Classification and calibration of metal exposure using leaf-level and simulated canopy-level reflectance data

Remote sensing technologies with high spatial and spectral resolution show a great deal of promise in addressing critical environmental monitoring issues, but the ability to analyze and interpret the data lags behind the technology. Robust analytical methods are required before the wealth of data available through remote sensing can be applied to a wide range of environmental problems for which remote detection is the best method, e.g. inaccessible areas, sites with potentially hazardous contamination, sites that need routine and frequent monitoring, and where information about the spatial context of such conditions is critical to understanding the location, distribution, or spread of adverse conditions.; In this study we compare the effectiveness of several statistical methodologies for the classification and calibration on a data set consisting of leaf-level reflectance from the salt marsh plant Frankenia that has been exposed to varying levels of copper toxicity. We also test these methodologies on simulated canopy-level reflectance produced by feeding the leaf-level data in the SAILH canopy reflectance model.; The classification methods compared were support vector classification (SVC) of exposed and non-exposed plants based on the reflectance data, and partial least squares compression of the reflectance data followed by classification using logistic discrimination (PLS/LD). PLS/LD was performed in two ways. We used the continuous concentration data as the response during compression, and then used the binary response required during logistic discrimination. We also used a binary response (−1 for non-exposed, and 1 for exposed) during compression followed by logistic discrimination. The statistic we used to compare the effectiveness of the methodologies was the leave-one-out cross validation estimate of the prediction error. We also calculated the theoretical (probably approximately correct or PAC) bounds on the prediction error of the support vector machine.; The calibration or regression methods compared were step-wise multiple regression (SMR) of the first derivative of log(1/ R) on the concentration, SMR of the second derivative of log(1/R) on concentration, and SMR of 9 optical indices currently thought to be effective in identifying changes in pigment content on concentration. We also use partial least squares regression, which uses partial least squares to compress the reflectance data to a few components and then uses least squares to regress the response (concentration) on the components. We performed all of the above regressions using both the concentration of the toxin to which the plants were exposed and the concentration found in the leaves of the plants at the end of the experiment. We used the predicted error sum of squares (PRESS) to compare the different regressions. We also calculate an estimate of the critical level, which is a value used in establishing detection limits in analytical chemistry.; PLS/LD using binary predictor variables during compression had the lowest estimated prediction (classification) error for both the leaf data. SVM classification had a lower prediction error for the simulated canopy level data. PLSR had the lowest PRESS statistic of all the regressions for both the leaf and simulated canopy data.