基于VTCI和分位数回归模型的冬小麦单产估测方法

条件植被温度指数(VTCI)是一种综合了归一化植被指数(NDVI)与地表温度(LST)的遥感干旱监测方法,在关中平原的近实时干旱监测中具有其适用性。分位数回归能全面反映因变量的条件分布在不同分位数处的特征,回归结果稳健可靠。为了进一步研究VTCI干旱监测结果与小麦单产之间的关系及提高冬小麦单产估测精度,构建了不同分位数τ(0.1,0.3,0.5,0.7,0.9)下关中平原各市2008—2014年的冬小麦主要生育期VTCI与单产之间的线性回归模型,并基于中位数(τ=0.5)回归模型对研究区域的冬小麦单产进行了估测。结果表明,分位数回归模型比较全面地反映了不同分位数下冬小麦单产分布与VTCI之间的相关程度,弥补了最小二乘估产模型回归结果单一、易受异常值影响等的不足。中位数回归模型的单产估测结果与实际单产之间的相对误差和均方根误差的最小值及平均值均低于最小二乘回归模型,估测精度较高。此外,中位数单产估测模型获取的冬小麦估产结果在年际变化规律与空间分布特征上与实际产量均较相符,说明分位数回归在研究VTCI与产量之间的关系及冬小麦单产估测中具有其适用性与可靠性。

Winter Wheat Yield Estimation Method Based on Quantile Regression Model and Remotely Sensed Vegetation Temperature Condition Index

Vegetation temperature condition index (VTCI) combines normalized difference vegetation index (NDVI) and land surface temperature (LST), and is applicable to a more accurate monitoring of droughts in Guanzhong Plain, Shaanxi Province, China. Quantile regression is a tool for comprehensively reflecting the conditional distribution characters under different quantiles, and its regression results are steady and reliable. In order to achieve a better correlation between winter wheat yield and the weighted VTCI as well as a higher yield estimation accuracy, linear regression models between the weighted VTCI and yields in the cities of Guanzhong Plain in the years from 2008 to 2014 were analyzed by using the quantile regression whose quantiles were set to be 0.1, 0.3, 0.5, 0.7 and 0.9, respectively. These quantile regression results roundly reflected the distribution of the yields under different drought conditions and were beneficial supplement of the linear regression from which the single fitted line and impressionable results from outliers were obtained. The wheat yield estimation model based on the median regression (quantile equalled to 0.5) was used to monitor the wheat yields in the cities of Guanzhong Plain from 2008 to 2014, the average and minimum values of the relative errors and the root mean square errors (RMSE) between the estimated yields and the actual yields were all lower than those derived from the ordinary least square method. Additionally, the characteristics of inter-annual evolution and spatial distribution of the estimated yields using the median regression model were in good agreement with the actual situation, which indicated that the quantile regression was feasible and reliable in the research of winter wheat yield estimation and the relationship between yield and drought.