Seagrass and Seaweed Detection Approaches Using Remote Sensing and Google Earth Engine: A comparative Analysis of Different Machine Learning Techniques

Abstract

Seagrasses and seaweed habitats contribute to crucial ecological services globally, from capturing carbon dioxide and supporting 20% of the world’s largest fisheries to sustaining the small, but many coastal communities [1]. Across the globe, an alarming decline in their wild distribution has been recorded, attributed to climate change and direct pollution [2]. Current estimates of how much the loss is are uncertain and mapping and monitoring efforts are costly, data-intensive, and lack scalability. Thus, freely available data and software in remote sensing, coupled with Machine Learning (ML) are deemed important means to leverage existing mapping of seagrasses and seaweed spatial distribution [3, 4]. This thesis explored a free and scalable workflow by comparing three different ML techniques mainly on Overall Accuracy (OA) and Tau(e) in classifying seagrass, seaweed, and water. These are supervised, unsupervised, and semi-supervised learning (SSL) which used data from the satellite, Sentinel-2 Level-2A, applied to a novel area of study, from Biddeford Pool to Small Point at the Coast of Maine, United States of America. Results showed that the SSL achieved the highest OA of 76% and Tau(e) = 0.72 on the hard test, in line with previous work. To the best of the author’s knowledge, this work contributes to the field of science by being the first in its field to use the geospatial analysis package ’geemap’, along with the software Google Earth Engine, and SSL for classifying seagrass and seaweeds. Through demonstration, this work shows the potential of free data in remote sensing, leveraged by ML to aid community monitoring in the environmental management of seagrass and seaweed. The results here can be considered as a starting point for further exploring the SSL paired with freely available data and community monitoring to lower costs, handle data scarcity, and scale up in the field of aquatic vegetation mapping and monitoring.

Seagrasses and seaweed habitats contribute to crucial ecological services globally, from capturing carbon dioxide and supporting 20% of the world’s largest fisheries to sustaining the small, but many coastal communities [1]. Across the globe, an alarming decline in their wild distribution has been recorded, attributed to climate change and direct pollution [2]. Current estimates of how much the loss is are uncertain and mapping and monitoring efforts are costly, data-intensive, and lack scalability. Thus, freely available data and software in remote sensing, coupled with Machine Learning (ML) are deemed important means to leverage existing mapping of seagrasses and seaweed spatial distribution [3, 4]. This thesis explored a free and scalable workflow by comparing three different ML techniques mainly on Overall Accuracy (OA) and Tau(e) in classifying seagrass, seaweed, and water. These are supervised, unsupervised, and semi-supervised learning (SSL) which used data from the satellite, Sentinel-2 Level-2A, applied to a novel area of study, from Biddeford Pool to Small Point at the Coast of Maine, United States of America. Results showed that the SSL achieved the highest OA of 76% and Tau(e) = 0.72 on the hard test, in line with previous work. To the best of the author’s knowledge, this work contributes to the field of science by being the first in its field to use the geospatial analysis package ’geemap’, along with the software Google Earth Engine, and SSL for classifying seagrass and seaweeds. Through demonstration, this work shows the potential of free data in remote sensing, leveraged by ML to aid community monitoring in the environmental management of seagrass and seaweed. The results here can be considered as a starting point for further exploring the SSL paired with freely available data and community monitoring to lower costs, handle data scarcity, and scale up in the field of aquatic vegetation mapping and monitoring.