Superpixel for seagrass mapping: a novel method using PlanetScope imagery and machine learning in Tauranga harbour, New Zealand

Seagrass ecosystem provides valuable ecosystem services and is significant blue carbon sink. This resource, however, has been degraded across the globe with a loss rate of 7% year(-1) to the end of twentieth century. The loss of seagrass meadows might lead to an unexpected emission of CO2 into the atmosphere, aggravating global warming and resulting in potential damages to regional ecology and economies. Accurate mapping of meadows extent in different coverages from remotely sensed data, therefore is in high demand as the first step in the strategy of monitoring, report, verification (MRV) that underpins large scale conservation of global seagrass. Despite the higher accuracy of seagrass mapping in recent years, several challenges still persist, particularly when dealing with degraded, sparse seagrass meadows. In this research, we propose a novel and high accuracy approach for mapping dense and sparse meadows of the small size Zostera muelleri seagrass, using high spatial resolution imagery (PlanetScope) at 3 m spatial resolution, and advanced machine learning (ML) models for a ten-fold cross-validation superpixel-based classification in Tauranga Harbour, New Zealand. We archive high mapping accuracy (overall accuracy = 0.913, Kappa coefficient (kappa) = 0.786, Matthews correlation coefficient (MCC) = 0.796 and F-1 = 0.908) using the LightGBM model from a set of superpixel image coupled with the Bayesian optimization for hyper-parameter tuning. Our proposed approach is solid and reliable with evidences of improving kappa (10%) and MCC (11%) when compared with pixel-based image classification, and is expected to provide novel, effective techniques for quantifying the spatial distribution and area of seagrass ecosystem worldwide.

Automated summary

This study proposes a novel and highly accurate approach for mapping dense and sparse meadows of Zostera muelleri seagrass using high spatial resolution imagery and advanced machine learning models. The proposed approach achieved high mapping accuracy through ten-fold cross-validation superpixel-based classification, and showed promising results in Tauranga Harbour, New Zealand. This approach is significant in addressing the challenges of seagrass mapping and is expected to provide effective techniques for quantifying the spatial distribution and area of seagrass ecosystem worldwide.