By using the Landsat remote sensing images, land surface temperature (LST) of Lanzhou city, a valley city in Northwest China, in the summer was retrieved during 1990—2015. Ordinary linear regressions (OLS) models and geographically weighted regressions (GWR) models were used to investigate the relationships between the proportions of land cover change and LST and analyzed the spatial non-stability. The results indicated that the high summer land surface temperature mainly focus on unused land in north and south mountains, the valley basin where the Yellow river runs across has low temperature. Urban-heat-island ratio index decreased firstly and then increased. The proportions of land cover change were significantly correlated to LST, but with spatial non-stability which is mainly due to the different geographical locations and surrounding environments of different areas. OLS model might overestimate or underestimate the adjusting ability of different cover types on temperature, which may decrease or increase LST. The results obtained by GWR models are better than those by OLS models. What’s more, GWR models could intuitively and accurately reveal the spatial non-stability of the relationships between the proportions of different land cover and LST.