Aims Phenology model is considered as the most efficient tool to assess the phonological responses of plants to future climate change. Furthermore, as an important component in dynamic ecological models, the performance of phenology model is of significance for the precision in simulating mass and energy exchanges between land and atmosphere. Combining long time series remote sensing data and climate data to construct regional phenology model may be the only way to solve the problem of deficiency in lacking in-situ observational data on phenology and species-specific phenology models. The objective of this study was to develop a new method of sample selections for constructing phenology in the arid zone of Central Asia where only sparse observational data are available. Methods Based the phenology data retrieved from 250 m-resolution MODIS images for the period 2000-2010, a new method is developed for constructing the vegetation phenology model in arid zone. A set of rules were built to select the representative pixels of PFTs (plant function types) surrounding the climate station, making sure that the climate of the representative pixels of PFTs could be represented by the observed meteorological data at the station, and then the phenology data on the representative pixels of PFTs were extracted from the MODIS images and the corresponding climatic data were set as the sample for model fitting and assessment. Forty-six representative pixels of PFTs for desert grassland vegetation and broadleaved deciduous forests were selected under the rules. The phenology model parameters were estimated using data from odd-numbered years and the simulation accuracy was assessed with the independent even-year data. Particle swarm optimization algorithm was used to parameterize the pre-selected model. The root-mean-square error and coefficient of determination were used to examine the performance of model with independent data. Important findings The best model for desert steppe vegetation was the modified temperature-precipitation model and the best model for deciduous broadleaved forest was the alternative model. Compared with other documented findings, the new method was proven feasible, and the results also suggested that this new method may improve the spatial match of climatic data and vegetation phenology data, and therefore contribute to improvement of the phenology model accuracy.