Impact of Reducing Polarimetric SAR Input on the Uncertainty of Crop Classifications Based on the Random Forests Algorithm

Abstract

Although the use of multidate polarimetric synthetic aperture radar (SAR) data for highly accurate land cover classification has been acknowledged in the literature, the high dimensionality of the data set remains a major issue. This study presents two different strategies to reduce the number of features in multidate SAR data sets: an accuracy-oriented reduction and an efficiency-oriented reduction. For both strategies, the effect of feature reduction on the quality of the land cover map is assessed. The analyzed data set consists of 20 polarimetric features derived from L-band (1.25 GHz) SAR acquired by the Danish EMISAR on four dates within the period April to July in 1998. The predictive capacity of each feature is analyzed by the importance score generated by random forests (RF). Results show that according to the variation in importance score over time, a distinction can be made between general and specific features for crop classification. Based on the importance ranking, features are gradually removed from the single-date data sets in order to construct several multidate data sets with decreasing dimensionality. In the accuracy-oriented and efficiency-oriented reduction, the input is limited to eight and three features per acquisition, respectively. On the reduced input, a multidate model is built using the RF algorithm. Results indicate a decline in the classification uncertainty when feature reduction is performed.