Analyzing the impact of red-edge band on land use land cover classification using multispectral RapidEye imagery and machine learning techniques

Abstract

RapidEye multispectral imager is the first satellite that consists of red-edge (RE) band. This study aims to analyze the effect of incorporating RE band information on land use land cover (LULC) classes. It further investigates the impact of integrating the most common vegetation indices: normalized difference vegetation index (NDVI) and its RE adaption, i.e., NDVI-RE in the classification process and sensitiveness of RE band information on LULC classes. In addition, this study also examines the potential of an advance ensemble technique, i.e., extreme gradient boosting (XGBoost) in comparison to the two state-of-the-art machine learning algorithms, random forest (RF) and support vector machine (SVM). A systematic comparison is performed using machine learning classifiers in case of inclusion and exclusion of RE and inclusion of NDVI versus NDVI-RE. Results show that inclusion of RE band improves accuracy of classification by +2.89 % , +3.49 % , and +3.03 % over without RE using XGBoost, RF, and SVM, respectively. The results obtained using all classifiers confirmed the effectiveness of NDVI-RE over NDVI for LULC classification. However, XGBoost outperformed RF and SVM by achieving highest overall accuracy of 92.41% when NDVI-RE is included as an input feature. For class-specific performance, incorporation of RE shows a significant rise in accuracy of all vegetation classes, whereas the inclusion of NDVI-RE reported a maximum increase in the accuracy of shrubland. Furthermore, results clearly indicated that XGBoost has great potential, and it may be beneficial for more complex classification problems.