A secondary mission objective of NASA’s Ice, Cloud and Land Elevation Satellite-2 (ICESat-2) is to derive vegetation heights that can be used to refine estimates of above-ground biomass at a global scale [1] & [2]. In this work we assess the accuracy of ICESat-2 derived canopy height (CH) retrieval in dense, late-succession, tropical forests in Mexico, Belize, Guatemala, Honduras, Costa Rica and Brazil. Over 28,000 ICESat-2 ATL-08 canopy height estimates (CHe) are compared against equivalent metrics derived from spatially coincident linear-mode high density airborne lidar (HDL) data obtained over a variety of forests exhibiting different structural characteristics and underlying terrain conditions. Our preliminary results indicate that in these high closure forests the ATL08 canopy height estimates (CHe) differ to the airborne lidar canopy heights (CHr) over a very wide range. The 5 to 95 percentile errors range between an underestimation of negative 26 m to an overestimation of 27 m; with an interquartile range (IQR) between -14 to 0.25 m. When the samples are stratified using the (CHr) into 5 meter classes (5-10, 10-15, …, 40-45 m), the IQR of the canopy heigh estimation errors grows exponentially in relation to the CHr (1.25e^0.057 ; R2=0.97), while the median for each class decreases linearly at a rate of -0.174m/m (R2 = 0.88). When the ATL08 CHe are normalized to the reference CH the median of the errors exhibit a mostly uniform behavior around an underestimation level of 16.4% . By analyzing the behavior of other ATL08 parameters such as the terrain elevation and top of canopy elevations (both absolute elevations above the ellipsoid) as a function of the CHr, it is apparent that the gradual underestimation of the vegetation height is due to the inability of the ICESat-2 sensor and algorithms to accurately detect the terrain under these dense late-succession tropical forests.