I. Introduction
Annual average daily traffic (AADT) or traffic count data are collected by permanent automatic traffic recorders (PATRs) and calculated as short-term traffic counts (daily or weekly) multiplied by seasonal adjustment factors at different statewide locations [1]. Accurate estimation of AADT is critical to traffic demand management, traffic-induced emission estimation, freeway performance measurement, and traffic accident evaluation [2]–[4]. However, the quality of AADT is hardly met due to the limited development of PATRs and the failures of traffic sensors. Zhong et al. [5] reported that approximately 50% of PATRs suffer from the issue of missing data [6]. A common countermeasure to address this issue is to either retake traffic counts or predict the missing data based on observed data. The collection of traffic counts is labor intensive and time consuming; thus, the design of suitable data imputation methods with acceptable estimation accuracies becomes a more cost-efficient strategy [7]. The temporal data prediction technique based on historical records may not be a feasible solution because of continuous and large-scale missing values. Considerable attention has been shifted to spatial interpolation approaches [8], [9]. Spatial interpolation methods utilize sparsely scattered samples to estimate missing data at unmeasured locations and assume the existence of spatial autocorrelation described in the First Law of Geography (i.e., the degree of correlation of the observations normally decreases as their spatial distance increases)” [10]. The law refers to the basic concept of spatial dependence and offers theoretical support for spatial interpolation [11].