应用PLS-GA-SVM构建的云南松林分蓄积量遥感估算模型

Single Seasonal Remote Sensing Model of Pinus yunnanensis Forest Volume by Using PLS-GA-SVM

为提高森林蓄积量遥感预测精度,结合机器学习算法构建单季节林分蓄积量估算模型。以云南省大理州2007年国家森林资源连续清查云南松林样地蓄积量为对象,根据冬季、春季和秋季遥感影像提取研究区的单波段、植被指数和纹理信息共74个因子,采用PLS提取前13个主成分作为自变量,经过GA优化c、g参数的SVM构建云南松林分蓄积量估算模型,探讨单季节SVM的训练效果和泛化能力。研究表明:各季节所有自变量与云南松林分蓄积量相关性较弱;春季遥感数据与秋冬季节遥感数据存在差异;高值低估现象普遍存在,冬季遥感影像构建的PLS-GA-SVM模型效果最好(训练集R^2=0.6690,ERMS=6.7345 m^3),泛化能力最佳;春季遥感数据复杂性较高,无法准确反映预蓄积量变化情况。

In order to improve the accuracy of remote sensing prediction of forest volume,a single season stand volume estimation model was constructed by combining machine learning.The 74 factors,including single band,vegetation index and texture information were extracted from the remote sensing images of winter,spring and autumn in Dali of Yunnan Province in 2007.The first 13 principal components were extracted by PLS as argument,and the estimation model of Pinus yunnanensis forest volume was built by SVM with GA optimized c and g parameters.The results showed that the correlation between all the argument of each season and the forest volume of P.yunnanensis was weak.There were differences between spring remote data and autumn winter remote data.High volume undervaluation was common,but the SVM model built by remote sensing image in winter has the best effect(training R^2=0.6690,RMSE=6.7345 m^3)and generalization ability.The complexity of spring remote data was high,which cannot accurately reflect the change of forest volume.