Rotation Invariance Regularization for Remote Sensing Image Scene Classification with Convolutional Neural Networks

Deep convolutional neural networks (DCNNs) have shown significant improvements in remote sensing image scene classification for powerful feature representations. However, because of the high variance and volume limitations available datasets, DCNNs are prone to overfit data used their training. To address this problem, paper proposes a novel framework based on deep Siamese network with rotation invariance regularization. Specifically, we design augmentation strategy model learn DCNN that is achieved by directly enforcing labels training samples before after rotating be mapped close each other. In addition cross-entropy cost function traditional CNN models, impose regularization constraint objective our proposed model. The experimental results obtained using three publicly-available datasets show method can generally improve performance 2~3% achieves satisfactory compared some state-of-the-art methods.