Besides their primary, real-time, navigational safety use, historical AIS data has enabled several impactful applications in marine informatics. One possible application, and one that is notably absent from the literature, is the extraction of intelligence regarding the demand/supply balance for freight contracts, which can be used to make informed decisions regarding fleet positioning with respect to expected spikes in freight rates in the spot market. In this paper we report on work-in-progress towards addressing this gap. Specifically, we present a methodology for constructing estimators of freight supply for a given area and time-frame. The application of this methodology to one year’s data from oil tankers in the US Gulf area is presented. If successfully validated on this dataset, the methodology requires minimal supervision and can be extended into a general tool for extracting freight supply estimates.

Kolmogorov-Arnold Networks (KANs) have been recently proposed as a machine learning framework that is more interpretable and controllable than the multi-layer perceptron. Various network architectures have been proposed within the KAN framework targeting different tasks and application domains, including sequence processing. This paper proposes seqKAN, a new KAN architecture for sequence processing. Although multiple sequence processing KAN architectures have already been proposed, we argue that seqKAN is more faithful to the core concept of the KAN framework. Furthermore, we empirically demonstrate that it achieves better results. The empirical evaluation is performed on generated data from a complex physics problem on an interpolation and an extrapolation task. Using this dataset we compared seqKAN against a prior KAN network for timeseries prediction, recurrent deep networks, and symbolic regression. seqKAN substantially outperforms all architectures, particularly on the extrapolation dataset, while also being the most transparent.

Style transfer methods are a powerful visualization tool that can be used to generate counterfactual explanations, plausible alternatives to the original input that leads to a different classification. In this paper we present xSTAE, a system that restyles a misclassified example into the correct class, in order to help the expert understand what patterns the classifier was looking for to assign the correct class, and failed to see in the instance. The system is based on an Autoencoder trained on a loss function that balances between identity loss (similarity with the original instance) and a classification loss derived from a pre-trained classifier, allowing xSTAE to remain completely agnostic with respect to the internals of the classifier it interprets. We present promising experimental results on sleep-stage classification decisions over EEG data, which validate the core of the idea and show future research directions.

Utilizing solar energy to meet space heating and domestic hot water demand is very efficient (in terms of environmental footprint as well as cost), but in order to ensure that user demand is entirely covered throughout the year needs to be complemented with auxiliary heating systems, typically boilers and heat pumps. Naturally, the optimal control of such a system depends on an accurate prediction of solar thermal production. Experimental testing and physics-based numerical models are used to find a collector's performance curve - the mapping from solar radiation and other external conditions to heat production - but this curve changes over time once the collector is exposed to outdoor conditions. In order to deploy advanced control strategies in small domestic installations, we present an approach that uses machine learning to automatically construct and continuously adapt a model that predicts heat production. Our design is driven by the need to (a) construct and adapt models using supervision that can be extracted from low-cost instrumentation, avoiding extreme accuracy and reliability requirements; and (b) at inference time, use inputs that are typically provided in publicly available weather forecasts. Recent developments in attention-based machine learning, as well as careful adaptation of the training setup to the specifics of the task, have allowed us to design a machine learning-based solution that covers our requirements. We present positive empirical results for the predictive accuracy of our solution, and discuss the impact of these results on the end-to-end system.

What sets timeseries analysis apart from other machine learning exercises is that time representation becomes a primary aspect of the experiment setup, as it must adequately represent the temporal relations that are relevant for the application at hand. In the work described here we study wo different variations of the Transformer architecture: one where we use the fixed time representation proposed in the literature and one where the time representation is learned from the data. Our experiments use data from predicting the energy output of solar panels, a task that exhibits known periodicities (daily and seasonal) that is straight-forward to encode in the fixed time representation. Our results indicate that even in an experiment where the phenomenon is well-understood, it is difficult to encode prior knowledge due to side-effects that are difficult to mitigate. We conclude that research work is needed to work the human into the learning loop in ways that improve the robustness and trust-worthiness of the network.

This article proposes a set of concrete technical requirements for trustworthy artificial intelligence methods. Although these requirements do not cover the social, ethical, or regulatory dimensions, they do cover the technical aspects of the complete life-cycle of an AI system, from its design and operational monitoring and control to its behaviour when it fails. The article concludes with an outline for an ambitious but realistic research plan that can advance the state of the art in the direction of reliable AI systems.

Accurate indoor free-space segmentation is a demanding task due to the intricate and dynamic nature of indoor environments. We propose an indoors free-space segmentation method that associates large depth values with navigable regions. Our method employs an unsupervised masking technique that, using positive instances, generates segmentation labels based on textural homogeneity and depth uniformity. Using the estimated free-space masks and a Dense Prediction Transformer (DPT) feature representation, a SegFormer model is fine-tuned on our custom-collected indoor dataset. Our experiments demonstrate sufficient performance in complex scenes where the identification of free space is challenging.

We propose a dual-stream, semi-supervised, attention-based approach that employs feature fusion of RGB and Laser Range Finder (LRF) modalities. Our method lever-ages the strength of two powerful transformer-based networks, i.e. Vision Transformer (ViT) and SegFormer, along with LRF information, to adequately predict whether the scene encountered in the image is safe for a robot to traverse. Towards this effort, we introduce an automated labelling system profiting from the combination of raw velocity readings and laser scanning information. Moreover, we show that overall GOINO-GO detection is enhanced by fusing RGB and laser modalities. Feature fusion is achieved through the employment of a Multi-Head Self-Attention (MHSA) module. Through cross-domain validation, we show that the proposed traversability estimation method can achieve decent amounts of transferability even with limited amount of training data.

Accurate indoor free-space segmentation is a challenging task due to the complexity and the dynamic nature that indoor environments exhibit. We propose an indoors free-space segmentation method that associates large depth values with navigable regions. Our method leverages an unsupervised masking technique that, using positive instances, generates segmentation labels based on textural homogeneity and depth uniformity. Moreover, we generate superpixels corresponding to areas of higher depth and align them with features extracted from a Dense Prediction Transformer (DPT). Using the estimated free-space masks and the DPT feature representation, a SegFormer model is fine-tuned on our custom-collected indoor dataset. Our experiments demonstrate sufficient performance in intricate scenarios characterized by cluttered obstacles and challenging identification of free space.

Training on simulation data has proven invaluable in applying machine learning in robotics. However, when looking at robot vision in particular, simulated images cannot be directly used no matter how realistic the image rendering is, as many physical parameters (temperature, humidity, wear-and-tear in time) vary and affect texture and lighting in ways that cannot be encoded in the simulation. In this article we propose a different approach for extracting value from simulated environments: although neither of the trained models can be used nor are any evaluation scores expected to be the same on simulated and physical data, the conclusions drawn from simulated experiments might be valid. If this is the case, then simulated environments can be used in early-stage experimentation with different network architectures and features. This will expedite the early development phase before moving to (harder to conduct) physical experiments in order to evaluate the most promising approaches. In order to test this idea we created two simulated environments for the Unity engine, acquired simulated visual datasets, and used them to reproduce experiments originally carried out in a physical environment. The comparison of the conclusions drawn in the physical and the simulated experiments is promising regarding the validity of our approach.

Background: Geospatial linked data brings into the scope of the Semantic Web and its technologies, a wealth of datasets that combine semantically-rich descriptions of resources with their geo-location. There are, however, various Semantic Web technologies where technical work is needed in order to achieve the full integration of geospatial data, and federated query processing is one of these technologies. Methods: In this paper, we explore the idea of annotating data sources with a bounding polygon that summarizes the spatial extent of the resources in each data source, and of using such a summary as an (additional) source selection criterion in order to reduce the set of sources that will be tested as potentially holding relevant data. We present our source selection method, and we discuss its correctness and implementation. Results: We evaluate the proposed source selection using three different types of summaries with different degrees of accuracy, against not using geospatial summaries. We use datasets and queries from a practical use case that combines crop-type data with water availability data for food security. The experimental results suggest that more complex summaries lead to slower source selection times, but also to more precise exclusion of unneeded sources. Moreover, we observe the source selection runtime is (partially or fully) recovered by shorter planning and execution runtimes. As a result, the federated sources are not burdened by pointless querying from the federation engine. Conclusions: The evaluation draws on data and queries from the agroenvironmental domain and shows that our source selection method substantially improves the effectiveness of federated GeoSPARQL query processing.

The European fusion research activities have, over recent decades, generated a vast and varied set of data. The volume and diversity of the data that need to be catalogued and annotated make the task of organising and making the data available within a broader environment very challenging. Nevertheless, there are strong scientific drivers as well as incentives and mandates from national research agencies suggesting that a more coherent approach to data referencing, dissemination and sharing would provide strong benefits to the fusion research community and beyond. Here, we discuss the technical requirements and developments needed to transition the current, and future, range of fusion research data to an open and Findable, Accessible, Interoperable, and Reusable data sharing structure guided by the principle 'as open as possible, as closed as necessary'. Here we propose a set of recommendations and technical implementations needed to form a European data sharing environment for the fusion research programmes. Consistency with the emerging IMAS (ITER Integrated Modelling and Analysis Suite) infrastructure is considered to facilitate future deployments.

Background: In geospatial query processing, spatial containment and intersection queries can be efficiently answered from the index. There is, however, a class of queries (such as within-distance) with a semantics that implies that every shape in the database is a potential match and should, in principle, be compared with the threshold. Naturally, this is impractical and optimizations have been developed that efficiently refine the set of candidate shapes before starting to actually compute distances and apply the threshold. In the case of the within-distance queries, many instances can be discarded in advance as too distant. Since geospatial databases organize data as a hierarchy of bounding boxes, this already provided the first direct optimization as the actual distance cannot be smaller than the distance between the bounding boxes. One can easily understand that there are shape configurations that give bounding boxes that are not very selective for near-by shapes. That is, configurations where there are shapes outside the requested distance but within the request distance from the bounding box. Methods: In this article, we investigate a further optimization in addition to and after comparing the bounding boxes, but before computing precise distances. We describe the distance optimizer operation currently used by PostGIS and show how the existing implementation prevails over approaches that use additional approximations. We implement a recursive algorithm to calculate the minimal possible largest inner rectangles of geometries. Results: We observe that the performance of the distance operation cannot be improved by using the inner approximations instead of the actual shapes. The overheads of the inner rectangles would not be recovered from calculating the distance between simpler geometries. Conclusions: The execution time of the distance operator has a small dependence on polygon complexity. Conclusively, an inner approximation for complex polygons cannot out-perform the standard PostGIS implementation.

We present a join optimization method for federated linked geospatial data. This optimization targets within-distance queries in cases where the shapes that are compared are served from different sources of the federation. This operation is computationally expensive, because it cannot be answered from the spatial index of any of the individual data sources in the federation. Our optimization augments the subqueries prepared for each source with additional restrictions on topological relations that must be satisfied and that (a) can be answered from the spatial index of the individual data sources; (b) do not change the semantics of the query. We evaluate our optimization on data and queries derived from a real-world workflow over land usage data. Evaluation shows that our optimization substantially improves query processing time.

We present a method for binary (go/no-go) indoors traversability estimation from 2D images. Our method exploits the power of a pre-trained Vision Transformer (ViT) which we fine-tune on our own dataset. We conduct experiments using a mobile robotic platform to gather image data. Our fine-tuning approach includes the use of a pre-trained Vision Transformer (ViT) en route towards developing a semi-supervised deep learning technique to enhance indoor traversability estimation for scenarios where only a small amount of data is available. We evaluate the accuracy and generalization power of our method against well-established state-of-the-art deep architectures for image classification such as ResNet, and show improved performance.

Traversability illustrates the difficulty of driving through a specific region and encompasses the suitability of the terrain for traverse based on its physical properties, such as slope and roughness, surface condition, etc. In this survey we highlight the merits and limitations of all the major steps in the evolution of traversability estimation techniques, covering both non-trainable and machine-learning methods, leading up to the recent proliferation of deep learning literature. We discuss how the nascence of Deep Learning has created an opportunity for radical improvement in traversability estimation. Finally, we discuss how self-supervised learning can help satisfy deep methods' increased need for (challenging to acquire and label) large-scale datasets.

Background: Geospatial linked data brings into the scope of the Semantic Web and its technologies, a wealth of datasets that combine semantically-rich descriptions of resources with their geo-location. There are, however, various Semantic Web technologies where technical work is needed in order to achieve the full integration of geospatial data, and federated query processing is one of these technologies. Methods: In this paper, we explore the idea of annotating data sources with a bounding polygon that summarizes the spatial extent of the resources in each data source, and of using such a summary as an (additional) source selection criterion in order to reduce the set of sources that will be tested as potentially holding relevant data. We present our source selection method, and we discuss its correctness and implementation. Results: We evaluate the proposed source selection using three different types of summaries with different degrees of accuracy, against not using geospatial summaries. We use datasets and queries from a practical use case that combines crop-type data with water availability data for food security. The experimental results suggest that more complex summaries lead to slower source selection times, but also to more precise exclusion of unneeded sources. Moreover, we observe the source selection runtime is (partially or fully) recovered by shorter planning and execution runtimes. As a result, the federated sources are not burdened by pointless querying from the federation engine. Conclusions: The evaluation draws on data and queries from the agroenvironmental domain and shows that our source selection method substantially improves the effectiveness of federated GeoSPARQL query processing.

In the SPARQL query processing community, as well as in the wider databases community, benchmark reproducibility is based on releasing datasets and query workloads. However, this paradigm breaks down for federated query processors, as these systems do not manage the data they serve to their clients but provide a data-integration abstraction over the actual query processors that are in direct contact with the data. As a consequence, benchmark results can be greatly affected by the performance and characteristics of the underlying data services. This is further aggravated when one considers benchmarking in more realistic conditions, where internet latency and throughput between the federator and the federated data sources is also a key factor. In this paper we present KOBE, a benchmarking system that leverages modern containerization and Cloud computing technologies in order to reproduce collections of data sources. In KOBE, data sources are formally described in more detail than what is conventionally provided, covering not only the data served but also the specific software that serves it and its configuration as well as the characteristics of the network that connects them. KOBE provides a specification formalism and a command-line interface that completely hides from the user the mechanics of provisioning and orchestrating the benchmarking process on Kubernetes-based infrastructures; and of simulating network latency. Finally, KOBE automates the process of collecting and comprehending logs, and extracting and visualizing evaluation metrics from these logs.

ExtremeEarth is a three-year H2020 ICT research and innovation project. Its main objective is to develop Artificial Intelligence and big data technologies that scale to the large volumes of big Copernicus data, information and knowledge, and apply these technologies in two of the European Space Agency (ESA) Thematic Exploitation Platforms (TEP): Food Security and Polar.