The surface renewal method offers an inexpensive way to estimate the sensible heat flux (H). When combined with measurements of net radiation and soil heat flux, latent heat flux (evapotranspiration) can be estimated indirectly. This study evaluated the performance of two surface renewal models to estimate H in a traditional coffee agroforestry system. The first model (Snyder96) is the commonly used approach in which H is calculated from the average temperature ramp amplitude and duration, which in turn are calculated from the 2nd, 3rd, and 5th order structure functions. The second model (Chen97) is a reformulation of Snyder96 in which H is calculated from the 3rd order structure function and friction velocity (u*). The models were calibrated and evaluated with eddy covariance data. Tests were conducted at three levels (above-canopy, above-coffee, above-ground), with each level having two replicate plots. Both models showed generally good agreement with the observations when using humidity-corrected sonic temperature data. Estimates from Chen97 agreed slightly better with measurements than those from Snyder96. In addition, the results suggest that for Chen97 the same calibration can be used for the above-canopy and above-coffee levels and for positive and negative H. Also for the above-ground level the combined empirical coefficient of Chen97 was similar to those for the other levels, but the variability between plots was too high to draw conclusions. Conversely, the results showed that Snyder96 requires separate calibrations for each level and for positive and negative H. For Chen97, similarly good results were obtained when using frequency response-corrected thermocouple data. For Snyder96, results were also good above the canopy, but not within the canopy, particularly close to the ground. Chen97 needs wind speed to calculate u*. Wind speed is needed for both models if the signals of thicker, stronger thermocouples are corrected for the frequency response error.