森林干扰在全球和区域碳平衡、气候变化、植被生产力、蒸散发等多方面都有着重要的影响.遥感技术以其在动态监测中经济便捷的优势而成为大尺度森林干扰监测的主要手段.本文以大兴安岭为研究区域,利用2006年1km分辨率的MODIS反射率、LST和NDVI数据,有效提取归一化森林干扰变化信息.针对本研究扰动象元点与非扰动象元点存在较大差异的不平衡性问题,对比了SVM,one class SVM(OCSVM),和bootstrapping SVM分类器在不平衡分类中的效果,结果表明,bootstrapping SVM能够获得更稳定的模型和更高的精度,总体精度达99.14%,kappa系数为0.87,说明基于MODIS粗分辨率数据和bootstrapping SVM算法可以克服不平衡分类问题、有效提取森林干扰区域,可作为一种经济可行的对大区域甚至全球森林干扰监测的方法.
英文摘要:
Forest disturbances play significant roles in carbon balance and global climate changes. Due to its advantages of macro-scale and cost-effectiveness, time-series MODIS data are a striking data source for monitoring forest cover and forest loss.With lkm resolution MODIS data in 2006,this study extracted feature metrics which capturing the salient features of phonological variations to reveal the forest disturbances in the Great Khingan,the largest forestry area in China. Due to the notably imbalanced " change " and " no-change "pixels,this study compared Support Vector Machine (SVM),one class SVM and bootstrapping SVM in forest disturbance detection. The results showed that bootstrapping SVM produced the best classification performance with its overall accuracy and kappa coefficient being 99.14% and 0.87, respectively. A bootstrapping SVM model,therefore,can be used as an effective tool for monitoring forest disturbances in large areas even for the global scale when the MODIS data is used.