This paper addresses the characterization of Land Surface Temperature (LST) variability according to Land Cover. It is a first step of a study which concerns the extraction of hydrological parameters in a semi-arid catchment applied located in Southern-Africa, and which includes image processing of satellite data. The main applicative interest of this work is to make available profiles of evapotranspiration (ET), which can be derived from LST, and to detect hydric stress by comparison between profiles of ET: potential ET simulated by an hydrological model and that estimated by satellite measurements. LST can be daily computed using the two thermal bands of NOAA/AVHRR. However, due to its coarse resolution (1.1 km at nadir), a NOAA/AVHRR pixel includes several land cover types and LST cannot be linked to a particular component. So, we process a data fusion between NOAA/AVHRR acquisitions and one high resolution land-use classification derived from Landsat-TM (30 meters at nadir), and consider a physical-based mixture model of the temperature pixel. Inverting this model on a learning area outputs individual temporal profiles of LST for each land cover type: bare soil, vegetated surface (grass, arable land, forest...). The obtained results with Landsat classification are then used to generate LST maps at spatial resolution of 30 meters and with a daily frequency.