Development of spectral indices for identifying glyphosate-resistant weeds

Abstract

Glyphosate as the most common and widely-used herbicide in agricultural crops has resulted in the explosion of resistant weeds around the world. An early detection of resistant weeds before observing the visible symptoms of glyphosate application in weeds can contribute to the crop protection in precision agriculture. This paper aims at developing and evaluating spectral weed indices (SWIs) to identify glyphosate-resistant weeds 72 h after herbicide application. A greenhouse experiment was conducted on three common weed species, namely, kochia (Kochia scoparia), ragweed (Ambrosia artemisiifolia L.) and waterhemp (Amaranthus rudis), including resistant and susceptible types to collect canopy spectral reflectance after glyphosate spraying. To generate SWI, the best weighted combination of single wavelength and a normalized wavelength difference in the range of 450–920 nm was established. Relief-F algorithm selected the most discriminative feature wavelengths, and two band normalized differences to differentiate weeds with higher degree of resistance, and changes in the hyperspectral signature caused by glyphosate application. The performance of optimized SWIs on resistant weeds identification was assessed by employing machine learning Random Forest (RF) method. The RF classification model achieved a classification accuracy of 96%, 100% and 97% in detecting resistant kochia, ragweed and waterhemp, respectively. A comparison between introduced SWIs in this study and previously published hyperspectral vegetation indices (VIs) indicated that the SWIs provided better agreement with real data in glyphosate-resistant weeds identification.