基于多尺度遥感影像纹理特征的森林蓄积量反演
汪康宁，吕杰，李崇贵

西安科技大学测绘科学与技术学院，陕西西安 710054

森林蓄积量;随机森林;多尺度;纹理特征;高分一号卫星

Inversion of growing stock volume using satellite image multiscale texture feature
WANG Kangning, LV Jie, LI Chonggui

Collage of Geomatics, Xi’an University of Science and Technology, Xi’an 710054, Shaanxi, China

growing stock volume; random forest; multi-scale; texture feature; GF-1 satellite

DOI: 10.14067/j.cnki.1673-923x.2017.11.014

备注

以国产“高分一号”卫星（以下简称 GF-1）获取的遥感影像数据与少量研究区样地数据为数据源，构建以光谱信息与多尺度纹理特征为特征变量的森林蓄积量反演模型，探讨不同尺度下提取的纹理特征对森林蓄积量估测模型准确度的影响，通过对特征变量的优选，寻求一种提高森林蓄积量反演模型的准确度的方法。首先，对覆盖研究区域的 GF-1遥感影像进行重采样，得到覆盖研究区域的不同分辨率的影像序列，基于不同窗口大小的灰度共生矩阵提取影像序列的纹理特征，与遥感影像光谱信息共同作为特征变量；然后，使用随机森林 (random forest，RF)算法构建森林蓄积量反演模型，对研究区域的森林蓄积量进行估测，分析不同特征变量与窗口大小对森林蓄积量反演模型准确度的影响；最后，通过比较特征变量重要性，确定森林蓄积量反演模型的最佳特征变量与窗口大小选择，对研究区进行森林蓄积量反演，得到研究区域的森林蓄积量分布图。当使用从 8m分辨率遥感影像提取的纹理特征与光谱信息作为特征变量时，森林蓄积量反演模型准确度明显优于使用其他特征变量。其中，当灰度共生矩阵窗口大小设置为 9×9时，森林蓄积量反演模型准确度最高，为 R2=0.70， RMSE=6.317。在根据重要性对从多尺度遥感影像提取的纹理特征进行选择后，所构建的森林蓄积量反演模型的准确度为 R2=0.74，RMSE=6.439。使用较高分辨率遥感影像提取的纹理特征作为特征变量，可以有效的提升森林蓄积量反演模型的准确度。将基于不同分辨率遥感影像提取到的纹理特征作为特征变量，其模型准确度优于使用单一分辨率遥感影像所提取的纹理特征。

Forest growing stock volume (GSV) is one of most important parameters to evaluate forest growth status. In this study, forest inventory data and Gaofen-1 images (GF-1) were used as data sources. The purposes of this study are to explore the effect of multi-scale texture feature on the estimation of GSV, and to improve the accuracy of GSV estimation model. Firstly, a group of different resolution image sequences over study area were obtained by resampling original satellite image. The texture feature of image sequences were extracted using different window sizes Gray-level Co-occurrence Matrix (GLCM). Secondly, a GSV estimation model was structured using Random Forest (RF) algorithm. The GSV of study area was estimated using this model. Finally, the best texture feature and window size were selected by compare importance of feature variables. The distribution of regional forest GSV was performed by the best estimation model. When using texture feature extracted from 8 m resolution image as feature variable, the accuracy of GSV estimation model was better than using texture from low resolution image significantly. With the GLCM window size was set to 9×9, the estimation model achieved the best accuracy(R2=0.70, RMSE=6.317). According to importance of variable, the best feature variables selected from multi-scale satellite to construct GSV estimate model(R2=0.74, RMSE=6.439). The results shown that using texture feature extracted from high resolution satellite image can improve the accuracy of GSV estimation model effective. When multi-scales texture features were used as characteristic variables, the model achieves higher accuracy than only using single scale.