基于优化BP神经网络的香格里拉高山松蓄积量模型研究

Models of Pinus densata Stock Volume in Shangri-La County Based on Optimized BP Neural Networks

以香格里拉县高山松为研究对象,以Landsat TM 8影像和DEM(30M)数据为信息源,结合森林资源二类调查数据和地面样地实测数据,借助MATLAB平台,在前期进行基于遗传算法(GA)和粒子群算法(PSO)优化BP神经网络模型基础上,采用决定系数(R2)、均方根误差(RMSE)及预测精度(P)3个指标对优化后的BP神经网络模型及进行评价,并建立了研究区高山松蓄积量估测模型。结果表明,遗传算法效率(耗时1.9h)低于粒子群算法(耗时1.4h);采用遗传算法优化后的BP神经网络模型R2、RMSE及P分别为0.636、4.216m3、81.748%,均优于粒子群算法。通过遗传算法优化后的BP神经网络模型估测香格里拉高山松蓄积量总量为13 317 879.7m3。

Taking the stock volume o{ Pinus densata in Shangri-La County, Yunnan as the research target, with the data of forest management inventory, Landsat TM8 images, DEM （resolution： 30 meters） and the ground sample data, we estimated the stock volumes o{ P. densata using BP neural networks which was optimized by genetic algorithms （GA） and particle swarm optimization algorithms （PSO） in MAT LAB. Three indices, such as the coe{{icient o{ determination （R2） , the root mean squared error （RMSE） and predict accuracy of model （P） were selected to evaluate optimized BP neural networks. The stock volume predict model was established. The results showed that the PSO which costed 1.4 hours was more effective than GA which costed 1.9 hours; but the R2 , RMSE and P of the model which was optimized by GA were 0. 636, 4. 126 m^3 and 81. 748%, respectively, better than the model＇s which was optimized by PSO. The stock volume of P. densata which was estimated by GA was 13 317 879.7 m^3.