基于多距离度量kNN模型的森林蓄积量反演

Forest stock volume inversion based on multi-distance metric kNN models

【目的】森林蓄积量是衡量森林质量和生长状况的重要指标。利用遥感技术进行森林蓄积量反演相比传统的森林调查能显著提高森林资源调查效率,对快速获取区域范围森林生长状况,进行高效的资源利用和森林经营管理具有重要意义。【方法】以陕西韩城市为研究区,利用森林资源二类调查数据库提取森林蓄积量实测数据,结合Sentinel-2遥感影像进行森林蓄积量反演。通过线性逐步回归法和重要性评价法分别进行变量筛选,构建多元线性回归模型、支持向量机模型、随机森林模型和基于欧式距离、曼哈顿距离和马氏距离构建的kNN模型进行森林蓄积量估测,通过精度评价比较最终选择估测精度最高的模型进行研究区森林蓄积量反演。【结果】1)马氏距离是最适合构建kNN模型的距离度量。基于马氏距离构建的kNN模型在所有模型中实现了最高的估测精度,决定系数R2为0.66,均方根误差RMSE为10.02 m^(3)/hm^(2),均方根误差相比随机森林模型、支持向量机模型和多元线性回归分别下降了3.9%、7.8%和29.9%;2)非参数模型在森林蓄积量估测中的精度显著优于参数模型。基于马氏距离构建的kNN模型、随机森林模型、支持向量机模型均方根误差相比多元线性回归分别降低了29.9%、27.0%和23.9%;3)研究区西北部森林生长情况较好,蓄积量值较大,东部和南部地区主要是水域和建筑用地,森林分布较少,森林蓄积量值较低。【结论】利用kNN模型结合Sentinel-2遥感影像能实现森林蓄积量反演和制图,为森林资源遥感估测研究提供参考。

【Objective】Forest stock volume is an important indicator to measure the quality and growth status of forests.Compared with traditional forest surveys,forest stock volume inversion using remote sensing technology can significantly improve the efficiency of forest resource surveys,which is important for quickly obtaining the forest growth status on a regional scale and conducting efficient resource utilization as well as forest management.【Method】The city of Hancheng in Shanxi Province was selected as the study area,and the forest management survey database was used to extract the actual measured data of forest accumulation combined with Sentinel-2 remote sensing images for forest stock volume inversion.Through the linear stepwise regression method and importance evaluation method,the variables were screened,and the multiple linear regression model(MLR),support vector machine model(SVM),random forest(RF)model and the kNN model based on Euclidean distance,Manhattan distance and Mahalanobis distance were constructed to evaluate the accuracy.Finally,the model with the highest estimation accuracy was selected for the forest stock volume inversion in the study area.【Result】1)The Mahalanobis distance was the most suitable distance metric for constructing the kNN model.The kNN model constructed based on the Mahalanobis distance achieved the highest estimation accuracy among all models,with a coefficient of determination(R2)of 0.66 and a root mean square error(RMSE)of 10.02 m^(3)/hm^(2),and the RMSE decreased by 3.9%,7.8%and 29.9%,respectively,compared with the random forest(RF),support vector machine(SVM)and multiple linear(MLR)models.2)The nonparametric model significantly outperformed the parametric model in the accuracy of forest stock volume estimation.The RMSEs of the kNN,RF and SVM models decreased by 29.9%,27.0%and 23.9%,respectively,compared with the MLR regression.3)The northwestern part of the study area had better forest growth and larger forest stock volume values,while the eastern and southern areas were mainly watersheds and construction lands,with less forest distribution and lower values of forest stock volume.【Conclusion】The kNN model combined with Sentinel-2 remote sensing images can realize forest stock inversion and mapping,which can provide references for forest resource remote sensing estimation.