Artificial neural network exploration of the influential factors in drainage network derivation

The interaction between landscape and stream processes has been well documented, and thus it is plausible to assume that landscape factors in a given watershed have a determinant function in stream network initiation and formation. Although past research has identified certain general relationships between stream drainage density and other environmental factors, no definite model for stream delineation has yet emerged. Particularly at smaller scales, drainage network extraction approaches have been devised that are generally independent of often widely heterogeneous landscape conditions. Utilizing digital terrain models alone for this purpose is, in many cases, problematic and inaccurate. This research effort involved using an artificial neural network (ANN) for the exploration of a number of environmental factors and information extracted from satellite images (Landsat 7 ETM+) with regard to finding the most indicative factors for drainage network location. The environmental variables included a series of available topographic, soil and lithology factors. Two hydrologically and geographically diverse case studies, one in Pennsylvania and the other in southern Italy, were used to this end. The results show that various spectral band ratios (1/4, 2/4 and 3/4) were mutually identified as significant in both case studies. Additionally, both case studies illustrated that different factors belonging to general parameter groups (soil and topography parameters) were identified by the ANN for each case, suggesting that future drainage network derivation methodologies should possibly consider that certain groups of environmental factors are more important for certain geographic regions than for others. Copyright © 2007 John Wiley & Sons, Ltd.