Extracting hydrographic objects from satellite images using a two-layer neural network

This paper presents a two-layer network for extracting hydrographic objects, such as rivers, from satellite images. The first layer is a locally connected network, which performs nonlinear smoothing. A unique property of the network is that the boundaries and junctions are presented with high accuracy while the noise within each region is greatly suppressed. A second layer is a locally excitatory globally inhibitory oscillator network (LEGION), which extracts the desired objects. The seeds of objects 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. Computationally, this approach is parallel and local and can be effectively implemented using hardware directly, the efficiency of which may provide a potential solution for real-time image processing. Experimental results using digital orthophoto images are provided