Detection of the type of trend in temperature data, by distinguishing between deterministic and stochastic trends, has important implications for understanding climate change. The Unit root tests (URTs) have been widely used for detecting the type of trend, but they do not consider the possibility of fractional integration and its influences. In this study, we detected the type of trends in observed surface air temperature during 1960–2019 at 558 stations across China, by considering fractional integration. The whole period was divided into three sub-periods by two structural breakpoints (denoted as SBP1 and SBP2). The fractional differencing parameter d was estimated by the Local Whittle (LW) function, and then three URTs (namely PP, KPSS, and ZA) were used to detect the type of trends in the temperature data during different sub-periods. The results indicated that the de-seasoned monthly temperature (DMT) series were fractionally integrated and, thus, exhibited long-range dependence characteristics, which significantly influenced the estimation of the slope of trend and detection of the type of trend. Compared with the LW function, the ordinary least squares method yielded biased estimation of the slope of trend, as it could not eliminate the influence of long-range dependence of the DMT series. The DMT series with weak long-range dependence or anti-persistent characteristics during 1960–SBP1, SBP1–SBP2, and SBP2–2019 were accurately detected as deterministic trends. However, the DMT series with strong long-range dependence during SBP1–2019 and 1960–2019 were detected to have stochastic trends by the KPSS test. Following the results of the fractional integration and URTs together, temperature over short periods in China were detected to have deterministic trends, but those at long periods were detected to have a combination of long-range dependence and deterministic trends.