Automatic Extraction of Hydrographic Objects in Digital Orthophoto Images

This paper presents a new approach to extracting hydrographic objects such as rivers, lakes, and other water bodies from digital orthophoto images. A family of directional-sensitive operators is derived and applied to suppress the noise while preserving the desired features. The unique property of this operator family is that boundaries and junctions are preserved with high accuracy while noise within each object is greatly suppressed. A region-based grouping technique, which was derived from locally excitatory globally inhibitory network (LEGION) dynamics, is then applied to extract desired objects, the seeds of which are selected separately. To find hydrographic objects, seed points are automatically identified from the original image, based on the assumption that water bodies are homogenous. This approach is general and can be used to find other objects when appropriate seed selections are incorporated. Computationally, this approach is parallel and local in nature and can be efficiently implemented using a neural network. Experimental results using real images are provided.