Discrimination of Maize Genotypes through Multi-temporal Object-based Remote-sensing Classification of Unmanned Aircraft System Images
Abstract
Remote-sensing classification in agriculture provides advanced techniques for expanding the capacity of high-throughput phenotyping in plant breeding and optimizing field management in precision agriculture. Methods applied in this study seek to advance remote sensing of plant breeding and precision agriculture through the classification of maize genotypes from unmanned aircraft system (UAS) images. Random forests (RF) and stochastic gradient boosting (SGB) algorithms were applied for classification of 12 maize genotypes at a row-scale analysis. Classification was achieved through the combined use of object-based image analysis (OBIA) and multi-temporal image analysis. The classification utilizes “object properties,” or variables (image layer statistics, object texture, object geometry, Structure from Motion (SfM) derived height measurements, and time-series measurements), to discriminate among genotypes. Object variables are evaluated for discriminative capacity in a maize-genotype context. Classification results with an accuracy of 81.25% supported the discrimination of 12 maize genotypes using only RGB images. This study further supports the use of SGB over RF in small class number classification, but RF in assessments of a larger number of classes. Multi-temporal dimensionality proved beneficial for classification; image and variable stack SGB classification produced results with an increase in accuracy of 30% and greater over single-date classification. Finally, study results identified the most discriminative object variables, such as mean canopy height model (Mean-CHM) and FRAGSTATS Cohesion metric, for classification of maize genotypes with the objective that future research will make use of derived variables and their association with plant characteristics for selection of optimal varieties of maize.
Citation
Walter, Andrew Crockett (2020). Discrimination of Maize Genotypes through Multi-temporal Object-based Remote-sensing Classification of Unmanned Aircraft System Images. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /191905.