The local connection characteristics of convolutional neural network (CNN) are linked with the local spatial correlation of image pixels for water depth retrieval in this paper. The method has greater advantages and higher precision than traditional retrieval methods. Traditional remote sensing empirical models require manual extraction of retrieval factors and the process is complex. This paper proposes a model based on CNN which uses different remote sensing images in four spectral bands, red, green, blue, and near-infrared, to retrieving the water depth. In general, CNN is mostly used for image recognition and classification tasks, which can make full use of the local spatial correlation between pixels. The method in this paper exploits this feature of CNN for water depth retrieval, taking into consideration the nonlinear relationship between the radiance value and water depth value from adjacent and central pixels. In this paper, remote sensing image data, measured water depth data, and lidar sounding data are used as input data to build the model. Then, the retrieval error is analyzed and the parameters are adjusted to further optimize the model. Quantitative analysis and experimental results show that the CNN model achieves higher precision in the retrieval of shallow sea areas compared with the traditional method. The mean absolute error can reach within 0.8m. Finally, the model is shown to be highly portable and capable of retrieving water depth data with resolution equal to the resolution of the remote sensing image using only a small amount of input water depth data.

To maintain national socio-economic development and maritime rights and interests, it is necessary to obtain the space location information of various ships. Therefore, it is important to detect the locations of ships accurately and rapidly. At present, ship detection is mainly carried out by combining satellite remote sensing imaging with constant false alarm rate (CFAR) detection. However, with the rapid development of satellite remote sensing technology, remote sensing data have gradually begun to show the characteristics of “big data”; additionally, the accuracy and speed of ship detection can be improved by analysing big data, such as by deep learning. Thus, a ship detection algorithm that combines CFAR and CNN is proposed based on the CFAR global detection algorithm and image recognition with the CNN model. Compared with the multi-level CFAR algorithm that is based on multithreading, the algorithm in this paper is more suitable for application to ship detection systems.

Shallowwater depth plays an important role in marine development, navigation safety, and environmental protection. It is an efficient and economical way to obtain water depth by remote sensing technology. At present, most empirical models based on multispectral image usually obtain water depth by the relationship between the sea surface reflectance (SSR) (a single pixel) and in situ water depth, it is a one-to-one correspondence between the reflectance and depth. However, seafloor substrate and inherent optical properties (IOP) will also have contribution to the SSR. In this article, we propose an adjacent pixels multilayer perceptron model (APMLP) model using adjacent pixels to weaken the influence of seafloor substrate and IOP.Datasets on Oahu Island (Sentinel-2B, LIDAR in situ data) and Saint Thomas Island (Sentinel-2A, LIDAR in situ data) are used to establish and verify the model. The APMLP model are also compared with the multilayer perceptron model (MLP) model, BP neural network model, and Log-ratio model. The overall root-mean-square error (RMSE) of APMLP model on Oahu Island is 0.72 m, which is much better than the other three models (MLP 1.07 m, BP 1.05 m, Log-ratio 1.52 m). Similar results are obtained from the Saint Thomas Island dataset, RMSE of APMLP model is 1.56 m, better than the other three (MLP 1.91 m, BP 1.89 m, Log-ratio 2.39 m). The study confirms that considering adjacent pixels in an artificial neural network model can effectively improve the performance of water depth retrieval.

Ship classification based on high-resolution synthetic aperture radar (SAR) imagery plays an increasingly important role in various maritime affairs, such as marine transportation management, maritime emergency rescue, marine pollution prevention and control, marine security situational awareness, and so on. The technology of deep learning, especially convolution neural network (CNN), has shown excellent performance on ship classification in SAR images. Nevertheless, it still has some limitations in real-world applications that need to be taken seriously by researchers. One is the insufficient number of SAR ship training samples, which limits the learning of satisfactory CNN, and the other is the limited information that SAR images can provide (compared with natural images), which limits the extraction of discriminative features. To alleviate the limitation caused by insufficient training datasets, one of the widely adopted strategies is to pre-train CNNs on a generic dataset with massive labeled samples (such as ImageNet) and fine-tune the pre-trained network on the target dataset (i.e., a SAR dataset) with a small number of training samples. However, recent studies have shown that due to the different imaging mechanisms between SAR and natural images, it is hard to guarantee that the pre-trained CNNs (even if they perform extremely well on ImageNet) can be finely tuned by a SAR dataset. On the other hand, to extract the most discriminative ship representation features from SAR images, the existing methods have carried out fruitful research on network architecture design, attention mechanism embedding, feature fusion, etc. Although these efforts improve the performance of SAR ship classification to some extent, they are usually based on more complex network architecture and higher dimensional features, accompanied by more time-consuming storage expenses. Through the analysis of SAR image characteristics and CNN feature extraction mechanism, this study puts forward three hypotheses: (1) Pre-training CNN on a task-specific dataset may be more effective than that on a generic dataset; (2) a shallow CNN may be more suitable for SAR image feature extraction than a deep one; and (3) the deep features extracted by CNNs can be further refined to improve the feature discrimination ability. To validate these hypotheses, we propose to learn a shallow CNN which is pre-trained on a task-specific dataset, i.e., the optical remote sensing ship dataset (ORS) instead of on the widely adopted ImageNet dataset. For comparison purposes, we designed 28 CNN architectures by changing the arrangement of the CNN components, the size of convolutional filters, and pooling formulations based on VGGNet models. To further reduce redundancy and improve the discrimination ability of the deep features, we propose to refine deep features by active convolutional filter selection based on the coefficient of variation (COV) sorting criteria. Extensive experiments not only prove that the above hypotheses are valid but also prove that the shallow network learned by the proposed pre-training strategy and the feature refining method can achieve considerable ship classification performance in SAR images like the state-of-the-art (SOTA) methods.

At present, most machine learning bathymetry retrieval models use the band reflectance as the inversion feature only, without considering features related to the water substrate and pixel spatial correlation. In this study, in addition to band reflectance, two features, Depth-Invariant Index (DII) and pixel location, are taken into account. Two machine learning algorithms, Random Forest (RF) and Back Propagation (BP) neural network are used to retrieve bathymetry. The effects of the two features on the accuracy and performance of bathymetry retrieval are explored. The results show that: (i) Machine learning algorithms are generally superior to the widely used Stumpf model. Stumpf model performs better only in the depth range of 8–16 m, with a Root Mean Square Error (RMSE) of 0.85 m, but has poor performance in other depth ranges. (ii) Compared with models that use Band Reflectance (BR) only, DII_b,g (blue-green DII) + BR model, Location and Location + BR models are all superior to the BR model for RF and BP algorithms. It means that DII and location features are very effective in improving the bathymetry retrieval accuracy.

High-frequency surface wave radar (HFSWR) can detect and continuously track ship targets in real time and beyond the horizon. Compared with transmit/receive (T/R) monostatic HFSWR, the T-R bistatic HFSWR has the advantages of flexibility, receiver concealment, and large coverage because of the separation between the radar transmitter and receiver locations. So far, there is little research on marine ship monitoring for bistatic HFSWR. This article examines ship monitoring performance with T-R bistatic compact HFSWR of small aperture based on measured experimental data. The first-order sea-clutter characteristics and detection blind zone for T-R bistatic HFSWR were investigated theoretically, and the formulas for the moving ships were derived and its position accuracy were analyzed from simulation results. The experimental results of ship target detection of bistatic compact HFSWR carried out in 2015 were presented, and the spectrum characteristics of the bistatic HFSWR were analyzed with measured experimental data. A method integrating detection and tracking was applied to the target detection data collected by the bistatic compact radar, and two targets examples tracked were given. Finally, the validity of the method and the tracing results were verified by using synchronous automatic identification system data, and the ship positioning accuracy of the bistatic compact HFSWR was statistically analyzed. The analysis shows that bistatic radars have larger errors than monostatic T/R radars, with a positioning error as large as 10 km.