Measuring performance in precision agriculture: CART—A decision tree approach

Abstract

Recently, there have been very rapid developments in hyperspectral remote sensing and interest is fast growing in the applications of hyperspectral data to precision farming. This paper investigates the potential of hyperspectral remote sensing data for providing better crop management information for use in precision farming by using an artificial intelligence (AI) approach. In this study, the ability of the classification and regression trees (CART) decision tree algorithm is examined to classify hyperspectral data of experimental corn plots into categories of water stress, presence of weeds and nitrogen application rates. In the summer of 2003, a three-factor split-split-plot field experiment representing different crop conditions was carried out. Corn was grown under irrigated and non-irrigated conditions with two weed management strategies: no weed control, and full weed control and with three nitrogen levels of 50, 150, and 250 kg N ha −1. The hyperspectral data was recorded (spectral resolution = 1 nm) with a hand-held spectroradiometer at three developmental stages of corn—early growth, tasseling, and fully maturity. The CART decision tree algorithm was able to classify the 12 treatment combinations with 75–100% accuracy at all 3 recorded stages of development, although the best validation results were obtained at early growth stage. When decision trees (DTs) were generated to classify the plots according to two and then only one of the three factors (irrigation, weeds or nitrogen), the classification accuracy was ever highest. With the spectra obtained at early growth stage and single factor analysis, the classification accuracy was 96% for the irrigation factor, 83% for the nitrogen, and 100% for the weed control strategies.