Mapping weed infestation in maize fields using Sentinel-2 data

Abstract

Weed management in maize farms is a time-specific activity and requires timely detection of weed infestations. The challenge to early detection of weeds is that many dicotyledonous crops and broad-leaved weeds often display similar reflectance profile in the early growth stage and requires hyperspectral data to detect them. However, the advent of Sentinel-2 sensor series, with enhanced spectral configuration featuring red-edge bands that are known for species-level discrimination of plants, presents an affordable opportunity to detect weeds using multispectral data. The present study explores the question of whether Sentinel-2 sensor with its advanced spectral configuration can differentiate weeds from maize (Zea mays) in the early growth stages of maize. The study recorded 165 GPS points of weeds, maize, and mixed class in six maize farms during the early stages of maize growth. These GPS points were overlaid on Sentinel-2 images acquired within two days of field data gathering to guide the collection of spectral signatures of the maize, mixed, and weed classes. Spectral signatures were divided into training (70%) and validation (30%) data in a Random Forest (RF) model with S-2 spectral bands and vegetation indices as predictor variables. Spectral signatures were firstly tested for spectral separability between classes using ANOVA. The results of spectral analysis showed that the weed class had higher interclass variability from the maize and mixed class particularly in the red-edge and NIR regions of Sentinel-2. The classification matrix consistently showed that weeds were detected with high user and producers’ accuracy of 95%. These results indicate the utility of the enhanced spectral configuration of Sentinel-2 data in the early detection of weeds in maize farms.