic habitat as they are distributed across the different protected or managed jurisdictions.

Percentage of benthic habitat class protected or managed by jurisdiction, based on The Nature Conservancy’s declared Caribbean marine protected or managed area database.

To assess the accuracy of the analysis, 2686 points collected in the field between 2010.2017 were compared against the 13 classes. An overall accuracy of 72% was achieved and was “calculated as the stratified (area-weighted) percentage of correctly classified sites in each sample drawn from the classified map." The greatest confusion between classes was observed in the classes dense vs. sparse seagrass and dense vs. sparse hardbottom given the difficulty to distinguish between them, especially at depth.

The regional results of the analysis can be publicly accessed using a web application provided by the TNC: http://caribbeanmarinemaps.tnc.org/. It is a Google Earth Engine app developed to share the results to non-technical stakeholders. As the authors say, “as increasing threats continue to degrade coastal habitats around the world, governments and conservationists greatly benefit from more accurate maps that can strategically guide decision-making, such as adopting new policies, expanding protected areas, increasing resilience, and restoring habitats at broad scales."

It has been exciting to watch this project develop over the past year and the progress the authors have made since the joint webinar in May 2020 on “Broad Scale Automated Mapping of Shallow Benthic Habitats in the Caribbean using Planet Dove Imagery." I look forward to seeing how the TNC will continue to use eCognition in their future projects!