基于光谱聚类的高分影像复杂地块特征提取方法研究

高空间分辨率遥感技术为大范围判识农用地利用类型提供了丰富的数据源。农用地类型多样性和复杂性给高效应用高分影像识别农用地类型带来很大挑战。地块矢量的引入可以帮助更好综合利用多元影像特征，进而提高农用地类型判识精度。但是，传统地块特征提取方法将地块视为一个整体，通过对地块内部像元特征平均实现地块特征表达，不能很好适用于地块内部像元光谱具有较强异质性的情况。针对内部光谱异质但具有较强规律的地块，提出一种基于光谱聚类的特征提取方法，将地块内部的光谱聚类特征作为地块的特征之一，对地块进行分类。利用上海崇明区内2个典型样区的BJ-2卫星影像和地面调查数据进行实验验证分析，结果表明：①该方法相对利用地块内部所有像元光谱平均的方法，能够有效提升地块分类精度；②通过引入地块内部光谱聚类特征到传统地块特征组合，可以进一步提升地块分类精度，对菜地和廊道等内部像元光谱混合比例变化较大的类别提升效果最为明显。该方法为复杂地块分类提供了新思路。

Spectral Clustering based Feature Extraction for Parcel Classification Using High Spatial Resolution Remote Sensing Images

High spatial resolution satellite remote sensing provides redundant data for large-scale agricultural land management. Due to the variety and complexity of land parcels， it is still a great challenge to accurately and timely map land use types. Introducing parcel boundary has proven an effective strategy to integrate spatial and spectral features， to improve the classification accuracy. However， current feature extraction methods always treat each parcel as a whole and use only mean feature values of all pixels inside each parcel. This approach cannot well adapt to the scenarios that parcels include more than one types of spectral similar target. To this end， this paper proposes a spectral clustering based feature extraction method to better model the complexity of parcels. BJ-2 images and ground surveying data from 2 typical areas in the Chongming County in Shanghai were selected to experimentally evaluate the proposed method. The results show that： （1） Compared with the direct spectral averaging method， the proposed method can effectively improve the accuracy of land parcel classification； （2） Introducing the clustering features into the typical feature combination can further improve the accuracy of land parcel classification. And the improvement mainly lies on categories with unstable mixing ratio of internal pixel spectrum， such as vegetable field and corridor. The proposed method provides an effective alternative to classify parcel types especially for parcels including more than one spectral similar target.