基于耦合推断的时间外延性辅助数据偏倚修正——以森林蓄积量估计为例

Rejuvenating the Shelf-Life of Outdated Model and Auxiliary Data for Remote Sensing-Assisted Forest Inventory:Taking Forest Volume as an Example

【目的】1)量化“过期”模型对推断总体参数(总体均值、方差)的影响;2)提出利用基于模型辅助的估计量修正模型保质期引起的推断偏倚;3)评估度量误差模型对推断偏倚的修正作用。【方法】在基于设计、基于模型和基于模型辅助3种推断框架下,应用度量误差模型,修正“过期”模型对总体参数(总体均值、方差)的影响。【结果】1)将二阶抽样的估计值作为参照,对比基于模型的统计推断估计值,无论是线性回归模型还是度量误差模型,其均值估计值与二阶抽样下总体均值估计值6.774 m^(3)·hm^(−2)接近,其方差估计值的平均值为0.117,远小于基于设计的方差估计值0.965,精度提升平均为87.93%;2)遥感数据“过期”引起模型失效,对总体均值估计产生较大偏差,总体均值估计值的偏移程度随着遥感数据获取时间的推移加剧;3)在基于模型辅助推断框架下,线性回归模型和度量误差模型的总体均值估计值均在6.5~6.8波动,波动范围较小,变化趋势相同,二者均值估计值差异较小,后者推断精度更高,精度提升范围为5.71%~22.50%,平均为13.34%。【结论】1)遥感数据作为辅助信息可有效提高估计精度;2)当遥感数据“过期”时模型失效,总体均值估计值偏倚增大,方差估计值被低估;3)基于模型辅助的统计推断可解决模型“过期”造成的推断偏倚问题,保持估计量近似无偏性,且精度与概率样本的样本量呈正相关;4)度量误差模型可降低总体方差估计值,但仅使用度量误差模型无法消除时间外延性遥感数据产生的推断偏倚。

【Objective】1)To quantify the effects of the“outdated”models on the population parameter estimates.2)To correct bias due to“outdated”model by model-assisted inference.3)To quantify the correction effects of the errors-in-variable(EIV)model on the inferential bias.【Method】Correcting for the effect of outdated models on estimates(the population mean,and the variance of the population mean)using EIV model under three inference frameworks:design-based inference,model-based inference,and model-assisted inference.【Result】1)Compare the estimates based on model-based inference with the estimates of two-stage sampling.The model-based estimates of the mean,,including the regular model and the EIV model,were similar to each other and to the two-stage sampling estimate of the mean,6.774 m^(3)·hm^(−2).The design-based and the model-based estimates of have great disparity,with two-stage sampling of 0.965 and model-based averaging 0.117,meaning that the average precision for model-based was improved by 87.93%.2)The adverse effects of temporally external independent variables are exacerbated over time with respect to a model and propagate to the population parameter estimation.3)In the model-assisted inference,the estimates of meanμ,including the regular and the EIV models,range from 6.5 to 6.8,with a small fluctuation range and the same trend,and the differences between the mean estimates of the two are small,but the latter has higher inference accuracy,with an accuracy improvement range of 5.71%−22.50%,with an average of 13.34%.【Conclusion】1)Remotely sensed data as auxiliary information can effectively improve the estimation accuracy.2)When remote sensing data are“out of date”,the model is invaild,the mean estimate bias increases,and the variance estimate is underestimated.3)Model-assisted inference solves the problem of inference bias caused by“outdated”model,maintains the approximate unbiasedness of the estimates,and the accuracy is positively correlated with the sample size of the probability sample.4)EIV model reduces the variance estimates,but the adverse effects of the“outdated”model cannot be eliminated by using the EIV model alone.