The sizes and shapes (physiognomy) of fossil leaves are widely applied as proxies for paleoclimatic and paleoecological variables. However, significant improvements to leaf-margin analysis, used for nearly a century to reconstruct mean annual temperature (MAT), have been elusive; also, relationships between physiognomy and many leaf ecological variables have not been quantified. Using the recently developed technique of digital leaf physiognomy, correlations of leaf physiognomy to MAT, leaf mass per area, and nitrogen content are quantified for a set of test sites from North and Central America. Many physiognomic variables correlate significantly with MAT, indicating a coordinated, convergent evolutionary response of fewer teeth, smaller tooth area, and lower degree of blade dissection in warmer environments. In addition, tooth area correlates negatively with leaf mass per area and positively with nitrogen content. Multiple linear regressions based on a subset of variables produce more accurate MAT estimates than leaf-margin analysis (standard errors of ±2 vs. ±3°C); improvements are greatest at sites with shallow water tables that are analogous to many fossil sites. The multivariate regressions remain robust even when based on one leaf per species, and the model most applicable to fossils shows no more signal degradation from leaf fragmentation than leaf-margin analysis.