The efficiency of four nonparametric species richness estimators - first-order Jackknife, second-order Jackknife, Chao2 and Bootstrap - was tested using simulated quadrat sampling of two field data sets (a sandy 'Dune' and adjacent 'Swale') in high diversity shrublands (kwongan) in south-western Australia. The data sets each comprised > 100 perennial plant species and > 10000 individuals, and the explicit (x-y co-ordinate) location of every individual. We applied two simulated sampling strategies to these data sets based on sampling quadrats of unit sizes 1/400th and 1/100th of total plot area. For each site and sampling strategy we obtained 250 independent sample curves, of 250 quadrats each, and compared the estimators' performances by using three indices of bias and precision: MRE (mean relative error), MSRE (mean squared relative error) and OVER (percentage overestimation). The analysis presented here is unique in providing sample estimates derived from a complete, field-based population census for a high diversity plant community. In general the true reference value was approached faster for a comparable area sampled for the smaller quadrat size and for the swale field data set, which was characterized by smaller plant size and higher plant density. Nevertheless, at least 15-30% of the total area needed to be sampled before reasonable estimates of S-t (total species richness) were obtained. In most field surveys, typically less than 1% of the total study domain is likely to be sampled, and at this sampling intensity underestimation is a problem. Results showed that the second-order Jackknife approached the actual value of S-t more quickly than the other estimators. All four estimators were better than S-obs (observed number of species). However, the behaviour of the tested estimators was not as good as expected, and even with large sample size (number of quadrats sampled) all of them failed to provide reliable estimates. First- and second-order Jackknives were positively biased whereas Chao2 and Bootstrap were negatively biased. The observed limitations in the estimators' performance suggests that there is still scope for new tools to be developed by statisticians to assist in the estimation of species richness from sample data, especially in communities with high species richness.