Aerial scene classification with convolutional neural networks

Abstract

A robust satellite image classification is the fundamental step for aerial image understanding. However current methods with handcrafted features and conventional classifiers have limited performance. In this paper we introduced convolutional neural network (CNN) method into this problem. Two approaches, including using conventional classifier with CNN features and direct classification with trained CNN models, are investigated with experiments. Our method achieved 97.4% accuracy on 5-fold cross-validation test of the UCMERCED LULC dataset, which is 8% higher than state-of-the-art methods.