基于Sentinel-1A和Landsat 8数据的区域森林生物量反演

Forest biomass retrieval based on Sentinel-1A and Landsat 8 image

【目的】结合主被动遥感数据,为基于不同遥感数据源、建模算法的亚热带森林生物量建模分析提供新思路。【方法】以湖南省郴州市桂东县2014年Landsat 8 OLI影像、2014年Sentine-1A影像、2014年43块森林资源连续清查固定样地数据为主要信息源,借助于ENVI、SNAP、R等软件,分别采用主动式遥感(Sentinel-1A数据)、被动式遥感(Landsat 8 OLI数据)、主被动相结合(Sentinel-1A数据结合Landsat 8 OLI数据)3种数据集和多元线性回归、随机森林、人工神经网络、袋装算法等4种模型,进行区域森林地上生物量特征变量选取、参数建模、模型精度评价、生物量空间制图。【结果】1)在特征变量选择上,红波段(B4)、红外波段(B5)反射率及纹理特征,归一化植被指数(NDVI),交叉极化(VH)后向散射系数及其纹理特征,在森林生物量反演中具有重要作用;2)4种遥感估测模型精度比较分析表明,无论是单一数据源还是二者结合,随机森林算法预测精度最高,人工神经网络、袋装算法次之,多元线性回归最低;3)3种不同数据源的遥感估测综合精度,按照由高到低的顺序排列,主被动结合>被动式遥感>主动式遥感;4)桂东县森林生物量平均值为53.68 t/hm2,生物量高(>90 t/hm2)的林分面积比例只有16.03%,主要分布在海拔较高、坡度较陡的东南、西南部。【结论】Sentinel-1A和Landsat 8数据的结合在估测森林生物量方面具有重要作用。

【Objective】Combine active and passive remote sensing data,provide new ideas for modeling and analysis of subtropical forest biomass based on different remote sensing data sources and modeling algorithms.【Method】In this study,the 2014 Landsat 8 OLI image and the 2014 Sentine-1A image of Guidong county,Chenzhou city,Hunan province and the data of 43 forest resources continuous inventory fixed sample plots in 2014 were taken as the main information sources.With the help of ENVI,SNAP and R software,used the following three methods of active remote sensing(Sentinel-1A data),passive remote sensing(Landsat 8 OLI data)and active-passive combination(Sentinel-1A data combined with Landsat 8 OLI data)and four models of multiple linear regression,random forest,artificial neural network and bagging algorithm to select characteristic variables of regional forest biomass,modeling parameter,model accuracy evaluation and making forest biomass spatial map.【Result】1)In the selection of characteristic variables,reflectance and texture features in red band(B4),infrared band(B5),normalized vegetation index(NDVI),backscattering coefficient and texture characteristics of cross polarization(VH)play an important role in forest biomass retrieval.2)The comparison and analysis of the accuracy of the four remote sensing estimation models show that regardless of the single data source or the combination of the two,the random forest algorithm has the highest prediction accuracy,followed by the artificial neural network and the bagging algorithm,and the multivariate linear regression has the lowest prediction accuracy.3)The comprehensive accuracy of remote sensing estimation from three different data sources is arranged in order from high to low as follows active-passive combination>passive remote sensing>active remote sensing.4)The average forest biomass of Guidong county is 53.68 t/hm2,and the proportion of forest area with high biomass(>90 t/hm2)is only 16.03%.It mainly distributes in the southeast and southwest of Guidong county with high altitude and steep slope.【Conclusion】The combination of Sentinel-1A and Landsat 8 data plays an important role in estimating forest biomass.