Evaluating the performance of object-based machine learning and deep learning models in classifying different maize genotypes with multispectral UAV imagery

Abstract

Modern remote sensing technologies play a critical role in agricultural applications, especially in recent years with the advances in unmanned aerial vehicle (UAV) technologies and artificial intelligence. Remotely sensed imagery is an invaluable data source for sustainable agricultural activities, such as precision agriculture. This study evaluates a comparative analysis of the performance of machine learning (ML) and deep learning models in classifying 12 different maize genotopies from multispectral UAV images. In this context, ortho mosaic and canopy height model obtained from UAV-mounted multispectral camera of the study area in Kirazca Agricultural Enterprise located in Arifiye district of Sakarya province were used as a main dataset. The object-based classification results show that the overall accuracy (OA) of the crop maps produced with the Rotation Forest (RotFor) and Canonical Correlation Forest algorithms was approximately 80%, while the OA value was 74.18% for Support Vector Machine algorithm. On the other hand, the popular U-Net model outperformed the ML-based models with an OA value of 97.61%. Individual class accuracy analyses revealed that the RotFor algorithm attained F-score values exceeding 90% for only 2 maize genotypes (i.e., Com. Sw. and Com. Ar.), whereas F-score values calculated with the U-Net model surpassed 95% for all 12 genotypes.