基于SPAN与NDVI的全极化SAR数据喀斯特地区土地类型划分

Land type classification of full polarization SAR data using SPAN and NDVI in Karst Areas, China

地貌复杂性、地物多样性等特征使得全极化合成孔径雷达(Synthetic Aperture Radar,SAR)数据的散射机制和散射强度相互交织,从而导致基于传统Wishart-H/α的全极化SAR数据难以实现喀斯特地区土地类型的有效划分。针对此问题,该研究首先用复Wishart距离测度对研究区土地类型样本进行聚类,同时利用H/α平面对研究区进行超盒聚类,然后根据超盒聚类结果平均相干矩阵与样本聚类结果平均相干矩阵间的复Wishart距离进行半监督分类,获得研究区土地类型划分的初步结果。在此基础上利用对建筑物与裸岩地敏感的极化总功率(Polarimetric-Total-Power,SPAN)和对林地、草地与耕地敏感的归一化植被指数(Normalized Differential Vegetation Index,NDVI)对初步结果继续进行划分,最终将研究区土地类型划分为水体、林地、草地、耕地、建筑地和裸岩地,总体分类精度为81.45%;采用另一地势相对平缓、地形相对单一的典型喀斯特地区全极化SAR数据进行验证,在实现该地区土地类型划分的同时总体分类精度为85.66%。这说明该研究方法能够实现喀斯特地区土地类型的准确划分。

Topographical complexity of landscapes and diversity of terrain features have caused to interweave with the scattering mechanism and intensity of full polarization Synthetic Aperture Radar(SAR)data.This makes it difficult for the traditional Wishart-H/α(polarization entropy/scattering angle)classification to effectively classify land types in Karst areas.In this study,a complex Wishart distance measure was used for the class clustering of land type in research areas.A super-box clustering was carried out using the H/αplane.The semi-supervised classification was also carried out,according to the complex Wishart distance between the average coherence matrix of super-box and sample clustering.The obtained data of land classification were further processed using the Polarimetric-Total-Power(SPAN)and the Normalized Differential Vegetation Index(NDVI),where the SPAN was sensitive to buildings and bare rock land,whereas,the NDVI was sensitive to wood land,grass land and cultivated land.An effective classification of land types was finally realized in the Karst areas.Specifically,the complex Wishart distance between two types of sample was calculated to determine the similarity of scattering characteristics in the samples of different land types,and the clustering was also performed during this time.Eight kinds of super-box clustering were divided into using the H/αplane.Three kinds of preliminary clustering were obtained,including the first type of water body,the second type of construction and bare rock land,as well as the third type of wood land,grass land and cultivated land.The SPAN was then used to classify the building and bare rock land,using a threshold in the way of super-box segmentation.The wood land,the grass land,and the cultivated land were classified by introducing a combination of the NDVI and Digital Elevation Matrix(DEM).The improved method can be used to effectively classify the water body,woodland,grassland,farmland,construction land,and bare rock land,with the overall accuracy of 81.45%.To verify the improved method,the another full-polarization SAR data was selected from the typical Karst areas,where the terrain was relatively flat,while the topography was relatively single.This case study demonstrated that the land classification was successfully implemented,where the overall classification accuracy reached 85.66%.The finding can provide a novel way to accurately classify various land types,and thereby serve as an ideal supplementary means to monitor rocky desertification evolution in Karst areas.