基于遥感光谱和空间变量随机森林的黄河三角洲刺槐林健康等级分类

Forest Healthy Classification of Robinia Pseudoacaciain the Yellow River Delta,China based on Spectral and Spatial Remote Sensing Variables Using Random Forest

对刺槐林健康状况进行准确分类制图,是进行刺槐林健康状况评估与生态修复的前提。以高分辨率IKONOS影像、基于影像提取的不同窗口、不同灰度共生矩阵纹理信息以及反映局部空间自相关的Local Getis-Ord Gi(Getis统计量)为数据源,结合实测生态样方数据,利用多决策树的组合分类模型随机森林(RF)对刺槐林健康进行分级,对6种方法的分类精度进行了比较且对分类变量的重要性进行了排序。结果显示:19m×19m是最佳纹理计算窗口;灰度共生矩阵均值是最优纹理变量;基于波段4计算的Getis统计量对RF分类具有最重要的作用;较之利用全部光谱、纹理和Getis统计量的80个波段/变量,利用前向选择得到的前16个重要性变量进行RF分类,获得了最高的分类精度(总精度为93.14%,Kappa系数为0.894)。研究证实了从高分影像提取的空间特征信息有助于提高对具有规则分布格局的人工刺槐林健康等级的分类精度;前向选择方法可以利用较少的预测变量获得较高的分类精度。

Accurate mapping health levels of Robinia pseudoacacia forests is the premise of forest health as- sessment and ecological restoration.Based on Multiple spectral bands,gray-level co-occurrence matrix （GL- CM） texture information, and Local Getis-Ord Gi statistical information extracted from High resolution IKONOS imagery, random forest classification method was used to identify and map health levels of Robin- ia pseudoacacia forests.Random forest method was applied to six combinations of the spectral, textural, and spatial features, and the contribution of predictive variables was ranked.The experimental results indi- cated that 19 m ×19 m is the best moving window size for GLCM texture information extraetion,GLCM mean calculated from IKONOS Pan band which is the best textural feature local Getis-Ord Gi statistical information calculated from IKONOS band4 which has the most important role.Compared with the combi- nation of all spectral,textural,and spatial features （total of 80 features）,the best model based on random forests with a forward variable selection process which selected onlyl6 variables of the original 80 variables and obtain the best predictive accuracy （overall accuracy of 93.14% ,kappa coefficient of 0.894）.Our results indicate that the spatial features extracted from high resolution imagery can improve classification accuracy of health levels of planted forests with a regular spatial pattern.Forecast sorting can utilize fewer predictors and obtain higher classification accuracy.