Short Courses

Roland Doerffer

GKSS Research Centre, Institute for Coastal Research

e-mail: roland.doerffer@gkss.de

Short Course Description

Artificial neural networks can be used to interrelate two sets of variables in a non-linear way. By this they function as a multiple non-linear regression method.
For coastal remote sensing we have the problem that a number of independent variables, such as phytoplankton pigment, yellow substance and inorganic suspended matter, determine the water leaving radiance reflectance spectrum.  To determine the concentrations or the IOPs of these components from a number of spectral bands requires an inversion scheme.
NNs can be used to describe the forward relationship, i.e. the reflectance in a number of spectral bands as a function of a number of water components, and they can be used as well to describe the inverse relationship.
To create a NN it is necessary to set up a data set of corresponding pairs of these variables, which should cover the range of interest of the independent variables (in our case the IOPs or concentrations) with sufficient density. This data set, which can be based on measurements or simulations, is then used to train a NN, i.e. to determine the coefficients of the “neurons” (analogue to the regression coefficients of a simple linear regression). Furthermore, before the training run, the architecture of the NN has to be defined (number of layers and neurons per layer), which depends on the complexity of the problem.
Finally, the NN has to be tested using an independent test data set. Tests have also to be performed also during the training run to avoid “overtraining” of a network.
In the short course the basic principles will be explained. Furthermore, we will work on an example and create a simple network for coastal water and test this with reflectance spectra.
Software and an example data set will be distributed. Participants should bring their notebook computer to work on the data. However, the hole procedure will be demonstrated, so that a participant will benefit from the demonstration also without a computer.