A case study on data fusion with overlapping segments

Abstract

With the continual improvement and diversification of existing sensing modalities and the emergence of new sensing technologies, methods to effectively and efficiently fuse the diverse and heterogeneous data sets are increasingly important. When different sensors acquire data over the same region of the Earth, a direct comparison between pixels acquired from one sensor to pixels acquired from a another sensor becomes difficult. For example, there could be different number of bands, or the sensors could measure drastically different spaces (hyperspectral and LIDAR). A solution to this problem is Feature Space Rotation, which realizes the sensor data independently in separate feature spaces via a machine learning algorithm and then a rotation is learned to bring the separate feature spaces into a common feature space. This rotation, in it original form, requires some amount of overlap between the data sets. We propose a study to determine the effect of decreasing the amount of overlap between the two sensors has on the classification accuracy. For this study, we shall rely on hyperspectral data that has been simulated to come from two disjoint sensors.