AlphaZero achieves superhuman performance through pure self-play without human expertise, but its dependence on sparse terminal rewards limits learning efficiency.This paper investigates integrating potential-based reward shaping into AlphaZero to accelerate learning while preserving optimality.We address whether reward shaping improves sample efficiency without compromising final performance, and which integration methods prove most effective.We present two implementation approaches: search-time shaping and auxiliary network heads, each targeting different components of the learning process.Experimental evaluation on Othello provides initial evidence of benefits, with ongoing work on comprehensive performance characterization across diverse environments.

Threat evaluation offers significant operational advantages in military and non-military contexts by reducing risks to personnel and enhancing the situational awareness. The accurate evaluation of threats, requires a comprehensive analysis of the behaviors of the agents. This paper presents a supervised learning approach to predict the intentions of agents in a simulated military environment, focusing on binary classification to determine whether an agent poses a threat or not. The model integrates multimodal data, including spatial information from a grid-based map and features related to agents, such as velocity and weapon possession. The temporal aspect of agents is considered. However, this yields limited improvements in prediction accuracy. The model is evaluated using self-generated wargaming data, and results show that deep learning approaches leveraging spatial-temporal data outperform traditional methods like Random Forest models, achieving an AUC score of 0.84. The proposed approach demonstrates the potential of using multimodal data fusion for improving threat identification. Future work will focus on expanding the diversity of scenarios and further enhancing the realism of data generation.

This paper addresses a critical issue in seabed characteri-zation with deep learning semantic segmentation using high-resolution Synthetic Aperture Sonar (SAS) data, that we call Catastrophic Receptive Field Overflow (CRFO). We propose novel methods, including Mosaic Augmentation and Homogeneous Patch Rejection, to (1) effectively mitigate CRFO and (2) enhance model performance. Through experiments on real-world SAS data, we investigate the origins of CRFO, revealing its dependence on model architectures and data characteristics. The presented solutions exhibit promising results, whether measured in terms of Overall Accuracy or the reliability of models in inference across various image input sizes or aspect ratios, in the face of new proposed metrics. These findings provide valuable insights for addressing CRFO challenges in tasks involving relatively homogeneous datasets.

Seabed characterisation consists in the study of the physical and biological properties of the bottom of the oceans. It is effectively achieved with sonar, a remote sensing method that captures acoustic backscatter of the seabed. Classical Machine Learning (ML) and Deep Learning (DL) research have failed to successfully address the automatic mapping of the seabed from noisy sonar data. This work introduces the Deep Supervised Semantic Segmentation model for Seabed Characterisation (D4SC), a novel U-Net-like model tailored to such data and low-label regime, and proposes a new end-to-end processing pipeline for seabed semantic segmentation. That dual contribution achieves state-of-the-art results on a high resolution Synthetic Aperture Sonar (SAS) survey dataset.

Following the launch of the Copernicus Sentinels, which has enabled the free access to petabytes of satellite data, change point analysis has caught the attention of the remote sensing community. In fact, the exploitation of satellite image time series has a number of advantages over the use of an image pair, as it allows a better understanding of how the process under study is evolving. Although this is a well-known area of research that spans different application domains, the majority of the change point detection methods have been designed for the analysis of univariate signal, and only a few of them can be used to process multidimensional data. In this paper, we present a novel change point detection method based on the combination of wavelets and mathematical morphology for the analysis of data cubes of satellite image time series that allows the user to reduce the data dimensionality at the input level. We conducted a preliminary performance assessment on 50 sites in Belgium using up to 5 different input features derived from Sentinel-1 and Sentinel-2 data.

The all-weather capabilities of SAR satellite sensors make them a powerful tool for the regular and consistent monitoring of large areas of interest. In particular, they could be used to gather information that would be otherwise difficult to obtain with in-situ campaigns, as for instance when it comes to assess the extent of damages in the aftermath of disastrous events. In this paper we show how the PELT change point detection algorithm can be used to analyze time series of Sentinel-1 images for rapid damage mapping in areas of conflict. Despite the sensor’s limited spatial resolution, the results show that, even with relatively short time series and using both polarizations, it is possible to achieve satisfactory detection rates.

The Walloon Region is currently managing a database of more than 2000 ‘redevelopment sites', i.e. urban sites that were previously used for industrial activities and/or housing and that are now abandoned. The administration needs to keep this inventory up-to-date so that the necessary urban planning could be done; however, at the moment, this information is obtained via time-consuming on field campaigns. Thanks to the launch of the Copernicus programme, free satellite data are now provided at high temporal resolution, and new monitoring approaches can be implemented. Leveraging a well-established changepoint detection method, this paper shows some preliminary results on how time series of Sentinel-1 and Sentinel-2 data could be jointly used to automatically detect changes in urban areas, thus providing the Walloon Region with a tool that can be exploited for a more efficient management of the ‘redevelopment sites'.

The “Compagnie Sucrière Sénégalaise” (CSS) wanted to upscale the field-level crop simulation model AquaCrop (FAO's agro-meteorological model) for the automated monitoring and management of its ±13.000 ha of irrigated sugarcane. A recently developed changepoint detector was applied to Sentinel-1 time series to identify key phenological crop stages for assimilation in AquaCrop. Field-specific emergence dates, varying from 10 to 45 days after planting, were assimilated in AquaCrop. Simulated sugarcane biomass had an R2 of 0.7 and an RMSE of 6.4%. The improved management support system was also able to identify potential irrigation mismanagements.

Sugarcane is the world's largest crop by production quantity, as reported by the Food and Agriculture Organization of the United Nations. Its growth cycle has a duration of about 12-14 months, and the same plantation can be generally harvested up to 7 times before replanting is needed. In order to both predict the yield and optimize the production processes, sugarcane industries need to be regularly updated on the harvest progress; however, they mainly rely on direct communications from farmers, and this has evident limitations.In this paper we present a method that exploits stacks of Sentinel-1 images for the automatic detection of sugarcane harvest dates. The method has been used to monitor, over a period of 21 months, a large cultivated area in Northern Senegal that comprises 719 sugarcane parcels.The results have shown that the method performs well in terms of both detection and estimation accuracy, and has the potential to be operationally used in the sugarcane production processes.

A methodology of intercalibration of C-band spaceborne scatterometers is developed and applied to European Remote Sensing satellite-1 (ERS-1), European Remote Sensing satellite-2 (ERS-2), and Meteorological Operational satellite (METOP) scatterometer data. Assuming that the differences between the instruments can be represented by an incidence-angle-dependent bias, this paper presents and discusses four methods, providing an estimate of that bias and of its standard deviation. Model-based methods performed more accurately than a direct comparison of σ0. The latter provides a systematic larger positive bias. The methodology is applied to ERS-1 and ERS-2 data acquired during the tandem mission in 1996. The same methodology is applied to ERS-2 and Advanced Scatterometer (ASCAT) data acquired in December 2008. Generally, the bias between the ERS-1 and ERS-2 scatterometers is smaller than 0.2 dB over most incidence angles, and the four methods provide relatively consistent results. The bias between ERS-2 and ASCAT is slightly higher, reaching 0.4 dB at certain incidence angles. These results suggest that these scatterometers need to be intercalibrated to achieve a consistent backscatter data.

Inductance-based self-sensing methods allow to estimate the rotor position by tracking the magnetic anisotropy linked to the rotor. The inductance can be quickly estimated by measuring the response on the injection of high-frequency signals in addition to the signals of normal-operating control. Compared to more conventional permanent-magnet synchronous machines, driving a low-inductance brushless DC-machine presents some issues that must be taken into account: significant inverter nonlinearities, impact of the stator resistance on the inductance estimation, important harmonic content in the magnetic field that leads to errors in the rotor-position estimation. In this paper, we propose a signal-injection solution that removes the stator-resistance impact. A solution to the inverter nonlinearities is also discussed. Experimental results demonstrate the robustness of the control scheme at low and higher speeds.

The importance of long-term, continuous, and homogenous time-series of satellite data is widely accepted and strongly fostered by the international scientific community. The various global projects and initiatives undertaken in the last few years are evidences of that effort. Among those are: the Long Term Data Preservation Working Group [1], the Permanent Access to the Records of Science in Europe (PARSE) [2], or the Global Climate Observing System (GCOS) [3]. One of the examples of long-term monitored variable is the wind vector. Since the European Remote-sensing Satellite (ERS)-1 launch in July 1991 and until ERS-2 decommissioning in July 2011, a continuous and consistent database of backscattering signal from the Earth surface has been built, and is now available. The Active Microwave Instrument (AMI) [4], which was one of the ERS-1 and ERS-2 payloads, provided radar backscattering coefficient measurements during the last 20 years by using its three nominal operational acquisition modes: Synthetic Aperture mode (SAR mode), Scatterometer mode (wind mode) and a special combination of the two over ocean where SAR and Scatterometer mode are interleaved (wind/wave mode). The main applications for data acquired in Scatterometer mode is related to the estimation of the wind vector over the sea surface. In that field the ERS-2 Scatterometer measurements give a very valuable contribution to the accuracy of the numerical weather forecast models, being assimilated in several meteorological weather forecast centers since the beginning of the mission. After the decommissioning of ERS-2, effort has been devoted to achieve a complete reprocessed database, including both ERS-1 and ERS-2 acquisitions [5]. The cross-calibration between these two satellites is a crucial task to obtain the homogeneousness of the wind vector database, and allow its long-term characterization. The approach followed by ESA in term of team organization, cross-calibration strategy and validation methodology towards this goal is presented in this paper as well as the preliminary results of the long-term characterization of the wind vector.

This paper presents an overview of the evolution of the European Remote-sensing Satellite (ERS)-2 scatterometer mission during the last 16 years, highlighting the changes in both satellite configuration and on-ground data processing algorithm. Instrument and on-ground data processor performances and evolutions are analyzed and commented; finally, future developments are emphasized. ERS-2 was launched in 1995 by the European Space Agency (ESA). Since then, the active microwave instrument, which is one of the ERS-2 payloads, is providing radar backscattering coefficient measurements by using its three nominal operational acquisition mode: synthetic aperture mode (SAR mode), scatterometer mode (wind mode), and a special combination of the two over ocean where SAR and scatterometer mode are interleaved (wind/wave mode). The main applications for data acquired in scatterometer mode are related to the estimation of the wind vector over the sea surface. In that field, the ERS-2 scatterometer measurements give a very valuable contribution to the accuracy of the numerical weather forecast models, being assimilated in several meteorological weather forecast centers since the beginning of the mission. Other applications of the ERS-2 scatterometer data are over land to retrieve information about the soil water content and over the sea-ice. A constant monitoring of the scatterometer performances is carried out since the beginning of the mission by ESA engineering teams located in ESTEC and ESRIN and the instrument manufacture (Dornier at launch time), in collaboration with several European research institutions, as the European Centre for Medium-range Weather Forecasts for product geophysical validation, the Belgian Royal Military Academy for data processing and calibration during the zero-gyro phase, and industrial partners, as Serco SpA for the routine data quality control activities since the beginning of operational phase. Results show outstanding performances even after the failure of several hardware components that has been properly compensated on-ground with evolution of the processor, and many years of operation, which permits the creation of a homogeneous database of wind vectors for the last 16 years (20 years if the ERS-1 mission is considered), in accordance with Global Climate Observing System recommendations.

We present a new method to estimate the clutter-plus-noise covariance matrix used to compute an adaptive filter in space-time adaptive processing (STAP). The method computes a ML estimate of the clutter scattering coefficients using a Bayesian framework and knowledge on the structure of the covariance matrix. A priori information on the clutter statistics is used to regularize the estimation method. Other estimation methods based on the computation of the power spectrum using for instance the periodogram are compared to our method. The result in terms of SINR loss shows that the proposed method outperforms the other ones.