Classification of hybrid maize seeds (Zea mays) with object-based machine learning algorithms using multispectral UAV imagery

Abstract

In recent years, detailed monitoring of different vegetation classes by using modern remote
sensing technologies has become one of the essential issues for smart agriculture activities. In
this study, using three advanced machine learning algorithms, namely canonical correlation
forest (CCF), rotation forest (RotFor) and support vector machines (SVM), and object-based
image classification techniques on multispectral (MS) unmanned aerial vehicle (UAV)
orthomosaics, the separability of 12 maize species were investigated. The investigations were
performed in Sakarya Maize Research Institute application area located in Arifiye district of
Sakarya province, Turkey. In maize monitoring, besides the five spectral bands (R, G, B, red
edge, NIR) of the MS UAV, the Normalized Digital Surface Model (NDSM) describing the height
of maize species was generated and included as an additional band to improve classification
performance, evaluated with F-score, overall accuracy (OA) and Kappa metrics. The results
demonstrated that CCF and RotFor algorithms provide similar OA as 76.61% and 76.75%,
respectively and the SVM algorithm has 74.18%. In parallel, the Kappa values of CCF and
RotFor are 0.75 and the SVM is 0.72. In terms of class-based F-scores, by all algorithms, C.
Sweet and C. Arifiye were identified with over 97% and 94% accuracies, respectively, that
prove the successful determination of their boundaries using object-based classification.