ABSTRACT
In this article, we propose an automatic procedure for classification of UAV imagery to map weed presence in rice paddies at early stages of the growing cycle. The objective was to produce a weed map (common weeds and cover crop remnants) to support variable rate technologies for site-specific weed management. A multi-spectral ortho-mosaic, derived from images acquired by a Parrot Sequoia sensor mounted on a quadcopter, was classified through an unsupervised clustering algorithm; cluster labelling into ‘weed’/‘no weed’ classes was achieved using geo-referenced observations. We tested the best set of input features among spectral bands, spectral indices and textural metrics. Weed mapping performance was assessed by calculating overall accuracy (OA) and, for the weed class, omission (OE) and commission errors (CE). Classification results were assessed under an ‘alarmist’ approach in order to minimise the chance of overestimating weed coverage. Under this condition, we found that best results are provided by a set of spectral indices (OA = 96.5%, weed CE = 2.0%). The output weed map was aggregated to a grid layer of 5 × 5 m to simulate variable rate management units; a weed threshold was applied to identify the portion of the field to be subject to treatment with herbicides. Ancillary information on weed and crop conditions were derived over the grid cells to support precision agronomic management of rice crops at the early stage of growth.
Acknowledgments
This study was founded in the framework of FP7 ERMES project. (http://www.ermes-fp7space.eu/) funded by the European Union Seventh Framework Program (FP7/2007-2013) under grant agreement 606983. The authors acknowledge Parrot company which allows for the use of the Sequoia sensor for scientific research. The authors also wish to thank Azienda Agricola Carlo Franchino (Rosasco, Italy), for hosting our field activity and Dr. Ramin Azar for his help in field data collection.
Disclosure statement
No potential conflict of interest was reported by the authors.