MODIS satellites provide continuous global observations on land surface temperature. It is more important in data-sparse
area, such as on the Tibetan Plateau (TP) with very few meteorological stations. Images with severe data missing or poor
quality pixels were often found in MODIS LST products, which mostly were caused by the influences of clouds. The
traditional geo-statistic methods, including ordinary Kriging and inverse distance weighted (IDW) methods, cannot well
interpolate missing-data pixels for a large area.
Assuming that the changes of LST at one location would be similar with that at the locations with similar features, a
novel method was proposed to interpolate the missing-data pixels by making use of other pixels with the most similar
features. MODIS/Terra LST covering TP in 2005 were used as experimental data, and pixels with cloud coverage,
average emissivity error greater than 0.04, and average LST error greater than 2K were identified as missing-data pixels.
The images with less than 10% missing-data pixels were selected as reference images, in which the missing-data pixels
were interpolated with IDW. Distances for different land surface features in images, such as DEM, slope, NDVI and LST,
from the interpolating pixel to the other pixels with known LST were calculated. Similar pixels are identified as the
distances less than a given threshold. Relationship of LST for those similar pixels was regressed, and was applied to
estimate LSTs for the missing pixels. Compared with IDW and Kriging, the proposed method could interpolate the
MODIS LST much better on the Tibetan Plateau.