Integration of multi-source data for water quality classification in the Pearl River estuary and its adjacent coastal waters of Hong Kong

Abstract

The spatial patterns of water quality were studied by integrating a Landsat TM image, 58 in situ water quality datasets and 30 samples from two concentration maps of water quality parameters derived from SeaWiFS and NOAA/AVHRR images in the Pearl River estuary and the adjacent coastal waters of Hong Kong. The reflectance of TM bands 1–4 was derived by using the COST method. The normalized difference water index (NDWI) was extracted from the raw image and the threshold segmentation was used to retrieve the water pixels of spectral reflectance. In order to study the spectral reflectance categories related to water quality, a dataset comprising 88 sampling points from four spectral bands of a Landsat TM image was used. The samples were positioned according to the availability of water quality parameters in the study area, and five reflectance classes could be finally distinguished by using the cluster analysis. Three supervised classifiers, maximum likelihood (MLH), neural network (NN) and support vector machine (SVM), were employed to recognize the spatial patterns of ocean color. All the 88 samples were divided into two data sets: 65 in the training data set and 23 in the testing data set. The classification results using the three methods showed similar spatial patterns of spectral reflectance, although the classification accuracies were different. In order to verify our assumption that the spatial patterns of water quality in the coastal waters could be indirectly detected by ocean color classification using the Landsat TM image, five optically active water quality parameters: turbidity (TURB), suspended sediments (SS), total volatile solid (TVS), chlorophyll-a (Chl-a) and phaeo-pigment (PHAE), were selected to implement the analysis of variance (ANOVA). The analysis showed that a statistically significant difference in water quality clearly existed among the five classes of spectral reflectance. It was concluded that the five classes classified by reflectance showed distinct water quality characteristics. Therefore, the ocean color classification based on landsat TM images and regular measurements of water quality may provide a reasonable spatial distribution for the coastal water quality. This may serve as an effective tool for the rapid detection of changes in coastal water quality and subsequent management.