Mapping Spatial Distribution and Biomass of Intertidal Ulva Blooms Using Machine Learning and Earth Observation

Abstract

Opportunistic macroalgal blooms have been used for the assessment of the ecological status of coastal and estuarine areas in Europe. The use of earth observation (EO) data sets to map green algal cover based on a Normalized Difference Vegetation Index (NDVI) was explored. Scenes from Sentinel-2A/B, Landsat-5, and Landsat-8 missions were processed for eight different Irish estuaries of moderate, poor, and bad ecological status using European Union Water Framework Directive (WFD) classification for transitional water bodies. Images acquired during low-tide conditions from 2010 to 2018 within 18 days of field surveys were considered. The estimates of percentage coverage obtained from different EO data sources and field surveys were significantly correlated (R2 = 0.94) with Cohen’s kappa coefficient of 0.69 ± 0.13. The results showed that the NDVI technique could be successfully applied to map the coverage of the blooms and to monitor estuarine areas in conjunction with other monitoring activities that involve field sampling and surveys. The combination of wide-spread cloud-coverage and high-tide conditions provided additional constraints during the image selection. The findings showed that both Sentinel-2 and Landsat scenes could be utilized to estimate bloom coverage. Moreover, Landsat, because of its legacy program, can be utilized to reconstruct the blooms using historical archival data. Considering the importance of biomass for understanding the severity of algal accumulations, an artificial neural networks (ANN) model was trained using the in situ historical biomass samples and the combination of radar backscatter (Sentinel-1) and optical reflectance in the visible and near-infrared (NIR) regions (Sentinel-2) to predict the biomass quantity. The ANN model based on multispectral imagery was suitable to estimate biomass quantity (R2 = 0.74). The model performance could be improved with the addition of more training samples.