基于高分二号的旺业甸林场蓄积量估测模型研究

【目的】为了探究国产高分二号(GF-2)影像在林分蓄积量估测中的潜力,并找到最佳的蓄积量估测模型。【方法】本次实验以内蒙古旺业甸林场为研究区,以高分二号卫星影像为信息源,结合2017年10月份调查的75块样地以及同时期的GF-2影像数据,提取波段特征、植被指数和纹理特征等43个遥感因子作为候选变量,利用Pearson相关系数选择出与蓄积量显著相关的6个变量,采用多元线性回归模型(MLR)、BP-神经网络模型(BP-ANN)、随机森林模型(RF)、支持向量机模型(SVM)和K邻近模型(KNN)进行蓄积量的估测。以决定系数(R^2)、均方根误差(RMSE)、相对均方根误差(RRMSE%)作为5种模型的评价指标,选择出旺业甸林场的最佳蓄积量估测模型,并绘制了研究区的森林蓄积量分布图。【结果】4种机器学习模型的结果明显优于传统的线性模型,其中随机森林(RF)模型和K邻近模型(KNN)均得到了较高的精度,其中RF模型的R^2为0.66,均方根误差为55.2 m^3/hm^2,相对均方根误差为28.1%,KNN模型的R^2为0.64,均方根误差为57.6 m^3/hm^2相对均方根误差为29.3%。【结论】在利用高分二号数据进行旺业甸林场蓄积量估测时,RF和KNN模型在估测针叶林蓄积量时相比于其他模型可以取得更好的结果。

Study on volume estimation model of Wangyedian forest farm based on GF-2

【Objective】In order to explore the potential of GF-2 image in estimating forest stock and find the best model for estimating forest stock.【Method】Taking Wangyedian forest farm in Inner Mongolia as the research area,took the satellite image of GF-2 as the information source,combined with 75 sample plots surveyed in October 2017 and GF-2 image data of the same period,extracted band characteristics,vegetation index and texture characteristics.With 43 remote sensing factors as candidate variables,Pearson correlation coefficient was used to select 6 variables significantly related to stock volume.Multivariate linear regression model(MLR),BP-ANN model,random forest model(RF),support vector machine model(SVM)and K-neighborhood model(KNN)were used to estimate stock volume.Taking the determination coefficient(R²),root mean square error(RMSE)and relative root mean square error(RRMSE%)as the evaluation indexes of the five models,the best stock estimation model of Wangyadian forest farm was selected,and the forest stock distribution map of the study area was drawn.【Result】The results show that the results of four machine learning models are significantly better than those of traditional linear models.Random Forest(RF)model and K-Nearest Neighbor Model(KNN)have higher accuracy.The R²of RF model is 0.66,the root mean square error is 55.2 m3/hm^²,the relative root mean square error is 28.1%,the R²of KNN model is 0.64,and the root mean square error is 57.6 m3/hm^²,the relative root mean square error is 29.3%.【Conclusion】Therefore,RF and KNN models can obtain better results than other models in estimating coniferous forest volume.