面向海岸线精确提取的高光谱影像流形学习研究

Accurate extraction of coastline directly affects the accuracies in the results of its long-term change detection, and even has strategic significance in both healthy development of coastal resources and sustainable development of coastal economy. Compared with other optical remote sensing technologies, hyperspectral imaging has high spectral resolutions and is helpful to differentiate main ground objects in the coastal zone. However, within the intertidal region, changes in spectral responses of main ground objects always occur because of the negative influences from some external factors, and hence the “different ground objects with similar spectrum” phenomenon seriously exists among all main ground objects. .Therefore, our study introduce manifold learning to extract low-dimensional manifold features from the coastal hyperspectral imagery, aim to expand the subtle spectral differences among main ground objects and finally to accurately extract the coastline. First, considering the geographical characteristics of the coastal zone and features of hyperspectral imaging, we propose new feature extraction methods using manifold learning for coastal hyperspectral imagery. Second, we study the interpretation symbols for each coastal shoreline and their recognition hierarchies in the three-level classification system of coastal zone, and then propose homologous strategies using the improved manifold learning method for each kind of interpretation symbol in order to make accurate extraction of coastline. Finally, using coastal zone in Xiang-shan Bay of Zhejiang province as our experimental area, we compares the extracted result of coastline using our proposed method with the coastline from the sea map, in order to completely verify the accuracy and reliability of our proposed method. Our research results from the proposal could promote the theory in feature extraction of hyperspectral imagery using manifold learning, and also could provide new technical supports for accurately extracting the coastline in the large-scale spatial space.

海岸线的精确提取直接影响海岸线的长期变化监测结果的准确性，更对海岸带资源的健康开发和沿海经济的可持续发展具有战略意义。相比其它光学遥感技术，高光谱影像特有高光谱分辨率有利于准确识别海岸带中主要地物。然而在潮间带中，各地物的光谱特性由于外界因素影响发生改变，存在严重的“异物同谱”现象。因此，本研究引入流形学习来提取海岸带高光谱影像中的低维流形特征，扩大地物间的细微光谱差异来精确提取海岸线。首先，从海岸带的地域特征和高光谱影像的特性出发，研究海岸带高光谱影像的流形学习的特征提取方法；其次，确定三级海岸带分类体系下海岸线的解译标志及其辨识度层次体系，研究对应的流形学习策略来提取海岸线；最后，以浙江省象山港海岸带为实验区，对比本文提取的海岸线与实测海岸线，验证提出的方法的准确性和可靠性。本课题的研究成果能够完善高光谱影像的流形学习特征提取理论，并为大空间尺度下海岸线的精确提取提供新的技术支持。

海岸线的提取对实现我国海岸带资源的健康开发和保障沿海经济的可持续发展具有重要意义。本研究依托高光谱遥感技术的高分辨率光谱优势，结合海岸带地形地貌以及地物的尺度和空间分布特征，研究非线性流形学习方法用于海岸线的精确提取问题，主要包括高光谱遥感流形学习方法、高光谱遥感特征选择和分类方法、海岸线提取与演化分析应用。高光谱遥感流形学习研究方面，考虑海岸带地物的光谱和空间特征，引入自适应加权综合核距离来构建改进拉普拉斯特征映射模型，提出稀疏低秩近似等距线性映射方法，挖掘海岸带高光谱遥感影像的低维流形结构。海岸带高光谱遥感特征选择和分类研究方面，提出一系列的海岸带高光谱影像特征选择方法，包括改进稀疏子空间聚类方法和差异性加权稀疏自表达方法等，选取适合海岸带地物区分的敏感波段，降低海岸带高光谱影像智能处理计算量的同时为海岸线的精确提取提供丰富的特征信息。针对海岸带两侧不同地物的敏感波段差异，提出波段加权的支持向量机分类模型，提升海岸带地物的识别精度。考虑到海岸带地物分布空间异质性强、破碎度高且地物类别不均衡问题，高光谱遥感影像中混合像元问题严重，提出参数子空间约束的泊松非负矩阵分解方法和鲁棒核原型分析的端元选取方法，实现海岸带地物的亚像元层面的精细提取。实例应用研究方面，利用提取海岸线来分析浙江省1990-2010年的围填海空间格局变化情况，剖析杭州湾南岸2000-2010年海岸带的景观格局的时空演变特征，利用GIS-Logistic的耦合模型来探讨海岸线附近景观演化的内在驱动力。