Image Classification Using RapidEye Data: Integration of Spectral and Textual Features in a Random Forest Classifier

Citation

Zhang, H., Li, Q., Liu, J., Shang, J., Du, X., McNairn, H., Champagne, C., Dong, T., Liu, M. (2017). Image Classification Using RapidEye Data: Integration of Spectral and Textual Features in a Random Forest Classifier. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, [online] 10(12), 5334-5349. http://dx.doi.org/10.1109/JSTARS.2017.2774807

Plain language summary

In this research, a Random Forest Decision Tree classify was used with a time series of optical satellite data (from the RapidEye satellite) to classify crops. The researchers included not only the reflectance measured by the RapidEye satellite, but also derived vegetation indices from the reflectance data, as well as measures of the spatial variance or texture of the reflectance. Classification accuracies improved by about 7% when the vegetation indices were used with the RapidEye reflectance, in the Random Forest classifier. A further 6% improvement in accuracy was reported when the textural information from the satellite data was also used.

Abstract

Information on crop types derived from remotely sensed images provides valuable input for many applications such as crop growth modeling and yield forecasting. In this paper, a random forest (RF) classifier was used for crop classification using multispectral RapidEye imagery over two study sites, one in north-eastern China and one in eastern Ontario, Canada. Both vegetation indices (VIs) and textural features were derived from the RapidEye imagery and used for classification. A total of 20 VIs, categorized into two groups with and without the red edge (RE) band in an index, were calculated. A total of eight types of textural features were derived using four different window sizes from both the RE and the near-infrared bands. To reduce redundancies among the VIs and textural features, feature selection using the principal component analysis, correlation analysis, and stepwise discriminant analysis was performed. Results showed that the overall classification accuracy was improved by ∼7% when the RE indices were combined with the five spectral bands in classification, as compared with that using the five bands alone. When textural information was included, the overall classification accuracy increased by ∼6% compared with that using the band reflectance alone. Furthermore, when all the features (band reflectance, VIs, and texture) were used, the overall classification accuracy increased by ∼12% compared with that using only the band reflectance. The RF importance measures showed that the RE reflectance was important for classification, as indicated by the high importance for the triangular vegetation index, transformed chlorophyll absorption in reflectance index, and green-rededge normalized difference vegetation index. The gray-level co-occurrence matrix mean is the most useful for classification among the textural features. The study provides a means to feature extraction and selection for crop classification from remote sensing imagery.