Improved crop classification using multitemporal RapidEye data

Abstract

Land Use/Land Cover (LU/LC) of agricultural areas derived from remotely sensed data still remains very challenging. With regard to the rising availability and the improving spatial resolution of satellite data, multitemporal analyses become increasingly important for remote sensing investigations. Even crops with similar spectral behaviour can be separated by adding spectral information of different phenological stages. Hence, the potential of multi-date RapidEye data for classifying numerous agricultural classes was investigated in this study. In an agricultural area in Northern Israel two complete crop cycles 2013 and 2014 with two cultivation periods each were investigated. In order to avoid a high number of classification runs, a pre-procedure was tested to get the multitemporal data set which provides best spectral separability. Therefore, Jeffries-Matusita (JM) measure was used in order to obtain the best multitemporal setting of all available images within one cultivation period. Eight classifiers were applied to compare the potential of separating crops. The three algorithms Maximum Likelihood (ML), Random Forest (RF) and Support Vector Machine (SVM) outperformed by far the other classifiers with Overall Accuracies higher than 90 %. The processing time of ML and RF, however, was significantly shorter compared to SVM, in fact by a factor of five to seven.