Urban thermal environment studies require high spatial and temporal resolutions

while currently available remote s ensors exhibit a compromise between spatial and temporal resolutions. Downscaling land surface temperature（ LST） data with a low spatial resolution but high temporal resolution is a method to derive LST product at application-demanded spatial and temporal resolutions and has become a critical issue in thermal remote sensing applications. This study aims at downscaling LST from coarser spatial resolution（ 960 m） to finer（ 120 m） resolution by constructing a trend surface of spectral indices over highly heterogeneous urban areas

to lay a foundation for potential downscaling of MODIS LST to TM spatial resolutions. To describe the complex composition of built-up areas and vegetated surface in urban and surrounding areas

normalized difference built-up index（ NDBI） and normalized difference vegetation index（ NDVI） are combined to construct a trend surface to downscale LST from 960 m to 120 m.Urban areas in Beijing is taken as a study case. The key is to construct a trend surface of spectral indices to fit the LST

that is

to derive a trend surface transfer function at 960 m

and then to use the transfer function established at 960 m to estimate LST at 120 m using spectral index trend surface at 120 m. To validate the applicability and stability of trend surface transfer function

the fitting degrees of spectral indices to LST are analyzed from the central urban areas to the sub-urban in Beijing and at varying scales from 120 m

240 m

480 m to 960 m; the scale effects of trend surface transfer function is also analyzed. The linear functions of transferring NDVI

NDBI and combined NDVI + NDBI to LST are established respectively to downscale LST from 960 m to 1 20 m.The results show that:（ 1） the LST in Beijing is the highest in urban central areas surrounded by the fifth ring road and is gradually decreasing from areas within fifth ring road to areas outside the sixth ring road; the built-up and bare soil areas have the highest average temperature values

while vegetation and water bodies have low average temperature values;（ 2） NDBI fitted the LST the best over the densely distributed built-up urban areas

while NDVI is more suitable to fit LST over naturally covered areas outside the sixth ring road; the combined use of NDVI + NDBI can promote the fitting degrees to LST

depending on the composition of land cover types;（ 3） the correlation of spectral indices and LST applies on four varying scales from 120 m to 960 m. The errors induced by scale effects of transfer function are much lower than the errors brought by the models themselves;（ 4） the accuracy of downscaled LST using combined NDVI + NDBI model is promoted

compared to those using a single spectral index. Yet the degree of accuracy enhancement depends on the composition of land cover types;（ 5） downscaling errors are related to aggregation errors.Although aggregation errors are higher than downscaling errors

they remain an important source of downscaling errors. The methods in this paper is also applicable to downscaling of MODIS LST data

while many factors such as temporal difference c orrection and geometric correction should be considered further.