Unsupervised Extraction of Drainage Network from Digital Elevation Data: A Self-Adaptive Approach

In this study, the problem of unsupervised extraction of drainage network from elevation data is addressed. A self-adaptive system and its application to the problem are described. The popularity of Geographic Information Systems and the availability of many high-resolution elevation data sets have increased many uses of elevation data. One of the more challenging and yet useful applications is the detection of drainage network and basins from these data sets. The detection can aid us in a more accurate measurement of the drainage network metrics. The manual and/or supervised classification of drainage suffer from long and tedious manual process as well as inter and intra operator error due to human factors. An automated unsupervised approach is required to process the data accurately and in a timely manner. A literature search resulted in identifying several significant attempts in automating the detection process. Jenson [1] used detection of local minimum and operator’s suggested threshold values to identify possible drainage pixels. The algorithm has several shortfalls such as operator dependency (supervised), dependency on data resolution, and feature size. Qian et al. [2] used an expert system, which combines the local operators and global reasoning to locate drainage network. The algorithm has a better performance compared to totally localized algorithms. However, its extraction is still a function of selected feature size, this becomes especially problematic when the feature size varies over the data set. Fern et al. [3] algorithm detects the drainage using connected network of overlapping local processors to overcome this problem and detecting a global network. The algorithm performance varies as a function of the algorithm parameters. The nine algorithm parameters dictated the system performance and their values need to be found experimentally for any given terrain type or the algorithm generates many false positive results. Two major types of elevation data are commonly used in remote sensing, the Digital Terrain Elevation Data (DTED) and the Digital Elevation Models (DEM) which are distributed by National Imagery and Mapping Agency (NIMA) and the U.S. Geological Survey (USGS) at a variety of resolutions. Figure 1 shows one of the DEM test files used in this study, it is produced by USGS and has a cell size of 7.5 X 7.5 min quadrangles, with a spatial resolution of 30m X 30m (1:24 000 scale). The drainage extraction algorithm proposed here is based on a nonlinear closeloop adaptive system. Adaptive systems’ self-organization, fault-tolerances, and generalization make them attractive classifiers. These systems have proven to be