Estimating Crop Physiological Traits From Remote Sensing Data Using Inverse Modeling.

High-throughput phenotyping for crop improvement requires that large amounts of field data be rapidly collected, processed, and analyzed for multiple crop genotypes.  Remote sensing and 'proximal' sensing have been identified as promising technologies for rapid characterization of phenotypes expressed by crop genotypes in the field.  However, traditional remote sensing instruments are largely dependent on quantifying reflectance of radiation from the crop canopy.  We hypothesize that a transformative increase in phenotypic data could be realized by using the remote sensing observations to invert physical or ecophysiological models.  Most numerical models are designed for forward modeling, where the attributes of a given system are converted to observable quantities based on the model's physical or biological state equations.  Inverse modeling does the reverse, using observed data to infer attributes of the system.  In practice, inverse modeling often requires an optimization algorithm, which runs the model forward many times, adjusting the input parameters to achieve some objective, such as minimizing error between an observed and simulated quantity.  Although inverse modeling appears to be a powerful tool for estimating crop phenotypes from remote sensing observations, the procedure does present challenges and risks.  Due to incomplete knowledge of system processes, simplifications in model design, and input parameter error, different configurations of a model may provide equally reasonable results, a condition known as equifinality.  A remaining question is thus whether a physical or ecophysiological model can be adequately constrained to provide meaningful estimates of crop phenotypes.  Issues of remote sensing data timing, availability, and spectral resolution will also likely impact the inverse modeling solution.  Additionally, the specific details of the inversion procedure will be important, including which parameters are allowed to be flexible, which parameters are constrained, the allowable range for flexible parameters, and the nature of the objective function used for model inversion.