Using Vegetation Indices as Input into Random Forest for Soybean and Weed Classification

Abstract

Weed management is a major component of a soybean ( Glycine max L.) production system; thus, managers need tools to help them distinguish soybean from weeds. Vegetation indices derived from light reflectance properties of plants have shown promise as tools to enhance differences among plants. The objective of this study was to evaluate normalized difference vegetation indices derived from multispectral leaf reflectance data as input into random forest machine learner to differentiate soybean and three broad leaf weeds: Palmer amaranth ( Amaranthus palmeri L.), redroot pigweed ( A. retroflexus L.), and velvetleaf ( Abutilon theophrasti Medik). Leaf reflectance measurements were acquired from plants grown in two separate greenhouse experiments conducted in 2014. Twelve normalized difference vegetation indices were derived from the reflectance measurements, including advanced, green, greenred, green-blue, and normalized difference vegetation indices, shortwave infrared water stress indices, normalized difference pigment and red edge indices, and structure insensitive pigment index. Using the twelve vegetation indices as input variables, the conditional inference version of random forest (cforest) readily distinguished soybean and velvetleaf from the two pigweeds (Palmer amaranth and redroot pigweed) and from each other with classification accuracies ranging from 93.3% to 100%. The greatest errors were observed between the two pigweed classes, with classification accuracies ranging from 70% to 93.3%. Results suggest combining them into one class to increase classification accuracy. Vegetation indices results were equivalent to or slightly better than results obtained with sixteen multispectral bands used as input data into cforest. This research further supports using vegetation indices and machine learning algorithms such as cforest as decision support tools for weed identification.