基于Landsat 8遥感影像的地上生物量模型反演研究

以吉林省延边朝鲜族自治州汪清县的主要针叶纯林树种为研究对象，结合Landsat 8 OLI数据和地面调查数据，通过提取半径为15 m圆形样地林分尺度下的遥感特征变量实现对地上生物量的估算。首先提取128块样地内的34个遥感特征，其次采用随机森林特征重要性分析遥感特征的贡献率，再利用BP神经网络算法的2种训练算法、SVM支持向量机的3种核函数构建地上生物量模型，最后利用32个测试样本评价模型的估算精度。结果表明，BP神经网络的L-M训练算法和贝叶斯正则化训练算法的R²分别为0.602 9、0.672 1，RMSE分别为5.096 9、4.263 7，MAE分别为4.166 9、3.211 8；SVM支持向量机的线性核函数、RBF核函数、多项式核函数的R²分别为0.585 8、0.561 9、0.487 7，RMSE分别为5.859 4、5.600 9、5.763 7，MAE分别为4.24、3.89、4.176。以贝叶斯正则化训练算法构建地上生物量模型的估测精度最佳；BP神经网络算法比SVM向量机更适用于本研究；同一种机器学习算法不同的训练函数存在差异性。

Aboveground Biomass Model Based on Landsat 8 Remote Sensing Images

Taking the main pure coniferous forest species occurring in Wangqing County,Yanbian Korean Autonomous Prefecture of Jilin Province as the research objects,combining with Landsat 8 OLI data and ground survey data,the above ground biomass was estimated with remote sensing characteristic variables which were extracted from the radius of with the scale of 15 m round sample land stand.Firstly,34 remote sensing features in 128 sample plots were extracted.Secondly,the contribution rate of remote sensing features was analyzed by using the importance of random forest characteristics.Then,two training algorithms of BP neural network algorithm and three kernel functions of SVM support vector machine were used to construct the aboveground biomass model.Finally,32 test samples were used to evaluate the estimation accuracy of the model.The R² values of L-M training algorithm and Bayesian regularization training algorithm of BP neural network were 0.602 9 and 0.672 1,respectively,the RMSE values were 5.096 9 and 4.263 7,respectively,and the MAE values were 4.166 9 and 3.211 8,respectively.The R² values of linear kernel function,RBF kernel function and polynomial kernel function of SVM were 0.585 8,0.561 9 and 0.487 7 respectively,RMSE values were 5.859 4,5.600 9 and 5.763 7,respectively,and MAE values were 4.24,3.89 and 4.176,respectively.The estimation accuracy of aboveground biomass model constructed by Bayesian regularization training algorithm is the best; BP neural network algorithm is more suitable for this study than SVM; different training functions of the same machine learning algorithm are different.