基于方差优化k最近邻法的森林蓄积量估测

森林资源调查是数字森林资源监测的基础,遥感技术可以克服传统方法如抽样调查的局限性,有效地缩短作业时间,提高效率。虽然目前森林蓄积量遥感估测方法很多,但随着样本数量的增加,这些方法无法保证估算的准确性。本研究拟提出一种基于方差速率优化的k最近邻法(k NN),以2017年10月Planet Labs影像为数据源,结合赤峰市旺业甸林场蓄积量实测数据建立反演模型,并与地理加权回归(GWR)模型、随机森林(RF)模型、普通k NN模型和距离加权k NN模型进行对比分析。在建立的森林蓄积量反演模型中,方差优化k NN模型得到最优精度[决定系数(R^2)为0.69,均方根误差(RMSE)为67.6 m^3·hm^-2,相对均方根误差(RRMSE)为32.04%],显著优于其他模型。结果表明,方差优化k NN模型相比其他模型更适用于森林蓄积量遥感估测,森林蓄积量遥感反演空间分布符合实际分布情况,可以满足建立反演模型的需求。同时,由于Planet Labs影像的鲜明特征(即具有高时间分辨率),该数据的时间序列数据对于森林季节变化有丰富的记录,在反演森林蓄积量方面有着很大潜力。

Forest stock volume estimation based on a variance-optimized kNN model

Forest inventory is an essential for digital forest resouce monitoring.Remote sensing technology can overcome the limitations of traditional methods such as sampling surveys and can effectively minimize operation time and improve efficiency.Although there exist many remote sensing methods for forest stock retrieval,these methods cannot ensure accuracy of estiamtion with an increase in the number of samples.In this study,a k NN method based on variance optimization is proposed.Planet Labs images from October 2017 are used as the data source,and an inversion model is established by combining the measured volume data for the Wangyedian Forest Farm in Chifeng City.The results of the inversion model are compared with those of the geographically weighted regression,random forest,ordinary k NN,and distance-weighted k NN models.The results showed that,among the established forest stock inversion models,the proposed variance-optimized k NN model obtained the best accuracy(R^2=0.69,RMSE=67.6 m^3·hm^-2,and RRMSE=32.04%)which was significantly better than that obtained by the other models.The results indicate that,compared with other models,the variance-optimized k NN model is more suitable for forest stock estimation by remote sensing.Spatial distribution of forest stock retrieved by remote sensing conforms to actual distribution and can meet the requirements of establishing an inversion model.Further,owing to the distinctive features of Planet Labs images(i.e.,high temporal resolution),the time series data can obtain abundant record of seasonal forest variations and has a considerable potential to invert forest stocks.