Seagrass mapping using high resolution multispectral satellite imagery: A comparison of water column correction models

Abstract

Satellite remote sensing is an efficient and economical technique for studying coastal bottoms in clear and shallow waters. Accordingly, the main objective of this study is the generation of benthic maps using high spatial resolution multispectral images from the WorldView-2/3 satellites. In this context, one of the main challenges consists of eliminating the disturbances caused in the signal by the atmosphere, the sea surface, and the water column. Regarding the water column correction, there is controversy about its effectiveness to improve the results achieved. To assess the impact of the water column correction in seagrass mapping, two coastal areas with different characteristics have been selected. Specifically, an analysis has been carried out consisting of the assessment of the Lyzenga and Sagawa water column correction models to identify the algorithm that provides the best mapping precision and, additionally, to seek if this pre-processing stage is helpful when classifying the seabed. The classification models selected for the study were: Gaussian Naïve Bayes (GNB), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Subspace KNN (S-KNN). Machine learning techniques have proven to achieve better results and, in particular, SVM and KNN models provide the best overall accuracy. The results after benthic mapping have demonstrated, that image classification without water column corrections provides better accuracy (95.36% and 99.20%) than using Lyzenga (73.49% and 97.80%) or Sagawa (82.04% and 99.10%), for Case 2 and 1 waters, respectively.