The use of non-destructive testing (NDT) equipment, such as the falling weight deflectometer (FWD), provides important estimates of road health and helps to optimize road management regimes. However, periodic road testing and post-processing of the collected data are cumbersome and require much expertise, a considerable amount of time, money, and other resources. This study attempts to develop a reliable prediction method for estimating the deflection basin area of different asphalt pavements using road temperature, load time, and load pressure as main characteristics. The data are obtained from 19 kinds of asphalt pavements on a 2.038-km-long full-scale field accelerated pavement testing track named RIOHTrack (Research Institute of Highway Track) in Tongzhou, Beijing. In addition, a chaotic particle swarm algorithm (CPSO) and a segmented regression strategy are proposed in this paper to optimize the XGBoost model. The experiment results of the proposed method are compared with those of classical machine learning algorithms and achieve an average of mean square error and mean absolute error respectively by 5.80 and 1.59. The experiments demonstrate the superiority of the XGBoost algorithm over classical machine learning methods in dealing with nonlinear problems in road engineering. Significantly, the method can reduce the frequency of deflection tests without affecting its estimation accuracy, which is a promising alternative way to facilitate the rapid assessment of pavement conditions.