Satellite imagery has become a fundamental part for maritime monitoring and safety. Correctly estimating a ship's identity is a vital tool. We present a method based on facial recognition for identifying ships in satellite images. A large ship dataset is constructed from Sentinel-2 multispectral images and annotated by matching to the Automatic Identification System. Our dataset contains 7.000 unique ships, for which a total of 16.000 images are acquired. The method uses a convolutional neural network to extract a feature vector from the ship images and embed it on a hypersphere. Distances between ships can then be calculated via the embedding vectors. The network is trained using a triplet loss function, such that minimum distances are achieved for identical ships and maximum distances to different ships. Comparing a ship image to a reference set of ship images yields a set of distances. Ranking the distances provides a list of the most similar ships. The method correctly identifies a ship on average 60% of the time as the first in the list. Larger ships are easier to identify than small ships, where the image resolution is a limitation.

Unidentified flying objects can be aircraft that do not continuously broadcast ADS-B. They pose a risk for air traffic safety, territorial violation, espionage, etc. In this study, we introduce a method for detecting and estimating the state of aircraft in Sentinel-2 multispectral satellite images. We construct a dataset of 579 ADS-B annotated aircraft from 69 Sentinel-2 images. A CNN is trained on the dataset to estimate the aircraft state vector i.e. position, velocity, heading, altitude. This work allows real-time monitoring of flying objects in satellite images.

The increasing importance of maritime surveillance, particularly in monitoring dark ships, highlights the need for advanced detection models that go beyond simple ship localisation. Current approaches largely focus on either detection or feature extraction, leaving a gap in unified methods capable of providing detailed ship characteristics. This study addresses this gap by developing a unified model for ship detection and characterisation from Synthetic Aperture Radar images, estimating features such as true length, true breadth, and heading. The model is designed to detect ships of varying sizes while simultaneously estimating their characteristics, and experimental results show a high detection accuracy, with a recall of 87.7% and an F1-score of 93.5%. The model also effectively estimates ship dimensions, with mean errors of 1.4 ± 16.2 m for length and 1.5 ± 4.5 m for breadth. Estimating the heading proved challenging for smaller ships, but was accurate for larger ships. A total of 50% of the heading estimates were within 15 degrees of error. This unified approach offers practical benefits for maritime operations. It is especially useful in situations where both ship detection and detailed information are needed, such as predicting future ship positions or identifying ships.

This paper presents an approach to increase the knowledge gained during maritime surveillance using high-resolution ICEYE Synthetic Aperture Radar (SAR) imagery, by estimating the three-dimensional features of vessels. Specifically, the height and mass, alongside traditional two-dimensional length, and width parameters. By analysing the SAR shadow and overlay, we calculate the height of the deck and bridge on the vessel, leading to an estimation of its cargo weight. This enhances maritime surveillance by providing a deeper understanding of vessel characteristics critical for security applications. While we utilize Ultralytics’ YOLOv8 deep learning for initial ship detection, the primary focus is on the detailed estimation of 3D features, a capability not previously demonstrated with SAR imagery.

Synthetic aperture radar (SAR) satellite images are used increasingly more for Earth observation. While SAR images are useable in most conditions, they occasionally experience image degradation due to interfering signals from external radars, called radio frequency interference (RFI). RFI-affected images are often discarded in further analysis or preprocessed to remove the RFI. However, few on-ground radars can cause RFI in SAR images and such information can thus increase domain awareness greatly over both land and sea, where, e.g., localizing and characterizing RFI signals in the ocean could help classify otherwise overlooked ships. The aim of the current study is to detect and localize RFI signals automatically in Sentinel-1 level-1 images and further characterize the on-ground radar. The spatial structure of RFI signals vary greatly. A convolutional autoencoder (CAE) was therefore developed to reconstruct RFI-free Sentinel-1 images. Conversely, RFI-affected images could not be well reconstructed. Anomalous heatmaps were then developed to automatically detect and localize RFI anomalies in the images under varying environmental and geographical conditions, whereafter the external radar characteristics were extracted manually from Sentinel-1 level-0 data. We could consequently classify and localize RFI signals believed to originate from both stationary radars and ship-borne radars. We further argue that the calculated ship-borne radar characteristics correspond to those of air-surveillance radars. Empirically, the method showed better detection results than those of previous studies. Our study shows that more information can be extracted from certain detected objects, such as ships, from SAR images.

This article presents a methodology for localizing radio frequency interference (RFI) signals in Synthetic Aperture Radar (SAR) images acquired from Sentinel-1 SAR satellites. RFI are caused by on-ground radars, and their detection and localization thus provide valuable information for decision makers. In this study, an unsupervised deep learning model based on a Convolutional Autoencoder is used to detect and localize RFI signals in SAR images. The CAE reconstructs the SAR images, excluding RFI signals and other large-scale anomalies. Anomalies are detected by comparing the original images with their reconstructions, and a secondary classification scheme is used to identify RFI signals among the detected anomalies. Results show that the proposed method detects and localizes RFI signals, even in complex regions. The automatic localization of RFI signals in SAR images can enhance various applications such as maritime domain awareness and border surveillance.

SAR satellites are used for monitoring ships worldwide. Moving ships are Doppler shifted by an amount proportional to their velocity. An offset between the ship and its wake is then produced during SAR processing. We present a novel automatic method for calculating the ship velocity. The method relies on multitask deep learning to estimate the offset and ship heading. From these parameters, the ship velocity can be obtained. A convolutional neural network is trained using a coupled loss function. The loss function allows both parameters to be estimated at the same time. We show the methods' effectiveness for ships in Sentinel-1 SAR images. For this purpose, a large dataset of 30,000 AIS annotated SAR ship images is collected. These images have 20 x 22 m pixel resolution and ships do not have a clear wake. The AIS provides the true ship velocity and allows the method to be evaluated. As a result, we can determine the ship speed with an accuracy of 1.1 m/s. The offset disappears near the azimuth direction of the SAR image. Yet, our method is reliable except for ships sailing within 2.5 degrees of the azimuth direction.

Ships may be found by different types of satellite sensors including Synthetic Aperture Radars (SAR), multispectral, Automatic Identification Systems (AIS) and RF electronic support systems. Several methods are described for detecting, classifying, and identify (ID) ships as well as discriminating them from, e.g., icebergs. For example, AIS data can be used in combination with other sensors both for annotation and for finding dark ships, i.e., ships that have turned off their transponder. The sensor fusion, data combination and matching of various sensor data spatially and temporally are then important. We analyze methods and results for various combinations of sensor data, where we use deep neural networks on annotated datasets for detection, classification, discrimination and ID of ships.

Non-cooperative vessels pose a challenge to traditional maritime surveillance systems. To overcome this challenge, alternative surveillance methods such as space-based monitoring sensors have been employed. However, the time-consuming process of satellite downlink hampers near-real-time applications. To address these issues, the use of onboard Artificial Intelligence for direct data processing has emerged as a key technology. This study explores the implementation of a lightweight Synthetic Aperture Radar ship detection model inspired by YOLOv8. The model achieves promising results on an annotated data-set, demonstrating the effectiveness of the approach for detecting both small and large ships. The study investigates the impact of atrous and depth-wise convolutions on the model’s performance and explores model quantization for further size reduction. Our final model has 0.3 million parameters and reached an average procession of 95.4 %. The results highlight the potential of lightweight models for onboard ship detection, offering comparable accuracy to larger models.

Maritime surveillance of the Arctic region is of growing importance as shipping, fishing and tourism are increasing due to the sea ice retreat caused by global warming. Ships that do not identify themselves with a transponder system, so-called dark ships, pose a security risk. They can be detected by SAR satellites, which can monitor the vast Arctic region through clouds, day and night, with the caveat that the abundant icebergs in the Arctic cause false alarms. We collect and analyze 200 Sentinel-1 horizontally polarized SAR scenes from areas with high maritime traffic and from the Arctic region with a high density of icebergs. Ships and icebergs are detected using a continuous wavelet transform, which is optimized by correlating ships to known AIS positions. Globally, we are able to assign 72% of the AIS signals to a SAR ship and 32% of the SAR ships to an AIS signal. The ships are used to construct an annotated dataset of more than 9000 ships and ten times as many icebergs. The dataset is used for training several convolutional neural networks, and we propose a new network which achieves state of the art performance compared to previous ship–iceberg discrimination networks, reaching 93% validation accuracy. Furthermore, we collect a smaller test dataset consisting of 424 ships from 100 Arctic scenes which are correlated to AIS positions. This dataset constitutes an operational Arctic test scenario. We find these ships harder to classify with a lower test accuracy of 83%, because some of the ships sail near icebergs and ice floes, which confuses the classification algorithms.

Maritime activity is expected to increase, and therefore also the need for maritime surveillance and safety. Most ships are obligated to identify themselves with a transponder system like the Automatic Identification System (AIS) and ships that do not, intentionally or unintentionally, are referred to as dark ships and must be observed by other means. Knowing the future location of ships can not only help with ship/ship collision avoidance, but also with determining the identity of these dark ships found in, e.g., satellite images. However, predicting the future location of ships is inherently probabilistic and the variety of possible routes is almost limitless. We therefore introduce a Bidirectional Long-Short-Term-Memory Mixture Density Network (BLSTM-MDN) deep learning model capable of characterising the underlying distribution of ship trajectories. It is consequently possible to predict a probabilistic future location as opposed to a deterministic location. AIS data from 3631 different cargo ships are acquired from a region west of Norway spanning 320,000 sqkm. Our implemented BLSTM-MDN model characterizes the conditional probability of the target, conditioned on an input trajectory using an 11-dimensional Gaussian distribution and by inferring a single target from the distribution, we can predict several probable trajectories from the same input trajectory with a test Negative Log Likelihood loss of −9.96 corresponding to a mean distance error of 2.53 km 50 min into the future. We compare our model to both a standard BLSTM and a state-of-the-art multi-headed self-attention BLSTM model and the BLSTM-MDN performs similarly to the two deterministic deep learning models on straight trajectories, but produced better results in complex scenarios.

Detection of aircrafts in satellite images is a challenging problem when the background is strongly reflective clouds with varying transparency. We develop a fast and effective detection algorithm that can find almost all aircrafts above and between clouds in Sentinel-2 multispectral images. It exploits the time delay of a few seconds between the recorded multispectral images such that a moving aircraft is observed at different positions due to parallax effects. The aircraft speed, heading and altitude are also calculated accurately. Analysing images over the English Channel during fall 2020, we obtain a detection accuracy of 80%, where the most of the remaining were covered by clouds. We also analyse images in the 1.38 μm water absorption band, where only 61% of the aircrafts are detected.

The European Space Agency Sentinel-2 satellites provide multispectral images with pixel sizes down to 10 m. This high resolution allows for fast and frequent detection, classification and discrimination of various objects in the sea, which is relevant in general and specifically for the vast Arctic environment. We analyze several sets of multispectral image data from Denmark and Greenland fall and winter, and describe a supervised search and classification algorithm based on physical parameters that successfully finds and classifies all objects in the sea with reflectance above a threshold. It discriminates between objects like ships, islands, wakes, and icebergs, ice floes, and clouds with accuracy better than 90%. Pan-sharpening the infrared bands leads to classification and discrimination of ice floes and clouds better than 95%. For complex images with abundant ice floes or clouds, however, the false alarm rate dominates for small non-sailing boats.