AID: A Benchmark Dataset for Performance Evaluation of Aerial Scene Classification
In case when higher-order statistic is used for local feature aggregation, final descriptor can have very high dimensionality. In this paper different methods for descriptor dimensionality reduction are evaluated for land-use classification. Concretely, aerial image classification accuracy is compared for the cases when dimensionality reduction is made per band with fixed and variable sizes. For both aerial image datasets, experimental results showed that even for 10 to 50 times reduced descriptor dimensionality, classification performance can be preserved. Moreover, reduction approach with variable-per-band descriptor sizes achieved slightly better results compared with traditional approach that include fixed-per-band descriptor sizes.