基于微波遥感极化目标分解的土地覆盖/土地利用分类

为有效利用微波遥感影像进行土地覆盖/土地利用分类,该研究以内蒙古河套灌区解放闸灌域为研究区域,采用春耕后试验区Radarsat-2全极化数据,利用极化目标分解方法提取得到了散射熵、平均散射角、反熵、平均特征值、单次反射特征值相对差异度、二次反射特征值相对差异度。结合实地数据,分析了各参数对于耕地、裸地、含植被水体、建筑等类别的可分离性。根据分析结果选取平均散射角、平均特征值、单次反射特征值相对差异度为分类特征变量,通过最小距离法计算了决策边界,最后结合树分类器对试验区影像进行了分类。整体分类精度93.89%,分类Kappa系数为0.914。结果表明,利用平均散射角可有效区分表面散射与二次散射及体散射;平均特征值可有效区分含植被水体与建筑物;单次反射特征值相对差异度参数可有效区分耕地与裸地。利用极化目标分解方法结合决策树分类器可精确地进行土地覆盖/土地利用分类。

Land cover/land use classification based on polarimetric target decomposition of microwave remote sensing

Abstract: The objective of this study is to combine polarimetric target decomposition and decision tree classifier for land cover/land use classification. Taking the Jiefangzha irrigation sub-district of Inner Mongolia Hetao Irrigation District as study area, based on the data of full polarization Radarsat-2 in study area after spring, the entropy, average scattering angle, anti entropy, average eigenvalue, characteristic value of relative difference of single reflection, and characteristic value of relative difference of secondary reflection were obtained by using polarization target decomposition method. Combined with the field data in the time of image acquisition, the separability of parameters on building area, bare land, cultivated land, water area containing vegetation was analyzed. The ground sample's mean values of the above parameters were calculated. By analyzing the mean of these parameters, the results show that the average scattering angle, the average eigenvalue, and characteristic value of relative difference of single reflection can be used for the characteristic quantities of classification. The decision tree decision boundary is determined by the minimum distance method. If the average scattering angle is greater than 36.61°, the area is divided into building area and water area containing vegetation, if not, is divided to bare land and cultivated land. In the building area and water area containing vegetation, if the average eigenvalue is greater than 0.18, the pixel is classified as building area, if not, the pixel is classified as water area containing vegetation. In the bare land and cultivated land, if the characteristic value of relative difference of single reflection is greater than 0.89, the pixels is divided into cultivated land, or else bare land area. The overall accuracy of image classification by decision tree is 93.89% and Kappa coefficient is 0.914. The results show that the average scattering angle can be used to accurately distinguish the single scattering, volume scattering and secondary scattering. Because secondary scattering always appeared in building area and vegetation area, the average scattering angle can be used to extract building area or vegetation area. The echo power of vegetation area is weak, so the average characteristic value associated with the echo power can be used to distinguish the building area and water area containing vegetation. Characteristic value of relative difference of single reflection is associated with terrain roughness, and it can be used to distinguish cultivated land and bare land area. Wrong classification of pixel occurred mainly between the cultivated land and bare land area and between the bare land and building area. Part of bare land surface has greater roughness, which is a major cause of confusion with cultivated land. The cause of wrong classification between bare land and building area is that part of bare land area has greater roughness resulting in secondary scattering. In addition, through the study it was found that the average scattering angle for the extraction of building area has a better effect, but if secondary scattering is produced, it will lead to the confusion between vegetation area and building area; it is easily confused between water area and building area if the water contains vegetation. According to the results, the methods of polarimetric target decomposition can fully explain the physical mechanism of object, and thus it can improve the land cover/land use classification accuracy.