Convolutional neural networks (CNNs) have been widely used in remote sensing scene classification due to their excellent performance in natural image classification. However, the complementarity of features extracted by different CNNs is seldom exploited, which limits the further improvement of classification accuracy. To solve this problem, we propose a classification method based on multi-structure deep features fusion (MSDFF). First, a data augmentation method based on random-scale cropping is adopted to achieve the expansion of limited data. Then, three popular CNNs are respectively used as feature extractors to capture deep features from the image. Finally, a deep feature fusion network is adopted to fuse these features and implement the classification. The effectiveness of the proposed method is verified on UC Merced, AID, and NWPU-RESISC45 datasets. The proposed method can achieve better performance than state-of-the-art scene classification methods. INDEX TERMS Convolutional neural network, scene classification, feature extraction, multi-structure deep features fusion.

No abstract

Abstract:In this work, a novel highly fabrication tolerant polarization beam splitter (PBS) is presented on an InP platform. To achieve the splitting, we combine the Pockels effect and the plasma dispersion effect in a symmetric 1x2 Mach-Zehnder interferometer (MZI). One p-i-n phase shifter of the MZI is driven in forward bias to exploit the plasma dispersion effect and modify the phase of both the TE and TM mode. The other arm of the MZI is driven in reverse bias to exploit the Pockels effect which affects only the TE mode. By adjusting the voltages of the two phase shifters, a different interference condition can be set for the TE and the TM modes thereby splitting them at the output of the MZI. By adjusting the voltages, the very tight fabrication tolerances known for fully passive PBS are eased. The experimental results show that an extinction ratio better than 15 dB and an on-chip loss of 3.5 dB over the full Cband (1530-1565nm) are achieved.

No abstract