Application of neural networks to waterline extraction in tidal flat from optic satellite images

The waterline extraction is one of the most effective satellite remote sensing tools for studying tidal flat environments and its changes, but has not been investigated in detail. A series of field surveys have been carried out to obtain grain size, moisture contents, field spectrometer measurement, and waterline tracking simultaneously with satellite observation. A neural networks algorithm was developed for extracting waterline in tidal flat from satellite-based remote. sensing data and applied to the Gomso tidal flat, Korea. Characteristics of tidal flats and spectral reflectance associated with waterline were analyzed first. We have chosen three bands as input data of neural networks: NIR reflects the amount of suspended sediment content at lower tidal flats; SWIR is sensitive to moisture content and is seriously affected by remaining surface water in sedimentary structures; and TIR appears to be the best among the three bands but its low spatial resolution reduces its utility. The neural network method developed here is independent of tidal situations and robust. The neural networks not only distinguish between tidal flats and seawater out of the images, but it also provides continuous outputs that represent mixed compositions of both features. The values of 0.3 - 0.4 were turned out to be waterline in neural networks output The neural networks output provided the closest to the ground truth and that of the ETM TIR band.