In 2015 Rachael Holmes conducted a master’s research project examining the use of remote sensing data to monitor change in seaweed habitat.
Remote sensing has high potential for the provision of information about the abundance and distribution of intertidal macroalgae species. Remote sensing can inform biodiversity conservation activities such as those outlined in the European Union’s (EU) Habitats Directive. The extent of seaweed habitat around the UK is estimated at 71,000 km2, with 18,700 km2 of this on rocky shores which can be difficult and dangerous to access (Yesson et al. 2015). In England there is a nationwide monitoring scheme in place run by the Channel Coast Observatory (CCO) that acquires very high-resolution imagery of the coastline every four years. The extent of the dataset provides the opportunity to develop spatially transferable models that reduce the need for ground surveys, creating a feasible and free method for intertidal vegetation mapping.
This study explores the application of false colour infrared (FCIR) aerial photography as a resource efficient method for quantifying exposed macroalgae at inaccessible locations using machine learning. Four ground surveys were carried out between January 2014 and May 2015 to acquire information on all cover types and were used for accuracy assessment and cross-validation. A support vector machine (SVM) based approach to supervised classification is used to determine the transferability of habitat classification between Thanet in southeast England, and the Isle of Wight in southern England. An SVM was chosen due to its ability to perform well with limited training data. The approach was validated at the local level against a dual approach using a minimum noise fraction (MNF) rotation and SVM classification and found to provide equal class agreement (Kappa=0.98).
Spatial transferability was successful when using one site to project into three unseen test sites with a total accuracy of 94.23%, with the lowest class accuracy being 70.83% for red algae. Three sites to characterise one unseen site increased the accuracy to 94.85%, raising the minimum class accuracy to 75.00%. The approach was most effective as a 6-class classification, with decreased accuracy when reducing the model to a binary classification for species’ of interest.
Holmes R (2015) Developing Spatially Transferable Models of Intertidal Macroalgae Distribution using False Colour Infrared Aerial Photography and Support Vector Machine Supervised Classification. Thesis submitted for consideration towards a degree of
MSc Remote Sensing, Dept. of Geography, UCL (University College London). Supervised by Chris Yesson & Juliet Brodie. (PDF)