The invention discloses an improved nuclear regression ground temperature observation data quality control method. The method comprises the following steps: S1, screening experimental data; S2, introducing a kernel regression method and improving ground air temperature observation data needing to be applied; S3, improving the applicability of the multi-site kernel regression method on the basis ofS2; and S4, introducing an adaptive algorithm and a particle swarm algorithm to improve a window width coefficient in a kernel regression method, and performing regression prediction and corresponding quality control according to the principle. According to the invention, on the aspect of the method, a kernel regression method is introduced, ground temperature observation data is improved, meanwhile, the window width coefficient in the kernel regression method is further improved through the self-adaption and particle swarm optimization algorithm, and the direction of a traditional quality control method is expanded; in the aspect of quality control effect, the method has good effects in the aspects of prediction precision, universality, error detection rate and the like in the field of quality control, and is more beneficial to research and application of quality control of ground air temperature observation data.