Near real-time soybean phenology detection using proximally sensed hyperspectral canopy reflectance and machine learning methods

To date, most of the research data on the selection of plant types based on crop growth phenology is obtained by expert observations in the field. The extraction of soybean phenology stages/phases from canopy spectral data might provide timing information for optimal field management. In this work, a simple high-resolution technique for collecting proximally sensed hyperspectral data at the canopy level was performed, and machine learning phenology models based on canopy spectral data plus physical predictors were developed. An empirical study of the effects of parameterization of machine learning models with physical ancillary predictors is conducted, estimating soybean phenology in datasets collected in a-posteriori time rather than the interval where the training data were gathered. The overall accuracy achieved for reproductive single stages from the beginning of seed development to full maturity was 84%; when these stages were grouped into phases, the detection accuracy reached 92%. The ability to identify soybean phenology in near real time allows extrapolation of this study to a regional scale by employing unmanned aerial vehicles (UAVs) outfitted with the same instruments used in this work.