Pre-planting weed detection based on ground field spectral data

Abstract

Site-specific weed management (SSWM) demands higher resolution data for mapping weeds in fields, but the success of this tool relies on the efficiency of optical sensors to discriminate weeds relative to other targets (soils and residues) before cash crop establishment. The objectives of this study were to (i) evaluate the accuracy of spectral bands to differentiate weeds (target) and other non-targets, (ii) access vegetation indices (VIs) to assist in the discrimination process, and (iii) evaluate the accuracy of the thresholds to distinguish weeds relative to non-targets for each VI using training and validation data sets. The main outcomes of this study for effectively distinguishing weeds from other non-targets are (i) training and validation data exhibited similar spectral curves, (ii) red and near-infrared spectral bands presented greater accuracy relative to the other bands, and (iii) the tested VIs increased the discrimination accuracy related to single bands, with an overall accuracy above 95% and a kappa above 0.93. This study provided a novel approach to distinguish weeds from other non-targets utilizing a ground-level sensor before cash crop planting based on field spectral data. However, the limitations of this study are related to the spatial resolution to distinguish weeds that might be closer to the one this study presented, and also related to the soil and crop residues conditions at the time of collecting the readings. Overall the results presented contribute to an improved understanding of spectral signatures from different targets (weeds, soils, and residues) before planting time supporting SSWM.

Luan Pierre Pott

PhD student in Agricultural Engineering

My research interests include digital agriculture, remote sensing, crop modeling and machine learning