Search

A major task in any discrimination scenario requires the collection and validation of as many examples as possible. Depending on the type of data this can be a time consuming process, especially when dealing with large remote sensing data such as Synthetic Aperture Radar imagery. To aid in the creation of improved machine learning-based ship detection and discrimination methods this paper applies a type of neural network known as an Information Maximizing Generative Adversarial Network. Generative Adversarial Networks pit a generating and discriminating network against each other. A generator tries to create samples that are indistinguishable from real data whereas the discriminator tries to identify whether a sample is real or generated. Information Maximizing Generative Adversarial Network extend this idea by extracting untangled latent variables as part of the discrimination process which help to classify the data in terms of categories/classes and properties such as ship rotation. Despite the limited size and class distribution of the dataset, the paper showed that the trained network was able to generate convincing samples from the three given classes as well as create a discriminator that performs similarly to state-of-the-art ship discrimination methods despite using no labels for training.

The well-being of the environment is one of the major factors that contributes to sustainability. Sustainable human settlements require local governance to plan, implement, develop, and manage human settlements expansions. This is important as the number anthropogenic activities is directly correlated to the increase in human population within a geographical region. Regional mapping of land cover conversion of natural vegetation to new human settlements is essential. In this paper we explore the effect which the length of a temporal sliding window has on the success of detecting land cover change. It is shown using a short Fourier transform as a feature extraction method provides meaningful robust input to a machine learning method. In theory, the performance is increased by improving the estimates on the features by increasing the length of the sliding window. Experiments were conducted in the Limpopo province of South Africa and were found that increasing the length of the sliding window beyond 12 months yield minor improves due to other seasonal and external factors.

Effective detection and discrimination of surface deformation features in Synthetic Aperture Radar imagery is one of the most important applications of the data. Areas that undergo surface deformation can pose health and safety risks which necessitates an automatic and reliable means of surface deformation discrimination. Due to the similarities between subsidence features and false positives, advanced discrimination methods are necessary in order to obtain meaningful results. Convolutional neural networks have shown to be effective discriminators in other image processing tasks by making use of the spatial relations and underlying characteristics of images in order to classify inputs into classes. This paper presents a Convolutional Neural Network tailored to process interferometric Synthetic Aperture Radar imagery to identify subsidence features. Initial results indicate that the network and its trained kernel weights are able to effectively identify false positives in a small dataset due to careful network selection. Future work includes improving the initial detections to reduce false alarms and making use of multi-channel Synthetic Aperture Radar data directly into the network.

The detection and classification of SAR imaged vessels at sea is a valuable ability for organisations interested in the marine environment or marine vessels. Matching the SAR detected vessels to their AIS messages allows vessels to be identified and context given to their activities. With sparse AIS data, or other identifying geospatial data, an amount of positional uncertainty is introduced that makes matching more difficult and the difficulty is only increased in high traffic areas such as ports or shipping lanes. Different vessel classes can have different behaviours; cargo and tanker vessels tend to move in the most efficient manner between ports while tugs and fishing vessels move in more haphazard tracks while performing their tasks. Using a single method or algorithm to match these different behaviours to SAR detections would result in sub-optimal matching for all classes. In this paper a general method of matching is described and optimised with tailored weights for each vessel class.

Detections and classification of non-AIS-compliant vessels is an important ability for countries or institutions interested in MDA. SAR has been proven to be an effective method but there exists a trade-off between the area that can be imaged and the resolution of each image pixel. Large swath SAR images are a cost effective method of performing maritime surveillance but classification or identification from the images remains a challenge. An algorithm to predict the AIS class of a vessel using historical AIS data and SAR derived features is described in this paper. The novel algorithm calculates the class probability by taking historical AIS data into account using a Bayesian algorithm. Features extracted from SAR imagery are then used with the AIS historical data to provide a list of class probabilities that can enhance other course resolution classification algorithms or to flag vessels that do not conform to historical class behaviour.

Oil spills present a major threat to the sea ecosystem and thus need to be monitored on a regular basis. Synthetic Aperture Radar (SAR) data is well known for ocean monitoring capabilities. SENTINEL 1 (SEN1) extra wide (EW) mode data and RADARSAT-2 (RS2) Maritime Satellite Surveillance Radar (MSSR) modes have been developed to further improve ocean surveillance. This data can monitor large areas (400 km for SEN1 EW and over 500 km for RS2 OSVN), with a finer resolution. These modes enable improved oil slick detection (including ship detection to identify the source) performance while reducing the number of needed scenes. Numerous studies have been proposed for SEN1 data due to its free access while less work has been done on oil spill detection methods using the RS2 OSVN data. In this paper, we evaluated a segmentation-based method on RS2 OSVN data.