As artificial intelligence (AI) increasingly plays a pivotal role in space applications—ranging from satellite operations to autonomous navigation and data analysis—it is crucial to establish a trustworthy framework for its deployment. Trustworthiness in AI is essential to ensure reliability, safety, and integrity in space missions. To achieve this, identifying and defining specific metrics based on explainable mechanisms and theories is fundamental. This approach facilitates transparency in AI decision-making, particularly when using complex models like neural networks. By focusing on interpretability and accountability, researchers can enhance the predictability of AI systems, allowing stakeholders to understand how decisions are made and ensuring that AI solutions align with mission objectives while mitigating risks.

In safety- and mission-critical space systems such as rovers and satellites, Deep Neural Networks (DNNs) are becoming a key component for computer vision tasks, particularly object detection and autonomous navigation. However, since the DNN behavior cannot be assessed through code inspection and analysis, test automation has become an essential activity to gain confidence in the reliability of DNNs.

Since DNNs for computer vision operate on large input spaces (e.g., any possible landscape for a Mars rover) traditional DNN testing approaches leveraging data sets with a limited number of pictures collected in the field might not enable the detection of all DNN limitations. For this reason, simulators, which may generate potentially infinite number of diverse images along with additional information to be used as ground-truth, have become a valid tool for DNN testing. But generating random images might not be the most effective solution.

This talk provides an overview of recent research approaches combining simulators with evolutionary search algorithms, which, driven by appropriate fitness functions, can determine what are the worst-case scenarios for the DNNs under test. In addition, we will show how the lack of fidelity can be effectively addressed by leveraging Generative Adversarial Networks (GANs), to transform the data generated by simulators into realistic input images. Such images can be used both to assess the DNN performance and to retrain the DNN more effectively. The talk will finally provide insights on the results achieved in the context of the ESA projects TIA.

Deep neural networks are increasingly important in safety-and mission-critical systems, including space systems, for example, in their perception layer to analyze images. Unfortunately, there is a lack of methods to ensure the functional safety of DNN-based components, specifically we identify three major challenges with existing practices: 1) situations that are under-represented in the test set may lead to serious safety violation risks but may, however, remain unnoticed; 2) characterizing such high-risk situations is critical for safety analysis; 3) retraining DNNs to address these risks is poorly supported when causes of violations are difficult to determine.

This lessons provides an overview of failure explanation approaches for DNNs that enable the detection of the distinct situations in which a DNN may fail and, consequently, enable the identification of countermeasures. Particular emphasis will be given to analysis pipelines that enable the identification of such situations by identifying cluster of similar failure-inducing images. They combine transfer learning, autoencoders, heatmaps of neuron relevance, dimensionality reduction techniques, and different clustering algorithms. The lesson will provide the results obtained in the context of the ESA project TIA and will include hands-on sessions to experiment with the presented solutions.

The analysis of Deep Neural Networks (DNNs), particularly those used as perception modules, is very challenging due to the networks’ complex and opaque decision-making processes. Multi-modal Vision-Language Models (VLMs) such as CLIP offer an exciting opportunity to interpret the representation space of vision models using natural language. VLMs have been trained on a large body of images accompanied by their textual description, and are thus implicitly aware of high-level, human-understandable concepts describing the images. In this talk, we report on on-going work that seeks to leverage VLMs for the formal analysis and run-time monitoring of requirements expressed in terms of natural-language concepts, as well as debugging of perception modules.

Explore the potential and challenges of Artificial Intelligence for Earth Observation (AI4EO). This session begins with insights from WASDI, showcasing real-world examples and case studies that highlight both the successes and limitations of AI4EO applications. From innovative use cases to persistent hurdles—such as ethical concerns, inflated expectations, algorithmic hallucinations, and reliability issues—we will examine the state of the art and identify key open problems. The lecture concludes with a detailed exploration of a specific case study, offering a balanced perspective on how AI can be leveraged for Earth Observation while addressing its current constraints.

This hands-on session will guide participants through the implementation of a basic downstream Earth Observation (EO) application in the cloud using WASDI. Starting from scratch, we’ll cover the process of algorithm design, data acces, and deployment in a cloud environment. Participants will learn how to leverage WASDI’s tools to simplify development, scale processing, and prototype EO services. By the end of this session, attendees will have the foundational skills to create and deploy their own EO applications, opening the door to scalable research and service-oriented projects.

Anomaly detection is an important problem in data analytics with applications in many domains. In recent years, there has been an increasing interest in anomaly detection tasks applied to time series. In this talk, we take a holistic view of anomaly detection in time series, starting from the core definitions and taxonomies related to time series and anomaly types, to an extensive description of the anomaly detection methods proposed by different communities in the literature. We will then present new benchmarks capturing diverse domains and applications for the purpose of evaluating anomaly detection methods. We will then conclude on Ensembling and Model Selection for time series anomaly detection, discussing different strategies applicable to automatically selecting the appropriate methods for a specific time series.

Anomaly detection is a fundamental task for time-series analytics with important implications for the downstream performance of many applications. In contrast to other domains where anomaly detection mainly focuses on point-based anomalies (i.e., outliers in standalone observations), anomaly detection for time series is also concerned with range-based anomalies (i.e., outliers spanning multiple observations). Nevertheless, it is common to use traditional point-based information retrieval measures, such as Precision, Recall, and Fscore, to assess the quality of methods by thresholding the anomaly score to mark each point as an anomaly or not. However, mapping discrete labels into continuous data introduces unavoidable shortcomings, complicating the evaluation of range-based contextual and collective anomalies. In this talk, we will dive into the existing evaluation measures to properly compare anomaly detection methods for time series. Finally, we will discuss how evaluation measures can be compared and evaluated.

Climate change is increasingly manifested by global increases in the frequency and magnitude of hydrological extremes (i.e., floods and droughts), with severe impacts on economies, livelihoods, and the environment. In a warming climate, both floods and droughts are likely to intensify (i.e., increasing in their duration, frequency, and severity). This can increase the likelihood of rapid shifts or overlap between these opposite hydrological extremes. For instance, in Northern Italy, a big flood affected Emilia-Romagna in May 2023 after the second winter drought in two years. The same region was again affected by floods in September and October 2024, while at the same time, southern Italy was affected by a severe drought. The latter, in turn, caused a significant shrinking of some lakes used for water supply of several municipalities.

High resolution remote sensing data provide important information for flood mapping and lake extent monitoring. Concerning flood mapping, they allow emergency managers to rapidly have a synoptic overview of the extent of the inundation, as well as to monitor its temporal evolution. In this context, synthetic aperture radar (SAR) sensors represent the most commonly used instruments, because they can provide data both day and night and in almost any meteorological conditions.

As for the monitoring of the extent of the water reservoirs, also optical data can be considered as a useful source of information. While they are often unsuitable for inundation mapping because of the frequent cloud cover associated with these disasters, they can be profitably exploited for a long-term lake extent monitoring (e.g., on a monthly basis), at regional or even national scale. This activity is helpful to identify possible situations of water scarcity.

The presentation will illustrate the basic principles of remote sensing of water surfaces from optical and SAR data and will show some applications performed in collaboration with the Italian Department of Civil Protection. For SAR data, the presentation will show examples of the use of both Sentinel-1 and COSMO-SkyMed (1st and 2nd generation). Regarding optical data, the presentation will focus on Sentinel-2.

Satellite data is very useful for evaluating the effects on the ground in the event of extreme events, such as floods, which increasingly affect our territory.
Usually, to search for possible useful acquisitions for emergency management and planning of interventions to help the populations of areas flooded or affected by an extreme event, it is necessary to access various portals and then summarize the best opportunities.
ORBIT allows us to carry out a search for the main satellite acquisitions available (e.g. Sentinel, COSMO SkyMed) on a particular geographical area selected by the user in order to plan those of greatest interest for the case study.

Students will be invited to practice with the ORBIT tool starting from a case study.

Precipitation data derived from satellite observations is a crucial resource for hydrology, meteorology, and climate research, particularly in regions with limited ground-based monitoring. However, satellite precipitation estimates are often limited in spatial resolution, posing challenges for local-scale applications such as flood forecasting, water resource management, avalanche danger forecasting and agricultural planning. To address this, convolutional neural networks (CNNs) have emerged as a powerful tool for downscaling coarse-resolution images to high-resolution grids, enhancing its utility for fine-scale analyses. This lecture reviews the potential of CNN-based downscaling techniques for increasing spatial resolution of satellite derived precipitation data. We discuss the architecture of CNN models, their training processes on diverse datasets, and their ability to capture spatial patterns and intricate features within precipitation fields. Results demonstrate that CNN-based downscaling improves the spatial detail and accuracy of satellite derived precipitation data, underscoring the value of combining satellite data with advanced machine learning techniques to overcome observational limitations in Earth sciences.