欢迎访问林业科学,今天是

林业科学 ›› 2019, Vol. 55 ›› Issue (9): 92-102.doi: 10.11707/j.1001-7488.20190910

• 论文与研究报告 • 上一篇    下一篇

高空间分辨率影像与SAR数据协同特征面向对象林分类型识别

毛学刚, 竹亮, 刘怡彤, 姚瑶, 范文义   

  1. 东北林业大学林学院 哈尔滨 150040
  • 收稿日期:2017-04-17 修回日期:2017-11-14 发布日期:2019-10-28
  • 基金资助:
    国家重点研发计划(2017YFB0502705);中央高校基本科研业务费专项资金(2572018BA02)。

Object-Oriented Classification for Tree Species Based on High Spatial Resolution Images and Spaceborne Polarimetric SAR Cooperation with Feature

Mao Xuegang, Zhu Liang, Liu Yitong, Yao Yao, Fan Wenyi   

  1. School of Forestry, Northeast Forestry University Harbin 150040
  • Received:2017-04-17 Revised:2017-11-14 Published:2019-10-28

摘要: [目的]研究对象特征对高空间分辨率遥感影像与星载全极化SAR数据协同面向对象林分类型识别的影响,评价2种数据协同林分类型识别的适宜性,为多源遥感影像结合面向对象分类技术提供科学依据。[方法]以QuickBird遥感影像和Radarsat-2数据为试验数据,选取福建省三明市将乐县将乐国有林场为试验区进行杉木、马尾松和阔叶林面向对象分类。在面向对象分类过程中,采用基于QuickBird多光谱波段分割、基于Radarsat-2数据分割和QuickBird & Radarsat-2协同分割3种分割方案,每种分割方案采用10种尺度(25~250,步长为25),应用QuickBird遥感影像和Radarsat-2数据提取的光谱、地形、高度和强度4方面32个特征指标,进行4种不同特征组合,运用支持向量机分类器进行面向对象林分类型分类,利用混淆矩阵计算的生产者精度、用户精度、总精度和Kappa系数4个指标对分类结果进行精度评价。[结果]所有组合的分类精度(Kappa系数)均随着尺度增大表现出先增加后降低的趋势,且以只使用单一光谱特征的分类精度最低,依次低于光谱+地形两特征和光谱+地形+高度三特征的分类精度,引入强度后的四特征组合分类与三特征组合无明显差异。QuickBird&Radarsat-2协同且在最优尺度参数为100时,结合对象光谱、地形、高度和强度四特征组合进行面向对象林分类型分类精度最高(OA=86%,Kappa=0.86)。[结论]高空间分辨率遥感影像(QuickBird)与SAR数据(Radarsat-2)协同最优尺度多特征组合进行面向对象林分类型分类优势明显,在光谱和地形特征中引入高度特征可进一步提高分类精度。本研究结果可提高面向对象分类中的特征选择效率和科学性,能够为其他影像的面向对象分类技术提供较好的参考依据。

关键词: 面向对象, 对象特征, SAR, QuickBird, 支持向量机

Abstract: [Objective] To clarify the effects of object features on object-oriented classification for tree species based on high spatial resolution remote sensing images and spaceborne polarimetric SAR data, and further evaluate the suitability using collaboration of two kinds of data to process classification for tree species, this study was conducted on Jiangle state-owned forest farm in Jiangle county, Sanming city, Fujian Province.[Method] In this study, QuickBird remote sensing image and Radarsat-2 data were used as the experimental data to carry out object oriented classification for Chinese fir, pine and broad leaved forest predominated in the study area. Three segmentation schemes(based on multispectral band of QuickBird remote sensing image segmentation, Radarsat-2 data segmentation, collaboration of QuickBird remote sensing image and Radarsat-2 segmentation)were adopted when processing object oriented classification and each segmentation scheme was divided to 10 scales(25-250, step size 25). For each scales of three segmentation schemes, 32 kinds of classification features from four aspects of spectral features, topographic features, height and strength features extracted from QuickBird remote sensing image and Radarsat-2 data was used. According to 3 segmentation schemes, 10 classification scales and 4 aspects of objects' features, 120 kinds of classification results for tree species were obtained by using the support vector machine classifier. Later the accuracy of each classification result was evaluated by 4 different evaluation indexes, namely, producer accuracy calculated by confusion matrix, user accuracy, classification accuracy and Kappa coefficient.[Result] The results showed that:the classification accuracy(Kappa coefficient)of all classification schemes showed a similar trend, or increased firstly and then decreased with scale increasing, each combination of classification scheme had an optimal scale. The highest Kappa coefficient based on spectral bands of QuickBird segmentation-classification scheme was 0.84 at an optimal scale of 150; the highest Kappa coefficient based on Radarsat-2 data segmentation-classification scheme was 0.68 at an optimal scale of 125; the highest Kappa coefficient of cooperative segmentation of QuickBird & Radarsat-2 classification scheme was 0.86 at a scale of 100. In all classification schemes, the classification accuracy that using spectral features individually was the lowest especially on small scales; which lowered than those using two collaboration features of spectral + topography, using three collaboration features of spectral + topography + height, using four collaboration features of spectral + topography + height + strength, respectively. The latter two classification schemes had no significant difference. It indicated that introducing more objected features than one spectral feature could improve classification efficiency. Based on cooperative segmentation of QuickBird & Radarsat-2 at its optimal scale of 100, object-oriented classification for tree species obtained the highest classification accuracy(OA=86%, Kappa=0.86)when simultaneously introducing spectral, topographic, height and strength features.[Conclusion] Classification for tree species by collaboration of high spatial resolution remote sensing images and spaceborne polarimetric SAR data combining with optimal scale and multi-objected features had manifest advantages. Combing height information with spectral and topographic features could further increase classification accuracy. This study could improve the efficiency of features selection and accuracy of object-oriented classification. Moreover, it could provide good reference basis for other images based on object-oriented classification technology.

Key words: object-based, object features, SAR, QuickBird, SVM

中图分类号: