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.

We propose and demonstrate a six-hole tapered silicon photonic crystal nanobeam cavity with a theoretical high quality (Q) factor and an ultrasmall mode volume based on machine learning. The crucial element to efficiently obtain high Q factors is to take the prediction result of the designed neural network as the fitness function of the genetic algorithm, whose evolution direction is developing towards higher fitness. Consequently, by combining the neural network and genetic algorithm iteration, an optimized photonic crystal nanobeam cavity with a theoretical Q factor (Q_th) as high as 1.2×10⁸ and an ultrasmall mode volume of 0.32(λ/n)³ is obtained. Leveraging the resonant scattering optical method, the cavity experimental Q factor (Q_exp) is measured as 2.17×10⁶, which is a record high experimental Q factor of silicon photonic crystal nanobeam cavity with maintaining an ultrasmall mode volume of 0.32(λ/n)³ and an ultra-compact device size of 6 μm². Owing to the ultra-high Q factor-to-mode volume ratio, the proposed photonic crystal nanobeam cavities could extremely enhance the interactions between light and matter, which have extensive important applications in low-threshold optical lasers, high-resolution filters and sensors.

We propose and experimentally demonstrate a narrowband notch microwave photonic filter (MPF) with an ultra-high all-optical tuning efficiency based on an optimized high-quality (Q) nanobeam photonic crystal (PhC) cavity. By optimizing the structural parameters, the hole shapes and the waveguide width of the cavity, a silicon nanobeam cavity with a theoretical high-Q factor of 4×10 ⁸ , a small mode volume (V) of 0.32(λ/ n ) ³ and a compact size of 6 μm ² could be obtained. Due to the extremely high Q/V ratio beyond 1×10 ⁹ , the nonlinear effects in the cavity could be efficiently excited, which is beneficial for on-chip low-power signal processing. In the experiment, by adjusting the wavelength of the optical carrier, the central frequency of the notch MPF could be widely tuned from 1 GHz to 40 GHz with maintaining the 3dB-bandwidth around 198 MHz and rejection ratios beyond 50 dB. On the other side, the central frequency of the MPF could be all-optically tuned based on the nonlinear effects in the cavity. Owing to the cavity extremely high Q/V ratio, the MPF frequency can be adjusted with an ultra-high tuning efficiency of 308.8 GHz/mW. To the best of our knowledge, whether the notch MPF narrow 3dB-bandwidth of 198 MHz or the all-optical tuning efficiency of 308.8 GHz/mW is a record value among the reported silicon-cavity-based notch MPFs with such a compact device size. With the dominant advantages of sub-gigahertz narrowband, wide tuning range, ultra-high all-optical tuning efficiency and extremely compact cavity size, the proposed MPFs illustrate competitive performance for on-chip microwave signal processing.

We propose and experimentally demonstrate sub-gigahertz microwave photonic filters (MPFs) with widely tunable bandwidths and ultra-high rejection ratios based on a silicon add-drop Mach–Zehnder interferometers (MZI) coupled ring array. To realize a ring filter with a widely tunable bandwidth and a flexibly adjustable extinction ratio, the two coupling waveguides of the silicon resonator are designed as bending waveguides and are both fabricated with microheaters. Consequently, the effective coupling coefficients between the add-drop ring and the coupling waveguides could be significantly manipulated, resulting in tunable bandwidths and extinction ratios. Capitalizing the designed resonators with broadly adjustable bandwidth, the experimental results show that the 3dB-bandwidth of the MPF could be widely tuned from 0.178 GHz to 22.7 GHz. What's more, owing to the ring flexibly tunable extinction ratios and the complete interference cancellation technology, the maximum MPF rejection ratios could realize beyond 75 dB. To the best of our knowledge, it is the first time to realize the notch MPFs with widely tunable 3dB-bandwidths ranging from sub-gigahertz to tens of gigahertz while maintaining the ultra-high rejection ratios. The proposed scheme paves a way for the notch MPFs with the dominant advantages of narrow bandwidths, wide bandwidth tuning ranges, ultra-high rejection ratios and complementary metal-oxide semiconductor (CMOS)-compatibility, which is competent to process dynamic radio frequency (RF) signals in on-chip microwave systems.

Ground-penetrating radar (GPR) is an effective tool for subsurface detection. Due to the influence of the environment and equipment, the echoes of GPR contain significant noise. In order to suppress noise for GPR data, a method based on singular value decomposition (SVD) of a window-length-optimized Hankel matrix is proposed in this paper. First, SVD is applied to decompose the Hankel matrix of the original data, and the fourth root of the fourth central moment of singular values is used to optimize the window length of the Hankel matrix. Then, the difference spectrum of singular values is used to construct a threshold, which is used to distinguish between components of effective signals and components of noise. Finally, the Hankel matrix is reconstructed with singular values corresponding to effective signals to suppress noise, and the denoised data are recovered from the reconstructed Hankel matrix. The effectiveness of the proposed method is verified with both synthetic and field measurements. The experimental results show that the proposed method can effectively improve noise removal performance under different detection scenarios.