Submitted to: Transactions of the ASAE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/25/2006
Publication Date: 10/20/2006
Citation: Kaleita, A.L., Stewart, B.L., Ewing, R.P., Ashlock, D.A., Westgate, M.E., Hatfield, J.L. 2006. Detection of maize pollen shed from hyperspectral reflectance data. Transactions of the ASAE. 49:1947-1954. Interpretive Summary: Detection of pollen shed from corn crops is critical to assess the potential for contamination of fields with pollen from different genetic material. Pollen can be transported in the wind across fields; however, the estimation of when fields begin to release pollen requires observations of the onset to the emergence of tassels and initial pollen release. To address this problem we collected a series of observations with a radiometer that measures the reflectance from canopies in a series of narrow wavebands that encompass the visible and near-infrared regions of the light spectrum. These data were collected from prior to tassel emergence through completion of pollen shed. Different methods of evaluating the reflectance data for their reliability in estimating both the tassels present and initial pollen shed were conducted on the data sets. Several of the methods were not reliable in their predictive capabilities because of various problems. The method that may provide the most useful and reliable approach combines statistical methods with an understanding of the shape of the reflectance curve obtained from these studies. The approach developed provides a method that can be used by remote sensing scientists to evaluate different methods of detecting plant characteristics.
Technical Abstract: Detecting the onset of pollen shed by a corn crop is important in estimating the potential for transfer of genetic material to neighboring fields. Because field scouting is time and resource intensive, remote sensing approaches for detecting the onset of pollen shed could be beneficial. In this study, several numerical methods were investigated for estimating percentage of plants with visible tassels (VT) and at initial pollen shed (IPS) using hyperspectral remotely sensed data. Correlation analysis indicated regions of the spectra influenced by tasseling and anthesis, but no single band, when considered alone, was predictive (maximum correlation less than 0.35 for both VT and IPS). Classification using an artificial neural network (ANN) was predictive, correctly classifying 80.6% and 88.3% of the VT and IPS data respectively. The extensive preprocessing necessary and the "black box" nature of ANN's, however, rendered analysis of different regions of the spectra difficult using this method. Partial Least Squares (PLS) analysis yielded models with high predictive capability (R2 of 0.858 for VT and 0.80 for IPS); however, the PLS coefficients did not exhibit a spectrally consistent pattern. We also evaluated a range operator enabled genetic algorithm (ROEGA), which considers the shape of the spectra. This novel approach had similar predictive capabilities to the artificial neural network and to the PLS, but has the added advantage of allowing information transfer for increased domain knowledge.