Multisensor data fusion and feature extraction for forestry applications

In this paper we discuss feature level multisensor data fusion with P-, L-, and C-band polarimetric synthetic aperture radar (PolSAR) data and multispectral Landsat Thematic Mapper (TM) data. The application is classification of Maritime pine age classes and bare ground in the Nezer forest in France. Multisensor data fusion is motivated by the complementary information available in SAR and optical data. Our objective is to investigate the choice of features among twenty six well known descriptors. First, we demonstrate the benefit of multisensor data fusion for improved classification performance over single sensor data classification with respect to forest monitoring. A comparison of the classification performances among the four different datasets reveals that the P-band SAR features yield the best results. By combining the P-band SAR features with the multispectral optical features, a significant classification accuracy improvement of 12.6% is achieved. Second, all twenty six features extracted in total from the four datasets are investigated for the purpose of identifying those features jointly possessing the highest discrimination power. Five features are found to preserve 98.5%of the classification information compared to classification based on the total set of features. This shows the advantage of feature selection with respect to preserving classification information while at the same time reducing the dimensionality of the feature space. A potential for improving the classification performance is also found by applying a thorough feature selection procedure.