Retrieving TSM concentration from multispectral satellite data by multiple regression and artificial neural networks

In this paper, we present different methodologies to estimate the total suspended matter (TSM) concentration in a particular area of the Portuguese coast, from remotely sensed multispectral data, based on single-band models, multiple regression, and artificial neural networks (ANNs). Simulations on different beaches of the study area were performed to determine a relationship between the TSM concentration and the spectral response of the seawater. Based on the in situ measurements, empirical models were established in order to relate the seawater reflectance with the TSM concentration for TERRA/ASTER, SPOT HRVIR, and Landsat/TM. Seven images of these three sensors were calibrated and atmospherically and geometrically corrected. Single-band models, multiple regression, and ANNs were applied to the visible and near-infrared (NIR) bands of these sensors in order to estimate the TSM concentration. Statistical analysis using correlation coefficients and error estimation was employed, aiming to evaluate the most accurate methodology. The chosen methodology was further applied to the seven processed images. The analysis of the root-mean-square errors achieved by both the linear and nonlinear models supports the hypothesis that the relationship between the seawater reflectance and TSM concentration is clearly nonlinear. The ANNs have been shown to be useful in estimating the TSM concentration from reflectance of visible and NIR bands of ASTER, HRVIR, and TM sensors, with better results for ASTER and HRVIR sensors. Maps of TSM concentration were produced for all satellite images processed.