Nonlinear neural-network-based mixture model for estimating the concentration of nitrogen salts in turbid inland waters using hyperspectral imagery

Abstract

The development of hyperspectral imaging instruments designed for water quality assessment, such as the DLR Reflective Optics System Imaging Spectrometer (ROSIS), has created a need for methods which are able to infer water quality parameters of turbid inland waters, and to use those parameters as indicators for water quality. It has been reported that the irradiance reflectance and, subsequently, the radiance collected by the sensor in such scenario is usually the result of an intimate mixture of sub-pixel components. As a result, the commonly used linear mixing model may not be appropriate to describe materials composition. In this work, we develop a nonlinear neural network-based algorithm for estimating water constituent concentrations, with special emphasis on the detection of chemical substances provided by agricultural and industrial sources. The proposed neural network architecture consists of a modified multi-layer perceptron (MLP) whose entries are determined by a linear function activation provided by a Hopfield neural network algorithm (HNN). The combined HNN/MLP supervised model has been used to estimate the concentration of water constituents by training the MLP with ground spectra of nitrogen salts, which are commonly used in extensive agricultural farms. Such spectra were collected using a Minolta spectro-photometer. The model was calibrated in our laboratory by using mixtures of water and nitrogen salt in different proportions. Hyperspectral images collected by the ROSIS imaging spectrometer over the Guadiloba reservoir in Cáceres, SW Spain, are also used in this study to estimate the concentration of nitrogen salts in turbid inland waters.