Seagrass species classification using random forest and U-Net models on multiple satellite images taken at various times and places

Seagrass meadows are important ecosystems that provide numerous benefits to both humans and marine life. They also play a vital role in mitigating climate change through their capacity to act as carbon sinks, with different seagrass species storing varying amounts of carbon. Effective monitoring and identification of seagrass meadows are essential for their conservation and management. Remote sensing imagery combined with machine learning offers a rapid approach for detecting seagrass habitats over large areas. This study is the first to map seagrass habitats at the species level using only high spatial resolution imagery with deep learning. We compared segmentation results from a deep learning model (U-Net) with pixel-based classification (Random Forest: RF) for detecting and classifying one non-seagrass class and five seagrass species: Cymodocea serrulata, Halophila ovalis, Halophila spinulosa, Syringodium isoetifolium and Zostera muelleri combined with Halodule uninervis. The study sites were located in Moreton Bay, Brisbane, Australia and Pak Meng, Trang, Thailand. Transfer learning was also explored in this study. The RF model performed well for species-level classification with straightforward predictions, while U-Net demonstrated better performance in detecting seagrass habitats across different image acquisition times, locations and resolutions. Both models showed the capability to map seagrass under varying environmental conditions.