Research On Coastal Area Analysis And Application Based On Interpretation Of Hyperspectral Data

Coastal area is a hotspot of geological survey and ecological research, a zone with rapid economic development, meanwhile also the foreland of modern warfare. Thus, quickly and precisely getting the information of large-scale coastal area, and timely processing and analyzing the information is quite significant for geological archeology, ecological conservation, development of coastal cities and national defense. However, due to the complicated coastal environment, manpower often fails in severe area. Remote sensing technique opens a new world for capturing the coastal information. It has many advantages such as rapidness, reiteration and large-scale measure, which is especially dominant in the severe area. From the long term, remote sensing method is more economic.Coastal area covers not only a large body of water but also costal land and tide lat. The size of small targets, such as reef, beacon and ship, is within 10 m, which requests a high spatial resolution of the remote sensor. While the groundcover is various, in order to identify a certain material and analyze the physical and chemical property of the material, the remote sensor is also asked to have a high spectral resolution. Hyperspectral remote sensing imager can get images with a series of spectra in narrow spectral margin. Hyperspectral image has very high spectral resolution and accepted spatial resolution, so that it becomes the preferred data for coastal area research.For pertinence, the coastal area is divided into background and target for research. The background is researched from the perspective of land and ocean. So land-ocean separation is first studied. According to the two characteristics of image segment, land-ocean segmentation is based on normalized difference water index which appropriately extracts water feature thereby reduces the complexity of algorithm a lot. Another is coastline extraction that is based on the spatial gradient of the whole hyperspectral vector thereby acquires precise and robust edge information. The research of land background focuses on a spectral classification complemented with spatial information. It takes advantage of the connectivity of region instead of extracting the spatial features, which has an obvious promotion in classified precision than spectral classification and has a lower complexity of design and timing than spatial features-complemented method. The research of marine background focuses on bathymetry retrieval based on support vector machine (SVM). The SVM is trained by spectra with bathymetrical information to make its output as bathymetry. The accuracy of bathymetry retrieval can be tuned during training. According to much interference in coastal area and sub-pixel target, the research of target in coastal area includes target recognition based on combined spectral signatures and generalized linear spectral mixture model (SMM). Combined spectral reflectance and spectral derivative, the target recognition algorithm tolerates the data characteristics with different spectral fluctuation, which makes the algorithm more precise and robust. Based on linear SMM, generalized linear SMM considers the interference from atmosphere and ground environment to make the manner of spectral mixing more precise by introducing nonlinear terms.Several groups of simulated and real hyperspectral data are used to verify the algorithms proposed in this thesis, the experimental results shows that the proposed algorithms outperform the traditional ones and have more robustness.