基于机器学习的高光谱湿地植被分类研究

Classification of Hyperspectral Wetland Vegetation Based on Machine Learning

为实现湿地植被的精细分类和高精度制图,为湿地管理部门提供准确的决策依据,以美国加州萨克拉门托—圣华金水域的典型湿地植被为研究对象,以高光谱影像为数据源,结合野外GPS采样点,对典型湿地植被的光谱反射率作一阶导数和二阶导数处理,基于均值置信区间原理筛选特征波段,基于单因素分析法筛选能够明显区分植被类型的植被指数。联合特征波段和植被指数构建特征集,利用机器学习C5.0决策树生成知识规则并提取湿地植被信息。结果表明,基于机器学习C5.0决策树的湿地植被提取总体精度为80.09%,Kappa系数为0.792,与最大似然法比较,总体精度提升10.79%,Kappa系数提升0.105,说明基于机器学习的C5.0决策树法能够实现植被的精细分类,方法切实可行。

In order to achieve fine classification and high - precision mapping of wetland vegetation and provide accurate decision - making basis for the wetland management departments, taking typical wetland vegetation in Sacramento - San Huajin waters of California as the research object, hyperspectral image as data source, combined with field GPS sampling points,the spectral characteristics of typical wetland vegetation were processed by first - order derivative and second - order derivative.The characteristic bands were screened based on the mean confidence interval principle, and the vegetation indexes which could clearly distinguish the vegetation types were screened based on the single - factor analysis method. The feature set was constructed by jointing characteristic bands and vegetation index, and the machine learning C5.0 decision tree was used to generate knowledge rules and extract wetland vegetation information. The results showed that the overall accuracy of wetland vegetation based on machine learning C5.0 decision tree was 80.09%,and Kappa coefficient was 0.792.Compared with maximum likelihood method,the overall accuracy was improved by 10.79%, and Kappa coefficient was increased by 0.105, which indicated that the C5.0 decision tree method based on machine learning realized the fine classification of vegetation,and the method was feasible.