Hyperspectral reflectance (350-2500 nm) measurements were made over two experimental rice fields containing two cultivars treated with three levels of nitrogen application. Four different transformations of the reflectance data were analyzed for their capability to predict rice biophysical parameters, comprising leaf area index (LAI; m 2 green leaf area m −2 soil) and green leaf chlorophyll density (GLCD; mg chlorophyll m −2 soil), using stepwise multiple regression (SMR) models and support vector machines (SVMs). Four transformations of the rice canopy data were made, comprising reflectances (R), first-order derivative reflectances (D1), second-order derivative reflectances (D2), and logarithm transformation of reflectances (LOG). The polynomial kernel (POLY) of the SVM using R was the best model to predict rice LAI, with a root mean square error (RMSE) of 1.0496 LAI units. The analysis of variance kernel of SVM using LOG was the best model to predict rice GLCD, with an RMSE of 523.0741 mg m −2 . The SVM approach was not only superior to SMR models for predicting the rice biophysical parameters, but also provided a useful exploratory and predictive tool for analyzing different transformations of reflectance data. The assessment of biophysical vegetation properties, such as leaf area index (LAI) and green leaf chlorophyll density (GLCD), is a major goal of remote sensing in agriculture. Remote-sensing-based assessments of these variables are made possible as a result of the strong contrast between spectral reflectances of vegetation and the soil background and the dramatic reflectance changes associated with changing vegetative cover. Based on this contrast, numerous vegetation indices (VIs) have been developed during the past few decades, which are highly correlated with the amount of vegetation. The most common of these indices use the red and near-infrared (NIR) canopy reflectances in the form of ratios, such as the ratio VI [1] and the normalized difference vegetation index [2], and as linear combinations of red and NIR reflectances [3,4]. These indices generally use averaged spectral information over broad bandwidths [5], resulting in the loss of critical information available in specific narrow bands [6], and potentially limiting the accurate estimates of agricultural crop and natural vegetation biophysical and biochemical variables [7,8]. In addition, many of these vegetation indices are strongly influenced by the soil background, resulting in soil-dependent VI-biophysical relationships [9,10]. Further improvements in quantifying vegetation are possible using spectral data from distinct narrow bands, as indicated by numerous hyperspectral studies using field spec-

No abstract