Detection and quantification of broadleaf weeds in turfgrass using close-range multispectral imagery with pixel- and object-based classification

Abstract

ABSTRACT The current practice used to evaluate broadleaf weed cover in turfgrass is visual assessment, which is time consuming and often leads to inconsistencies among evaluators. In this study, we investigated the effectiveness of constructing Random Forest models (RF), either pixel-, object-based (OBIA) or a combination of both to detect and quantify broadleaf weed cover. High resolution multispectral images were captured of 136 turfgrass plots, seeded with five species of Festuca L. and overseeded with either clover (Trifolium repens L.), daisy (Bellis perennis L.), yarrow (Achillea millefolium L.), or a mixture of all three weeds. Ground measurements of vegetation cover and bare soil were taken with a point quadrat and digital image analysis. Weeds were detected with 99% accuracy by OBIA, followed by the combined approach (98%) and Pixel-based approach (93%). Accuracy at distinguishing among weed species was somewhat lower (89%, 81% and 90%, respectively), with yarrow contributing most to the decrease in accuracy. The predictions based on ground measurements were further compared to field measurements. For both soil and weed classification, models that used shape features (OBIA and combined) resulted in better agreement with field measurements compared to Pixel- based classifications. Our study suggests that broadleaf weed cover comprised of species such as clover and daisy can be accurately quantified with high resolution multispectral images; however, quantifying yarrow cover remains challenging.