基于SVM雷达卧龙大熊猫栖息地森林成图

SVM-based forest mapping of Wolong Giant Panda Habitat using SAR data

卧龙自然保护区(世界自然遗产地)是大熊猫最主要的栖息地之一。结合雷达遥感全天时、全天候观测优势,以及森林覆盖对栖息地生境评价的重要性,开展多时相、双极化雷达数据森林精细成图研究就显得尤为重要。本研究首先对雷达数据进行辐射地形校正;然后选用5个时相ALOS PALSAR数据,采用支持向量机(support vector machine,SVM)方法进行森林精细成图。研究选取了5个多时相、双极化典型特征信息参与初始训练和分类,即HH_m,HV_m,TSD,HH_m-HV_m和HH_m/HV_m;接着通过对不同信息组合分类精度的试验与对比,获取了最优特征组合HH_m,HV_m,TSD,HH_m-HV_m。对应分类总体精度、森林及非森林类别用户精度分别为86.90%,82.34%和92.83%,显著优于单时相单极化数据分类结果(分类总体精度55.47%)。研究结果验证了多时相、双极化雷达遥感数据在自然遗产地森林精细成图中的有效性,并揭示了雷达遥感在多云多雨地区生境监测与评价中的潜力与应用价值。

Generally,the conservation of Wolong Giant Panda Habitat( Natural World Heritage site) is significant for the sustainability of the rare species of Giant Panda. As we know,forest coverage can be an essential impact on the suitability of the habitat. Owing to the all-weather,all-day operation capability of radar systems,in this study,the authors investigated the performance of Synthetic Aperture Radar( SAR) images in fine mapping of forests using multi-temporal/polarization PALSAR data. The authors firstly corrected the radiometric distortion of SAR data induced by the cliffy topography; then the authors selected 5 different temporal acquisitions for the forest mapping using the Support Vector Machine( SVM) approach. 5 multi-temporal/dual-polarization indexes,i. e.,HH_m,HV_m,TSD,HH_m-HV_mand HH_m/HV_m,were applied for the training and classification. Experimental results demonstrated that the combination of HH_m,HV_m,TSD and HH_m-HV_mderived an optimal classification( e. g.,total accuracy and user accuracy of forest/non-forest were 86. 90%,82. 34% and 92. 83%,respectively),better than the single-temporal/polarization mode( total classification accuracy of 55. 47%). This study shows the effectiveness of multi-temporal/polarization SAR data in forest fine mapping,particularly in the monitoring and evaluation of natural heritage sites located in cloudy and rainy environments.