Protein, lignin, cellulose, hemicellulose, sugar, and starch contents per unit leaf area of fresh leaves were related to the corresponding reflectance and transmittance spectra using multiple linear regressions. When the regressions are evaluated over the data set used for calibrating, they show relatively good performances. However, applying these relations to an independent data set led to very poor performances. It was concluded that the detailed canopy biochemistry was not accessible in a robust way from fresh leaf optical properties measurements over a large range of leaf types. The only variables that can be accurately derived from leaf reflectance or transmittance measurements are water and dry matter (i.e., the specific leaf weight) contents per unit leaf area. Transforming reflectance (rho) or transmittance (tau) values into the corresponding absorbance (log(1/rho) or log(1/tau)) values improves the accuracy of the estimates. Using transmittance rather than reflectance provides better retrieval performances. We investigated the sensitivity of the relationships to the radiometric noise associated to reflectance or transmittance measurements. It appears, particularly for water, that the estimates are quite sensitive to the radiometric resolution of the instrument used. We propose a technique that minimizes the sensitivity of the estimates to the radiometric noise and improves their robustness. It consists of enlarging the calibration data set by adding random instrumental noise similar to that observed over the test data set. Results show that, this way, three wavebands (1910, 1380, and 900nm for water; 2310, 2160, and 1870 nm for dry matter) allow good estimates of water (RMSE 0.0011g cm 2) and dry matter contents (RMSE 0.0008 g cm 2).