Evapotranspiration (ET) plays a crucial role in the land-atmosphere interaction and climate variability, especially in arid and semi-arid areas. Accurate estimates of ET are important in hydrological and climate modeling. This study evaluates eight ET data products from different models used for ET estimation. The data products are classified into three main categories depending on the type of modeling approaches: namely process-based land surface model, empirical models, and satellite data derived estimates. The different model estimates are evaluated against in situ measurements from the Skukuza flux tower which is situated in a semi-arid savanna in South Africa. The correlation score and cantered root mean square error computed on monthly ET averages indicate that the satellite-derived model and land surface model estimates are closer to the observed ET signal for the Skukuza site, both in-phase and magnitude. The empirical models' outputs tend to reflect a relatively pronounced departure from observations in magnitude. The normalised mean bias computed for different seasons reveals that the estimates from all modeling approaches are close to the observed signal during the transition period (March–May) to the austral summer. In general, all models overestimate ET during summer and underestimate it in winter. A qualitative analysis of the year-to-year variation for different seasons reveals that all model estimates are qualitatively consistent with the observed seasonal pattern of the signal. Satellite and process-based land surface models (LSMs) also show a response to extremes events such as drought years. The study identifies satellite-derived model outputs as a candidate for understanding spatio-temporal variability of ET across different landscapes within the study region, and process-based models to potentially be used for climate change impact studies on ET.