Compact Deep Color Features for Remote Sensing Scene Classification

Abstract Aerial scene classification is a challenging problem in understanding high-resolution remote sensing images. Most recent aerial approaches are based on Convolutional Neural Networks (CNNs). These CNN models trained large amount of labeled data and the de facto practice to use RGB patches as input networks. However, importance color within deep learning framework yet be investigated for classification. In this work, we investigate fusion several models, using representations, We show that combining significantly improves recognition performance compared network alone. This improvement is, however, achieved at cost high-dimensional final image representation. propose an information theoretic compression approach counter issue, leading compact feature set without any significant loss accuracy. Comprehensive experiments performed five benchmarks: UC-Merced with 21 classes, WHU-RS19 19 types, RSSCN7 7 categories, AID 30 NWPU-RESISC45 45 categories. Our results clearly demonstrate features always overall standard features. On large-scale dataset, our provide absolute gain 4.3% over