Can suspended sediment concentrations be estimated from multispectral imagery using only image-derived information?

Abstract

Sampling for suspended sediment concentrations (SSC) in inland waters is traditionally based on collecting samples at sparse locations and in limited intervals. A number of investigators explored the utility of earth-observing satellites and air-borne sensors for monitoring of SSC over vast areas. Two approaches are commonly deployed: (1) empirical relationships between a chosen remotely sensed quantity and the actual in-situ SSC; and (2) bio-optical models founded on radiative transfer modeling. Unfortunately, in-situ measurements are often unavailable for direct image calibration, and inherent optical properties of optically active constituents (specific scattering and absorption coefficients) are usually unknown. This paper examines the possibility to retrieve SSC from multispectral satellite imagery without any in-situ data, i.e. using only image-derived information. The fundamental principle of image selfcalibration relies on the fact that in the visual domain of wavelengths (∼400–700 nm) the at-sensor reflectance becomes “saturated“ at high SSC, whereas the near-infrared domain (∼700–900 nm) remains almost perfectly linearly related to sediment concentrations. The core idea of the self-calibrating procedure is rather simple and is based on fitting an exponential function between reflectance and SSC, with SSC replaced by a linear relationship between SSC and reflectance in the near-infrared domain. As a first approximation of the non-linearity between reflectance and SSC levels in the 400–700 nm range, we used the equation proposed by Schiebe et al. (1992), although other equations, especially those arising from optical theory could be used as well. The technique is illustrated on a moderately sediment-laden reservoir and two scenes acquired from Landsat ETM+. The standard error of the estimated SSC was below 15 mg/L (i.e. ∼25 % relative error for the observed range of SSC). Although the proposed algorithm does not yield better results than other models mentioned in the literature, the primary advantage of the outlined methodology is that no in-situ measurements (water sampling nor spectral profiling) are needed — i.e. only image-derived information is used.