Widya, L. K., Kim, C. H., Do, J. D., Park, S. J., Kim, B. C., & Lee, C. W. (2023). Comparison of satellite imagery for identifying seagrass distribution using a machine learning algorithm on the eastern coast of South Korea. Journal of Marine Science and Engineering, 11(4), 701. Click (a) GeoEye-1 based seagrass distribution map : A precise detection of seagrass habitat boundaries along the Uljin-gun coast, achieved by applying the SVM algorithm to high-resolution (2m) GeoEye-1 satellite data. (b) Sentinel-2 based seagrass distribution map : Classification results using medium-resolution (10m) multispectral Sentinel-2 data, stably identifying the overall distribution pattern. (c) Landsat-8 based seagrass distribution map : Results using 30m spatial resolution Landsat-8 OLI data, intuitively demonstrating the differences in spatial resolving power and vegetation mapping details compared to the previous sensors. (d) GeoEye-1 confusion matrix : Among the three satellite sensors, GeoEye-1 yielded overwhelmingly superior detection performance, achieving an overall accuracy of 92% and a Kappa coefficient of 0.89. The confusion matrix analysis quantitatively proves that even in complex coastal environments, 2m ultra-high-resolution data can most perfectly separate seagrass from other topographical features (Land, Others). This study evaluates the performance of three optical satellite sensors with varying spatial resolutions—GeoEye-1, Sentinel-2, and Landsat-8 OLI—for monitoring the distribution of seagrass, a key indicator of coastal ecosystem health on the eastern coast of South Korea (Uljin-gun). Following rigorous preprocessing steps including ACOLITE atmospheric correction, Hedley sunglint removal, and Lyzenga water column correction, a Support Vector Machine (SVM) algorithm was applied to classify seagrass habitats.The results indicated that GeoEye-1, with the highest spatial resolution (2m), outperformed other sensors in delineating complex seagrass boundaries, achieving an overall accuracy of 92% and a Kappa coefficient of 0.89. Notably, the ROI extraction process was optimized by employing 'Random points inside polygons' in QGIS, ensuring that multiple sampling points within each polygon better represented the spatial distribution compared to single-point sampling. This research establishes an optimized satellite sensor selection and analytical workflow for large-scale coastal vegetation monitoring and provides a critical foundation for blue carbon ecosystem conservation. Continue reading...
