Comparison of Data Mining and GDD-Based Models in Discrimination of Maize Phenology

Abstract

Data mining approaches are designed for classification problems in which each observation is a member of one and only one class. In this study, a non-deterministic approach based on C5.0 data mining algorithm has been employed for discriminating the phenological stages of maize from emergence to dough, in a field located in Karaj, Iran. Two readily-available predictors i.e. accumulated growing degree days (AGDD) and multi-temporal LANDSAT7-extracted normalized difference vegetation index (NDVI) was used to build the decision tree. The AGDD was calculated based on three cardinal thresholds of temperature i.e. effective minimum, optimum, effective maximum. The NDVI was compared with two recently developed indices namely, enhanced vegetation index2 (EVI2) and optimized soil adjusted vegetation index (OSAVI) using the signal to noise ratio (SNR) criterion. Findings confirmed that these three remotely sensed indices do not have significant differences, therefore, the smoothed time series of NDVI was used in the C5.0 algorithm. The precisions of classification by C5.0 data mining algorithm in partitioning of training and testing data were approximately 90.51 and 81.77%, respectively. The mean absolute error (MAE) values of the onset of maize phenological stages were estimated about 2.6–5.3 days for various stages by C5.0 model. While corresponding values for the classical AGDD model were 3.9–10.7 days. This confirms the skill of data mining approach in comparison with commonly-used the classical AGDD model in applications of real time monitoring.