Committee:
     PRESIDENT: LÓPEZ MORENO, CARMEN
     SECRETARI: DUFFO UBEDA, NURIA
     VOCAL: BATLLÓ ORTIZ, JOSEP
Thesis abstract: Earthquakes are among the most destructive natural disasters, causing significant infrastructure damage and casualties. Between 1998 and 2018, seismic events resulted in approximately 846000 fatalities and caused economic losses totaling US$661 billion, emphasizing their profound socioeconomic impact. While thousands of earthquakes occur each year globally, only a small number are significant enough to be detected by monitoring systems or felt by people. Although earthquakes cannot be prevented, efforts have been made to reduce their consequences through risk assessment and public preparedness initiatives. Despite these advances, a reliable early warning system for earthquakes remains insufficient. The absence of consistent, deterministic precursors to seismic events is a critical challenge in developing such systems. However, research has identified small detectable geophysical signal anomalies that may appear days to weeks before major earthquakes. These anomalies involve changes in thermal infrared emissions, ionospheric scintillation, disturbances in magnetic fields, etc. While these signals are not usually present, their detection could enhance forecasting capabilities. Remote Sensing (RS) is a promising technique that provides broad spatial coverage, high temporal resolution, and the capability to observe otherwise inaccessible areas, such as oceans, deserts, and mountains. RS systems allow for continuous monitoring of the Earth's surface and atmosphere, enabling the detection of potentially earthquake-related anomalies across the lithosphere, atmosphere, and ionosphere. This Ph.D. thesis studies the use of RS techniques for earthquake precursor detection and their recent advancements. It explores Lithosphere-Atmosphere-Ionosphere Coupling (LAIC), a multidisciplinary framework that explains how seismic stress and rock deformation in the lithosphere can trigger cascading effects in the atmosphere and ionosphere. The thesis also presents results from ongoing research into short- and medium-term earthquake forecasting using Earth observation data. The Ph.D. thesis examines several satellite-derived parameters: Land Surface Temperature (LST) anomalies, Ionospheric Scintillation (IS) indices, geomagnetic field variations, and space weather data. This work aims to contribute to understanding earthquake precursors and seeks to develop practical tools for better predicting seismic events, enhancing mitigation and early warning efforts.

Committee:
     PRESIDENT: BURGOS GARCÍA, MATEO
     SECRETARI: JOFRE ROCA, LUIS
     VOCAL: ALVAREZ PEREZ, JOSE LUIS
Thesis abstract: The SAR satellite missions from Geostationary Orbits (GEOSAR) offer short revisit times of less than 24 hours, which are essential for monitoring fast-evolving phenomena. A first mission has recently been launched in China, and other missions are being studied or proposed in Europe and the USA, with applications ranging from water cycle monitoring to terrain stability assessment. All of these missions share the need for precise orbit determination at Geostationary Orbits (GEO) to enable the formation of well-focused SAR images.By employing Field-Programmable Gate Array (FPGA) technology, digital correlators have been designed to be used in an interferometric orbit determination system. The system has been validated for obtaining orbital estimates using phase observables from the correlator outputs, which can be used to determine the satellite's trajectory. The implemented system tracks TV broadcasting telecommunication satellites, using the transmissions as signals of opportunity (SoOp). Urban reflectors are utilised to form large interferometric baselines with respect to the developed receiver.Precise orbital determination is crucial for Synthetic Aperture Radar (SAR). Simulation studies were conducted for a monostatic configuration and a ground-based bistatic configuration to assess the impact of orbital parameter estimation errors on SAR image focusing. An autofocusing technique based on entropy minimization optimisation has been studied to improve the focusing of the GEOSAR images affected by residual orbital errors, allowing for the relaxation of orbit precision requirements to the order of hundreds of metres. Using the same transmitters of opportunity as in the interferometric system, a ground-based SAR (GB-GEOSAR) passive system has been designed and implemented. A multichannel coherent receiver is employed to capture echo data from a close by urban scene. A range compression module, based on the design of the interferometric correlators, is employed to process the echo data by correlating it with the main satellite signal, which acts as a reference for the matched filter process. Azimuth compression, based on the orbital data from the interferometric system, is then applied to obtain SAR-focused images. These images can be used to study GEOSAR concepts and serve as proof of concept for the capabilities of a passive system, which could be employed for monitoring local regions with low deployment costs.Interferometric and polarimetric techniques can be applied to GB-GEOSAR images, which could be used in topographic or earth deformation mapping and surface type classification. An experimental evaluation based on images of urban and mountainous areas near the university campus has enabled the identification of the challenges, potentials, and limitations of the developed system.

Committee:
     PRESIDENT: MARTIN NEIRA, MANUEL
     SECRETARI: DUFFO UBEDA, NURIA
     VOCAL: FORTUNY GUASCH, JOAQUIM
Thesis abstract: Radio Frequency Interference (RFI) has emerged as the most critical challenge to the effective exploitation of the microwave spectrum for Earth observation. RFI consists of non-natural emissions that obscure the radiometric signals measured by microwave radiometers. The presence of RFI degrades the accuracy and completeness of radiometric measurements. This is evident, for example, in the European Space Agency’s (ESA) Soil Moisture and Ocean Salinity (SMOS) mission, which, despite operating in the protected L-band, is severely affected by interference. This PhD thesis presents innovative approaches to mitigate this issue, with a particular focus on Synthetic Aperture Interferometric Radiometers, which have been insufficiently studied in the context of RFI.This PhD thesis introduces several novel RFI detection and mitigation techniques, including autocorrelation-based detection, non-linear decomposition using Empirical Mode Decomposition (EMD), and data-adaptive mitigation via the Karhunen-Loève Transform (KLT). These methods address some of the limitations of existing approaches and have been evaluated with simulated data using standard performance metrics. Furthermore, a new algebraic interpretation of mitigation is provided, bridging the gap between RFI detection and signal transformation methods, and introducing novel performance evaluation metrics.In addition to the theoretical contributions, this thesis includes experimental validation of the proposed techniques using real-world data. The practical feasibility and performance of whiteness-based detection and KLT-based mitigation techniques are demonstrated in realistic scenarios. The research also explores the effects of signal quantization on mitigation techniques, a relevant concern for digital radiometer implementations.

Committee:
     PRESIDENT: MERINO MARTINEZ, MARIO
     SECRETARI: GARCÍA RIGO, ALBERTO
     VOCAL: ÁLVAREZ LAGUNA, ALEJANDRO
Thesis abstract: Spacecraft entering planetary atmospheres are enveloped by a plasma layer with high levels of ionization, caused by the extreme temperatures in the shock layer. The charged particles in the plasma interact with electromagnetic waves emitted by onboard antennas, resulting in communication and tracking difficulties during flight and causing radio communication blackouts that can last several minutes. The magnetic field alleviation technique is a proposed solution to reduce the plasma layer effects on radio signal propagation. This technique involves superimposing a magnetic field onto the plasma flow, transforming it into an anisotropic medium and altering its refractive index. This enables the creation of an extraordinary wave capable of propagating at plasma frequencies higher than the radio signal frequency.The Horizon 2020 Magnetohydrodynamic Enhanced Entry System for Space Transportation (MEESST) project aims to design and test a high-temperature superconducting (HTS) magnet to mitigate radio communication blackout and reduce surface heat flux. This research focuses on the experimental study of the radio communication blackout phenomenon and its mitigation using an applied magnetic field, in the VKI plasma wind tunnel. A tailored communication system, designed to accommodate the ionization levels encountered in ground testing, is developed and fully characterized in the UPC anechoic chamber.The results are divided into three main parts, each addressing a specific research question. The first part is dedicated to studying signal propagation in a plasma flow. Both air and CO2 plasma flows are characterized using emission spectroscopy. Radial temperature profiles are derived from the intensity of an oxygen atomic line, while electron number density profiles are determined from the Stark broadening of the Hb line. Radio signal magnitude profiles are measured in various configurations (side-to-side and stagnation) across a frequency range of 33 GHz to 40 GHz, under a wide range of testing conditions. Attenuation and Faraday rotation caused by the plasma are estimated, and the results show a strong correlation with the measured plasma frequencies.The second part examines the effects of an applied magnetic field on the plasma flow and radio signal propagation. The probe houses the HTS magnet, which can generate a maximum magnetic flux density of approximately 0.45 T at the front of the antenna. Characterization of the magnetized air plasma flow reveals an increase in radiance, stagnation heat flux, and plasma temperature and frequency, attributed to the applied magnetic field. The higher electron density near the antenna and the increased attenuation of the extraordinary wave result in negligible differences in signal attenuation profiles. However, changes are observed in the reflection coefficient depending on the magnetic field strength, and in the Faraday rotation depending on the field direction.The third part explores methods to duplicate flight conditions in ground facilities, concerning aerothermochemistry and radio communication blackout. A relevant condition related to radio blackout during the ARD reentry flight is selected, at an altitude of 60.6~km. The shock layer around an ARD model is replicated in the JAXA HIEST shock tunnel. The condition is numerically simulated for both flight and ground scenarios, highlighting the limitations of the binary scaling methodology. For the same flight condition, the local heat transfer simulation methodology is applied and tested in the Plasmatron facility, successfully replicating the same flow conditions. Finally, the electron number density profiles from both facilities are compared with those from flight.

Committee:
     PRESIDENT: ROMAN DIEZ, ROBERTO
     SECRETARI: MUÑOZ PORCAR, CONSTANTINO
     VOCAL: HERRERA, MILAGROS
Thesis abstract: The improvements in retrieving microphysical, optical, and radiative properties of aerosols can contribute to more detailed local observations and research, support local or global climate modeling and air quality forecasting, validate satellite products, and contribute to public policies. This PhD thesis aims at achieving such improvements through synergies between ground-based remote-sensing instruments. The objectives are to (1) evaluate the potential of an inversion algorithm, namely GRASP, for retrieving aerosol properties from the synergy between a polarized Sun-sky-lunar photometer and a state-of-the-art ACTRIS/EARLINET multi-wavelength lidar, and (2) estimate long-term aerosol radiative effects by combining net radiative fluxes from pyranometers and Aerosol Optical Depth (AOD) from photometers, and assessing GRASP’s ability to retrieve such effects on a case-by-case basis. To achieve Obj. 1, sensitive tests (noise-free and random noise retrievals) were performed with GRASP for several combinations of ground-based observations, including polarization information as GRASP inputs, for three aerosol scenarios in Barcelona, Spain. The sensitive tests showed that the addition of the Degree of Linear Polarization (DoLP) improved the noise-free and random-noise GRASP inversions, particularly under higher AOD. The gains were most notable for the coarse mode of the optical properties, and for the coarse mode of the Real Refractive Index. To achieve Obj. 2, a direct method was applied to estimate spectral Aerosol Forcing Efficiency (AFE) and Aerosol Radiative Forcing (ARF) for a 14-year database in Barcelona, Spain. The AFE-ARF estimations confirmed a dominant cooling forcing over Barcelona, with the strongest cooling effect caused by mineral dust, followed by urban/industrial-biomass burning and mixed aerosols. Furthermore, the GRASP inversions with DoLP were more accurate for the ARF, agreeing with the direct method in the high-AOD cases.

Committee:
     PRESIDENT: VIPIANA, FRANCESCA
     SECRETARI: O\'CALLAGHAN CASTELLA, JUAN MANUEL
     VOCAL: ZENG, XUEZHI
Thesis abstract: This thesis develops and validates an innovative algorithm for colorectal cancer (CRC) detection using microwaves, integrating it into endoscopic procedures. Validation includes ex-vivo, in-vivo, and clinical trials. The algorithm enhances polyp detection by exploiting dielectric contrast between healthy and cancerous tissues, addressing colonoscopy limitations. It normalizes signals, calibrates artifacts, reconstructs images, and applies automatic detection to improve sensitivity and specificity.Key calibration strategies include First Frame Subtraction, Similar Frame Subtraction, Antenna-Pair Subtraction, and Previous Frames Averaging to isolate polyp scattering. Automatic detection methods, such as Exponential Smoothing and Target-to-Clutter Ratio Thresholding, enable polyp identification regardless of anatomy. The algorithm undergoes extensive preclinical testing with phantom models, ex-vivo tissues, and in-vivo pig trials, evaluating detection accuracy through sensitivity, specificity, and mismatch metrics. A pilot clinical trial validates real-world performance.Optimization involves refining a phantom model to simulate colonoscopy conditions and tuning algorithm parameters for a second clinical trial. These advancements position microwave-based detection as a promising CRC screening alternative. The research demonstrates the potential of microwave imaging to improve diagnostics, offering real-time, automated polyp detection with meaningful clinical impact.

Committee:
     PRESIDENT: SCAGLIOLA, MICHELE
     SECRETARI: MALLORQUI FRANQUET, JORDI JOAN
     VOCAL: MAKHOUL VARONA, EDUARDO
Thesis abstract: The capability to measure ocean surface topography from space emerged in the 1970s with satellite radar altimeters. Since then, it has become essential in Earth observation, supporting a range of applications, from coastal water-level monitoring and sea-ice elevation, to quantifying sea level rise. Advances in radar altimetry have significantly improved the along-track resolution, progressing from kilometre scale in Low Resolution Modes to approximately 300 metres with Synthetic Aperture Radar (SAR) techniques using delay/Doppler processing, and even metre-scale resolution with Fully Focused SAR (FF-SAR). While Low Resolution and delay/Doppler algorithms are now operationally mature, FF-SAR algorithms remain relatively new and computationally demanding, limiting their operational use. This thesis introduces two novel algorithms designed to enhance the computational efficiency of FF-SAR radar altimetry processing, making it feasible for current and future satellite missions.The first alternative is a frequency based algorithm, named Omega-K Closed-Form algorithm, providing up to 4000 times of improvement in computational efficiency compared to the classic FF-SAR Back-projection algorithm. The algorithm is validated using data from point targets and open ocean. Additionally, the Omega-K algorithm is applied to estimate wavelength of long traveling ocean waves.The second alternative is an Accelerated Back-projection algorithm, which significantly reduces the computational runtime of classic Back-projection on a CPU-based architecture by a factor of 28, and a factor up to 13 on a GPU-based architecture. With respect to the classic Back-projection on a CPU, combining GPU processing with the Accelerated Back-projection achieves an overall computational improvement up to 1500. This method preserves accuracy and enables near-real-time processing, validated extensively with transponder and open ocean data. The Accelerated Back-projection algorithm is slower than the Omega-K algorithm, but it is more versatile, robust, accurate, and precise.Additionally, an alternative approach to delay/Doppler processing is introduced and extensively evaluated. This method, named Sub-looked Back-projection algorithm, is a variation of the classic Back-projection method. It generates waveforms with a resolution of 300 metres, comparable to those obtained from conventional delay/Doppler methods, but without relying on the typical approximations commonly employed by standard delay/Doppler processors to simplify computations. Furthermore, a comparison between delay/Doppler algorithms and the Sub-looked Back-projection method is conducted, assessing biases as a function of Significant Wave Height (SWH) in selected regions. Results indicate that the Sub-looked Back-projection method reduces SWH biases by 20% compared to conventional delay/Doppler processing. When implemented on GPU architectures, this method achieves processing speeds comparable to current operational delay/Doppler processors.Collectively, these methods represent a substantial step toward faster, as accurate as possible, and more versatile radar altimetry processing, paving the way for the next generation of operational satellite missions.

Committee:
     PRESIDENT: HERNÁNDEZ GUTIERREZ, JOSÉ ALBERTO
     SECRETARI: VELASCO ESTEBAN, LUIS DOMINGO
     VOCAL: NATALINO, CARLOS
Thesis abstract: The increasing complexity and demands of modern telecommunications networks necessitate the development of autonomous and secure systems to ensure efficient, reliable, and secure communications. The integration of advanced technologies such as Artificial Intelligence (AI) and Machine Learning (ML) together with Quantum Key Distribution (QKD) into optical networks addresses these needs. This integration enables the creation of networks that can efficiently automate their operations while ensuring the highest standards of security. In this context, this thesis explores the use of Software Defined Networking (SDN) for the advancement of autonomous and secure optical networks, in particular Elastic Optical Networks (EONs). The research focuses on enhancing network efficiency and security to meet the growing complexity and demands for high-capacity, low-latency, and secure communications.The PhD thesis investigates the application of ML, specifically Deep Reinforcement Learning (DRL) and Graph Neural Networks (GNN) to tackle key challenges in the management and optimization of EONs. The primary goal is to develop autonomous and intelligent solutions for dynamic service provisioning, resource allocation, and spectrum management. A significant contribution of this work is the development of novel DRL-based approaches for Routing and Spectrum Assignment (RSA). These methods are designed to adaptively manage network resources in real-time, overcoming the limitations of traditional, static RSA algorithms. By considering latency as a key factor, the DRL-based RSA mechanism ensures the efficient provisioning of latency-sensitive applications and improves overall network performance metrics, such as latency and throughput. The thesis also examines the dynamic provisioning and optimal placement of Virtual Network Functions (VNFs) using DRL and GNNs. This combination of technologies enables a more efficient mapping of resource requirements to the physical infrastructure, facilitating scalable and flexible network management systems.The research also includes an experimental validation of the proposed solutions. A proof-of-concept (PoC) was implemented to demonstrate the integration of DRL models within an SDN control plane framework. This involved externalizing path computation to a dedicated entity that assists the SDN controller in the path and spectrum selection function. The experimental results confirmed the practical applicability of the DRL approach in supporting selected control functions in operational EON infrastructures.Furthermore, the research explores the coexistence of Continuous Variable Quantum Key Distribution (CV-QKD) and classical channels within EONs, which is essential for ensuring secure communications in the quantum computing era. To address the challenge of noise interference from high-power classical channels on sensitive quantum channels, the thesis introduces dynamic spectrum allocation strategies leveraging SDN. These strategies optimize the use of spectrum resources and minimize noise interference, ensuring secure and efficient operation of the integrated network.In summary, this thesis provides significant advancements in the field of autonomous and secure optical networks by integrating advanced ML techniques, contributing to the development of agile, high-capacity, reliable, and secure EONs for future telecommunications.

Committee:
     PRESIDENT: GAMBA, PAOLO ETTORE
     SECRETARI: VALL-LLOSSERA FERRAN, MERCEDES MAGDALENA
     VOCAL: MARCELLO RUIZ, JAVIER
Thesis abstract: The increasing availability and miniaturization of satellite sensors have enabled CubeSats to emerge as powerful platforms for Earth Observation (EO). Despite their reduced size and low cost, CubeSats can provide high-resolution geophysical information, provided that their intrinsic limitations, limited onboard power, processing capacity, storage, and downlink bandwidth, are effectively addressed. This thesis explores the development and implementation of efficient data processing methodologies that can operate under such constraints. By shifting part of the EO processing pipeline from the ground segment to the satellite itself, the goal is to enable autonomous, real-time analysis onboard, thus reducing the volume of data transmitted and increasing the timeliness and scientific value of satellite products.To this end, the thesis presents a series of algorithmic contributions spanning both traditional signal processing techniques and machine learning approaches, including neural networks. First, a lightweight deblurring methodology is proposed to mitigate image degradation caused by Attitude Determination and Control System (ADCS)-induced jitter in high-resolution optical CubeSat imagery. The method relies exclusively on inertial telemetry, estimating a spatially variant blur kernel and performing regularized deconvolution to restore image sharpness with minimal computational overhead.Second, the work introduces a two-stage neural network architecture to retrieve Sea Ice Concentration (SIC) and Sea Ice Extent (SIE) from L-band radiometric, and GNSS-Reflectometry (GNSS-R) data acquired by the FSSCat mission. The first network estimates coarse SIC/SIE maps using radiometry and auxiliary variables, while the second enhances resolution over GNSS-R reflection points, effectively combining the complementary strengths of both sensors.Third, the thesis proposes the Sequential Band Selection and Ranking (SBSR) algorithm for hyperspectral dimensionality reduction. This unsupervised method employs simple, interpretable metrics, entropy and spectral correlation, to iteratively select bands that maximize information content while minimizing redundancy. Its low computational complexity makes it well-suited for onboard execution.Finally, the SBSR approach is extended through a Convolutional Neural Network (CNN)-based model capable of performing band selection directly on coarsely aligned hyperspectral data, as typically encountered in pushbroom sensors onboard CubeSats. The CNN processes cross-correlation and autocorrelation matrices to produce band relevance scores, enabling robust selection without prior spatial co-registration.All methods were validated using real-world datasets from the PRISMA and FSSCat missions, as well as benchmark hyperspectral scenes (Indian Pines, Houston 2018). The results demonstrate that the proposed algorithms achieve comparable or superior performance to existing approaches while drastically reducing computational cost and data volume.

Committee:
     PRESIDENT: SANTAMARÍA CABALLERO, IGNACIO
     SECRETARI: SALA ALVAREZ, JOSE
     VOCAL: PEREZ NEIRA, ANA ISABEL
     VOCAL: SECO GRANADOS, GONZALO
     VOCAL: KOCH, TOBIAS
Thesis abstract: Diversity is a well-established concept in wireless communications whose purpose is to quantify the potential robustness of a receiver when multiple independent copies of the informative signal are received. Indeed, there exists a formal definition of this concept within the context of wireless communications that takes into account its practical usage, i.e. it is defined with respect to the symbol error probability averaged over the channel statistical fluctuation. However, there is no consensus on the generalization of the previous definition to other forms of signal processing applications. For this reason and being inspired by an intuitive definition of diversity extracted from the multimodal data fusion framework, the purpose of this dissertation is to explore the concept of diversity through the lenses of Information theory, a numerical optimization framework based on the Majorization-Minimization principle and the Grassmann manifold. The motivation behind the Majorization-Minimization algorithms is that they fit perfectly to the optimization problems arising from information theoretic cost functions, while the Grassmann manifold emerges naturally in the context of sparse-aware signal processing problems that exhibit some sort of diversity. All these ideas are surveyed through three different scenarios: the multisensor fusion, the Covariance Conversion from wireless MIMO communications and the detection of correlation. All of the scenarios share the fact that the intrinsic dimension of the data is much smaller than the ambient space dimension.In the multisensor fusion problem, we analyze the intuitive definition of diversity in a straightforward manner for three fusion policies. Firstly, the Covariance Intersection principle is reviewed to highlight its connection to the minimum error entropy criterion and the waterfilling algorithm for optimal power allocation in communications. Secondly, we derive a bounded descriptor based on the R\'enyi entropy of a sensor network contamination worst-case scenario (unbounded variance). Thanks to the aforementioned descriptor, it is possible to provide an operational interpretation to the commonly used L0 norm regularization particularized for this problem. Finally, we consider a fusion scheme that incorporates a subspace-based regression technique into the fusion operation. This proposal, which is inspired by a duality with the problem of unstructured interference mitigation in navigation receivers, is motivated by the fact that it is possible to obtain a measure of the fusion integrity when the temporal redundancy of the measurements and the intersensor covariance matrix are estimated in a joint manner.Besides, a different kind of diversity is unveiled in the Covariance Conversion problem for Frequency Division Duplexing schemes from wireless communications. In essence, this problem consists in the estimation of the Downlink channel covariance matrix using a prior estimation of the Uplink channel covariance matrix. Particularly, we are interested in those cases where sparsity can be defined on the second-order statistics, which are found in the mmWave and ultra-wide band channels. Through a detailed analysis of this problem, we show a promising conversion algorithm founded on the Alternating Direction Method of Multipliers.Lastly, the detection of correlation between two Gaussian vectors problem serves as a way to explore an information theoretic approach for the quantification of diversity. In fact, we transform this setting into a Mutual Information estimation problem of M parallel Gaussian channels to yield the aforementioned information theoretic measure. However, the Maximum Likelihood estimation of the Mutual Information suffers from bias when a subset of these channels provide no information. In light of this, we propose the adoption of model-order selection rules, well-known in other areas, as a means for estimating information under a bias-variance trade-off.

Committee:
     PRESIDENT: GÜNDÜZ, DENIZ
     SECRETARI: PASCUAL ISERTE, ANTONIO
     VOCAL: POOR, VINCENT
Thesis abstract: This dissertation addresses two fundamental and increasingly connected challenges in modern information processing: nonlinear learning and task-oriented communications. Both fields require compact, interpretable and efficient representations of nonlinear functions. These representations are essential for building learning models as well as for transmitting information relevant to a downstream task. Traditional approaches, often dominated by black-box architectures and data-driven approaches, face limitations in interpretability, scalability and robustness. This motivates the search for structured and principled alternatives.At the core of this work lies the discrete cosine transform (DCT), a classical signal processing tool reinterpreted here as a unifying framework for nonlinear modeling and communication. By leveraging its mathematical properties, such as orthogonality, energy compaction and oscillatory structure, the DCT is shown to offer advantages far beyond its traditional role in image compression. The dissertation explores how the DCT can support both efficient learning algorithms and the design of novel modulation strategies that jointly encode data and computation.The study begins with the integration of the DCT in adaptive learning algorithms for univariate nonlinear functions, where the structural properties of the DCT enable fast convergence and low complexity solutions. This method is successfully applied to the compensation of nonlinearities in wireless communication channels. The approach merges the structure and interpretability of classical signal processing blocks with the flexibility and performance of data-driven methods.The work then extends to multivariate settings through the introduction of the Expressive Neural Network (ENN), a multilayer perceptron that integrates the DCT within its activation functions. This architecture preserves the interpretability of the DCT while enabling expressive and scalable modeling. Thanks to the structure provided by the DCT, even small networks can learn complex nonlinear mappings.On the communications front, the dissertation investigates the emerging paradigm of task-oriented communications, where the focus shifts to transmitting data relevant to a computation task. To this end, the oscillatory nature of the DCT is employed to design modulation schemes that embed computation within the physical waveform. A novel DCT-based modulation is proposed, and classical schemes are revisited through the lens of the DCT, enabling function computation through frequency modulations. These concepts are further extended to multi-user wireless systems, where the DCT guides the design of a new multiple access scheme tailored for federated edge learning (FEEL). This approach offers robust, low-power and efficient aggregation of distributed data for collaborative edge learning.Altogether, this work demonstrates that the DCT, a classical tool introduced nearly 50 years ago, can be revitalized as a foundational building block for modern signal processing. Its structure offers a powerful, interpretable and efficient foundation for addressing contemporary challenges in intelligent information processing. By bridging classical signal processing with contemporary machine learning and communication challenges, the dissertation presents a unified framework for nonlinear representation, learning and communications.

Committee:
     PRESIDENT: FERRO-FAMIL, LAURENT
     SECRETARI: LOPEZ MARTINEZ, CARLOS
     VOCAL: BOVENGA, FABIO
Thesis abstract: Ongoing changes in climate and weather patterns have emphasized the need for improved observation and understanding of rapid land–atmosphere interactions, particularly those related to the water cycle. This thesis explores the use of high-temporal-resolution SAR to monitor soil and crop dynamics, aiming to support the development of future Earth Observation missions, especially from geosynchronous orbit with short revisit times.The research was conducted within the framework of the HydroSoil project, funded by ESA, which established a unique ground-based SAR observation facility. The system allows for continuous monitoring of fields with a temporal resolution of 10 minutes and a spatial resolution of 1 m2 in the boresight. Complemented by extensive meteorological and ground truth data, the facility offered an ideal setup to study short-term dynamics of SM and crop development, focusing on two crops: barley and corn.The first part of the study examines diurnal cycles in radar backscatter time series and its relationship with environmental variables such as atmospheric humidity, temperature, and dew formation. Time-frequency analysis using wavelets was applied to quantify how radar signals correlate with specific meteorological variables at different time scales. Following this, a two-layer surface scattering model was applied to better capture the effects of diurnal SM variations on radar backscatter. In particular, the model accounts for the formation of a transient wet layer at the soil surface during nighttime, followed by daytime evaporation. When integrated with a short-term change detection method, the model enabled the retrieval of soil moisture time series with improved accuracy, particularly under conditions of high incidence angles and during rapid transitions following precipitation.The second part of the thesis focuses on the radar response to crop development. It is shown that radar backscatter and interferometric coherence are sensitive to plant height and phenological stage, though these relationships weaken under dense vegetation. A new metric, the daily growth rate, was introduced to better characterize crop evolution, correlating better with radar observables than traditional static parameters. Additionally, diurnal cycles in backscatter, primarily driven by dew formation and atmospheric humidity, were consistently observed and quantified.The final part of the thesis investigates the scattering mechanisms underlying the radar signal using several polarimetric decomposition techniques. These tools allowed for a detailed temporal characterization of how radar signals respond to environmental variables and crop growth stages. While all decomposition methods provided valuable insights, each also presented limitations, particularly in high biomass conditions where traditional models tended to overestimate volume scattering due to saturation effects.To address these limitations, the NNED was applied, reducing inconsistent scattering artifacts and improving robustness especially under dense canopies. The study also highlights that attributing all volumetric scattering to cross-polarized channels is overly simplistic. It is suggested that future research should explore combinations of different decomposition methods and leverage time-series analysis to refine interpretations of crop–environment interactions.In conclusion, this work demonstrates the potential of high-temporal-resolution SAR for monitoring short-term variations in soil and vegetation conditions. By integrating advanced modeling, coherence analysis, and polarimetric decompositions, the thesis provides a comprehensive framework for improving the understanding of land surface processes. The findings contribute to the current development of SAR-based methods for precision agriculture and offer practical guidance for the design of future radar missions targeting environmental and agricultural monitoring applications.

Committee:
     PRESIDENT: PEREIRA GUIOMAR, FERNANDO PEDRO
     SECRETARI: FÀBREGA SÁNCHEZ, JOSEP MARIA
     VOCAL: CANO VALADEZ, IVAN NICOLAS
Thesis abstract: Optical fiber communications have revolutionized the telecommunications networks in all segments, leveraging a wide set of advantages ranging from THz capacity to low-loss, security and form factor. In long-haul and core networks, optical links use a plethora of properties of the electromagnetic field jointly to maximize the throughput of the deployed infrastructure, such as its frequency, phase, polarization and amplitude, enabled by coherent detection. In contrast, in access and fronthaul systems, cost-effectiveness has typically constrained the dimensions used for data transmission to amplitude and coarse frequency multiplexing, based on direct-detection. While both approaches have succeeded by optimizing link properties for their respective requirements, the growing traffic near the edge of the networks forecasts the need to implement the advantages of coherent detection and full field modulation, such as scalability, increased capacity and network flexibility, even in cost-constrained segments.This thesis aims to develop and assess architectures and technologies to simplify full-field modulation and coherent detection for edge segments like fixed access and mobile front-haul, especially over passive optical networks. In particular, single-sideband techniques of signals modulated over electric carriers are used to achieve single-laser transceivers and spectrally efficient sub-carrier multiplexing with increased sensitivities and avoiding relevant link setbacks such as Rayleigh backscattering. To further reduce the cost and consumption of the systems, analog processing for the phase-noise compensation is included, and the links’ performances are evaluated. The thesis focuses on low-cost and footprint devices, suitable for urban and local access networks, such as fiber to the home, business, antenna, etc. using a wavelength-to-the-user/antenna approach, to deliver the highest sensitivity and spectral efficiency possible with the lowest expression of coherent detection and transmitter complexity. Therefore, the main topics discussed in the thesis are simplified coherent detection and field modulation, phase noise cancelling, coherent passive optical networks power budget, scalability and spectral efficiency, and single-sideband techniques as a key-enabling tool for the accomplishments of the aforementioned targets.

Committee:
     PRESIDENT: KSENTINI, ADLEN
     SECRETARI: CHERGUI, HATIM
     VOCAL: VARDAKAS, JOHN
Thesis abstract: This thesis addresses the emerging challenges in resource management for 6G networks by proposing intelligent, scalable, and explainable solutions using Deep Reinforcement Learning (DRL) and related AI techniques. With the evolution from 5G to 6G, the increasing heterogeneity of applications and services introduces complex requirements in terms of latency, bandwidth, computational efficiency, and end-to-end quality of service (QoS). The research presents a suite of AI-driven solutions for dynamic and adaptive resource allocation tailored to network slicing scenarios in Open and Programmable architectures.The work begins by developing a DRL-based, QoS-aware slice resource allocation framework, integrating user association parameterization for beyond-5G O-RAN environments. A hierarchical DRL model is introduced to manage global-local resource trade-offs efficiently. This is extended by proposing a multi-time scale resource management framework under an AI-as-a-Service (AIaaS) paradigm to serve heterogeneous 6G services.To improve interpretability and trust in automated network operations, the thesis incorporates Explainable Reinforcement Learning (XRL) techniques into RAN slicing and management strategies. Finally, the use of transfer learning in DRL is explored to enhance policy adaptation in intra- and inter-slice domains, accelerating learning and improving performance in diverse and dynamic network conditions.The thesis includes extensive simulations and experimental validation to demonstrate the superiority of the proposed methods in terms of scalability, efficiency, and generalization over state-of-the-art (SOTA) baselines. The overall contributions provide a technically innovative and practically applicable roadmap for intelligent, trustworthy, and adaptive resource management in future 6G wireless systems.

Committee:
     PRESIDENT: CLOSAS GÓMEZ, PAU
     SECRETARI: PAGES ZAMORA, ALBA MARIA
     VOCAL: BERARDINELLI, GILBERTO
Thesis abstract: Wireless communications have become the central nervous system of the Factory of the Future (FoF). According to this, wireless infrastructure in industries serves the dual purposeof providing connectivity and positioning industrial assets such as Automated Guided Vehicles (AGVs). The Non-Line-of-Sight (NLoS) and multipath-dominant environments, such asthose found in industries, hinder wireless signal propagation, leading to inaccuracies in wireless infrastructure-based positioning. While advances have been made in developing approaches to enhance wireless positioning accuracy in complex indoor scenarios, they may not be readily applicable to industrial environments without or with minimal modification. This motivates the study towards developing accurate and precise wireless infrastructure-based positioning approaches tailored explicitly for industrial settings. This PhD thesis aims to address key challenges in wireless positioning for industrial environments and develop algorithms to enhance the accuracy of wireless infrastructure-based positioning by proposing both model-driven and data-driven approaches.This PhD thesis presents simulation analysis and practical experiments to validate the efficiency of the proposed approaches using a range of different wireless infrastructures, including Fifth Generation (5G), Ultra-Wideband (UWB), and Sixth Generation (6G) mobile communication, including Integrated Sensing and Communication (ISAC). This work begins by analyzing the achievable 5G-based positioning accuracy in C-band and Millimeter-Wave (mmWave) bands within a specific model of a dense, cluttered industrial environment, utilizing high-fidelity ray tracing simulations to capture the impact of NLoS and multipath propagation. The findings highlight limitations of 5G positioning and, in general, wireless positioning in complex and cluttered industrial settings. To overcome these limitations, this work advances the State of the Art (SotA) by developing novel enhancement approaches based on sensor and/or data fusion for UWB-based and ISAC-assisted positioning systems. These approaches leverage the power of Deep Neural Networks (DNNs), Long Short-Term Memory (LSTM) networks, and Graph Neural Networks (GNNs) to model spatial and temporal relationships more effectively, yielding substantial improvements in positioning accuracy, robustness, and adaptability to dynamic industrial scenarios. Furthermore, the application of the proposed approaches is illustrated through an Automated Guided Vehicle (AGV) use case. This PhD thesis lays a strong foundation for advancing future research and real-world applications, offering valuable insights that can shape the next generation of industrial wireless positioning system design and deployment.

Committee:
     PRESIDENT: TRIANTAFYLLOPOULOU, DIONYSIA
     SECRETARI: RUIZ BOQUE, SILVIA
     VOCAL: LIOTOU, EIRINI
Thesis abstract: The Massive Internet of Things (MIoT) characterizes a communication scenario where a massive number of battery-operated devices perform infrequent, primarily uplink-oriented small data transmissions to a network. The connectivity and management of a massive number of such devices pose challenges, and this thesis aims to tackle the issues related to the initial access mechanism of MIoT in the context of 5G and beyond networks. The thesis begins by conducting a thorough examination of both licensed and unlicensed spectrum technologies for the MIoT. Additionally, we investigate different test beds and simulators employed for the analysis. In the initial phase of the thesis, our focus is on examining existing standardized technologies in Massive IoT, with specific attention given to the Narrow-Band Internet of Things (NB-IoT) as a critical component for supporting MIoT in 5G networks. The performance of the NB-IoT access procedure is thoroughly analyzed under different network densities and combinations of configuration parameters. The findings lead to the development of an adaptive access mechanism that adjusts the access parameters based on the anticipated network density, improving the overall performance of the access procedure. In the subsequent stages of the study, we qualitatively analyzed network-based and User Equipment (UE) based relaying solutions for IoT applications, motivated by the fact that effective support of IoT requires cellular service in deep coverage areas. Relaying is a promising solution to extend the coverage while at the same time meeting the battery life requirements of IoT devices. While various relay-ing architectures and technologies have emerged in research and standardization, it is not yet clear which is best suited for IoT applications. Hence, we examine relay architectures and evaluate the suitability of UE-based relaying and network-based relaying solutions for IoT applications. We analyze the relaying solutions in terms of radio aspects, such as connection establishment, protocol architecture, and resource allocation. Our findings suggest that even though the standardized layer-2 network-based relaying solution (e.g., integrated access and backhaul networks) may be better suited to handle large numbers of low-complexity, low energy-consuming devices with better interference management, any optimizations aimed at supporting IoT traffic will directly affect existing networks, such as the introduction of new control plane signaling over the Uu interface. Conversely, the standardized layer-3 UE-based relaying solution can be optimized for handling massive IoT relaying with minimal impact on existing networks, since the optimizations can be made on the PC-5 interface with a subdued effect on the Uu interface. Based on the outcome of the qualitative study of relaying architecture, we analyze the suitability of layer-3 relaying over the New Radio (NR)-PC5 interface to support massive IoT applications. More precisely, we study the unicast connection establishment mechanism over the NR PC5 interface in a partial coverage scenario. Further, a set of optimizations on the NR-PC5 procedure to effectively support massive IoT applications are proposed and analyzed. The obtained performance evaluation results, which are presented in terms of data success probability, device power consumption, and signaling overhead, quantify how effectively the NR-PC5 interface can support the requirement of IoT in the 5G and beyond era. The proposed sidelink small data transmission and frame-level access provides the largest gain overall and can reduce the device power consumption by an average of 68 percent and signaling overhead by 15 percent while maintaining a data success probability of more than 90 percent in an International Mobile Telecommunications (IMT)-2020 defined IoT traffic scenario.

Committee:
     PRESIDENT: MARTIN NEIRA, MANUEL
     SECRETARI: VALL-LLOSSERA FERRAN, MERCEDES MAGDALENA
     VOCAL: PIERDICCA, NAZZARENO
Thesis abstract: The research presented in this paper-based dissertation focuses on the development of products for the ESA HydroGNSS mission, including the Level 2 surface inundation algorithm, the exploitation of the coherent channel for the L1B product, and the development of a simplified coherent channel simulator.The first contribution of this work is the development of a water detection algorithm that is applicable to high and low sampling rate data. The integration of two low-sampling rate power-derived observables, the reflectivity and the power spread ratio, enables a more comprehensive characterization of surface properties. And, the incorporation of auxiliary background information mitigates the impact of spatial heterogeneity. The random forest classifier effectively captures nonlinear relationships, achieving results that meet accuracy requirements, with false positives primarily associated with wetland areas. Another advancement is the integration of two high-sampling rate complex-derived observables, one of which is introduced for the first time -- the normalized amplitude (NA). NA remains stable over land and shows minimal sensitivity to changes in coherence integration time. The findings indicate that the NA has greater feature importance than the coherence coefficient (CC), underscoring its relevance for water classification.The second contribution of this work involves an analysis of the impact of the right-hand circularly polarized signal on the water detection algorithm. The normalized power spread ratio in the right-hand circularly polarized signal exhibits a clearer distinction between land and water surfaces and shows higher consistency across various land types compared to its left-hand counterpart. Reflectivity in the right-hand circularly polarized exhibits increased overlap between land and water surfaces, which may negatively impact classification performance. For non-inland water surfaces, the surface type contributes negatively to the output, indicating that higher coherence indicators are required to mitigate its influence. Furthermore, posterior probabilities derived from Bayes’ theorem demonstrate that the combination of normalized power spread ratio and reflectivity enhances classification accuracy by reducing false positives.The third contribution focuses on the exploitation of the HydroGNSS coherent channel. An analysis was conducted to assess the sensitivity of coherence indicators to water surfaces. CC and NA were compared to the `full entropy' metric available in the CYGNSS product. The results indicate that NA exhibits a similar sensitivity to water surfaces as full entropy and remains stable regardless of the coherent integration time. In contrast, the CC demonstrates higher sensitivity to small water bodies but is strongly influenced by the integration time. Additionally, a new variable, high sampling rate reflectivity (HSRR), was developed from normalized amplitude applying a track-wise exponential fit to the low-rate calibrated reflectivity. HSRR exhibits good agreement with both CYGNSS L1 reflectivity and calibrated raw IF reflectivity, further validated through a forward model. To enhance its robustness, a noise threshold based on low sampling rate noise floor was introduced. Adjacent track analyses confirmed that it provides a higher resolution than current 1–2 Hz products, improving its applicability for detailed surface characteristic monitoring.The development of a simplified coherent channel simulator has contributed to better understanding the behavior of this new GNSS-R mode of operation, and is included as an annex. Until now, simulating high sampling rate complex GNSS-R signals reflected over land required complicated models with computationally expensive implementations. Being able to reduce the complexity of the modeling and its implementation and computing time has facilitated the generation of synthetic data that mimic the properties of the coherent channel.

Committee:
     PRESIDENT: CIOMPI, FRANCESCO
     SECRETARI: CASAS PLA, JOSEP RAMON
     VOCAL: SÁNCHEZ GUTIÉRREZ, CLARISA
Thesis abstract: Histopathology is the gold standard for cancer diagnosis, where pathologists examine tissue samples to assess cellular morphology, tissue architecture, and biomarker expression. The digitization of histological slides has enabled the application of artificial intelligence to assist in this process. However, deploying deep learning models in pathology remains challenging due to the gigapixel-scale nature of whole slide images, variability introduced by staining protocols and tissue morphology, and the scarcity of high-quality annotations.This thesis focuses on the development of deep learning methodologies for the identification and characterization of cells in histopathological slides, addressing these core challenges. Inspired by the way pathologists assess slides, the proposed methods target both individual cells and their spatial organization within the tissue. A transformer-based model, CellNuc-DETR, is introduced for efficient cell nuclei detection and classification, achieving state-of-the-art performance while significantly reducing inference time. To address annotation scarcity in immunohistochemistry, the model is adapted using unsupervised domain adaptation techniques, enabling knowledge transfer from hematoxylin and eosin-stained slides without requiring annotated immunohistochemistry data.To incorporate broader spatial context into cell classification, the thesis explores graph-based representations of histological tissue. It proposes a novel graph construction strategy that integrates both cell-level and patch-level nodes, enabling the fusion of cellular features with visual and contextual information in a scalable manner. To support annotation-efficient learning, the thesis introduces Regularized Graph Infomax, a self-supervised learning algorithm designed for representation learning on graphs without manual annotations. This algorithm is then applied to the proposed cell–patch graphs to enable efficient, large-context cell classification with minimal supervision.Finally, the thesis demonstrates the clinical relevance of these methods in a study on endometrial carcinoma. A weakly supervised whole slide classifier is trained to predict molecular subtypes, and the developed cell detection and classification tools are used to extract cell-level biomarkers from predicted regions of interest. These biomarkers enable the quantification of subtype-specific cell spatial organization differences, linking computational predictions to interpretable biological features.By integrating transformers, graph-based representations, and self-supervised learning, this thesis introduces scalable and annotation-efficient methodologies for computational pathology. These approaches enhance the efficiency, robustness, and generalization of artificial intelligence models, contributing to the development of clinically relevant tools for cancer diagnosis and characterization.

Committee:
     PRESIDENT: WICKERT, JENS
     SECRETARI: VAZQUEZ GRAU, GREGORIO
     VOCAL: TEGEDOR, JAVIER
Thesis abstract: Radio Frequency Interference (RFI) poses significant challenges to remote sensing systems, particularly affecting Global Navigation Satellite System-Reflectometry (GNSS-R) and microwave radiometry. This thesis advances two interconnected areas of remote sensing: geophysical parameter retrieval using signals of opportunity and RFI impact assessment on Position, Navigation, and Timing (PNT) systems. The research develops novel approaches for RFI detection, mitigation, and excision across multiple remote sensing platforms, with particular emphasis on space-qualified systems.The first part of this work advances GNSS-R instrumentation through the development of an experimental testbed and simulator. Laboratory measurements in controlled environments provided critical insights into GNSS-R signal characteristics, enabling systematic validation of signal processing algorithms. The research demonstrates the feasibility of continuous Earth observation through opportunistic platforms, specifically integrating GNSS-R instruments in commercial aviation for applications including weather nowcasting and maritime route optimization. Analysis of polarimetric data revealed significant limitations in current receiver architectures, particularly in processing Left-Hand Circular Polarization (LHCP) signals and reconstructing Stokes parameters. This continued with work conducted at JPL in 2023, analyzing data from the SMAP-R mission which modified its radar payload for GNSS-R reception. The dual capability for forward and backward scattering measurements from the same platform enabled comparative analysis of scattering mechanisms. Leveraging SMAP-R's fully polarimetric reflectometry measurements, we investigated bistatic radar scattering characteristics across varied surface conditions. The analysis focused on two aspects: comparing scattering regime behaviors across different terrains, and developing a prediction model that derives backward-scattering characteristics from forward-scattering measurements. This work extends the theoretical understanding of bistatic scattering mechanisms while providing validation through spaceborne measurements.The second part addresses RFI challenges through three complementary approaches: spectrum monitoring systems, hardware-efficient algorithms, and polarimetric detection methods. A key contribution is the development of an automatic RFI Detection, Location, and Classification (RFIDLC) system for GNSS bands, utilizing a six-sector antenna array design for real-time processing between 1525 and 1625 MHz. For Synthetic Aperture Radar (SAR) applications, novel algorithms operating on 1-bit quantized signals enabled RFI mitigation despite severe quantization effects. The FPGA implementation achieved significant resource reduction through innovative overclocking and serialization techniques while maintaining space qualification requirements.A major advancement is the development of the Polarimetric Kurtosis detector for microwave radiometry, extending conventional kurtosis-based detection to the full-polarimetric domain through a four-dimensional measurement derived from Stokes parameters. This technique demonstrates enhanced sensitivity to polarized interference signals, particularly in scenarios where conventional detection methods show limitations. The research validates these approaches through experimental testing on both ground-based and spaceborne platforms, establishing a foundation for future RFI mitigation systems in space-based Earth observation missions.

Committee:
     PRESIDENT: LABAKA INTXAUSPE, GORKA
     SECRETARI: ESPAÑA BONET, CRISTINA
     VOCAL: NOZZA, DEBORA
Thesis abstract: This thesis delves into various aspects of natural language processing, focusing on domain-specific neural machine translation, specialized word embeddings, data augmentation, recommender systems, document embedding techniques, and hierarchical classification with large language models. The work is structured around a couple of key contributions.Chapter 4 introduces a novel data preparation and tokenization method (Hybrid-BTS) for biomedical neural machine translation, alongside a post-specialization technique leveraging Wasserstein GANs to enhance word embeddings with multilingual constraints. Chapter 5 proposes a domain adaptation strategy for biomedical translation involving forward translation, BPE optimization, and term frequency manipulation. It also details the development of two fashion e-commerce recommender systems: one content-based and one collaborative, both integrating practical style rules. Finally, Chapter 6 presents an evaluation of document embedding models (Doc2vec, SciBERT, Longformer, LLaMA-3, GEMMA-2B) for long document classification, and a modular pipeline designed for multi-label hierarchical patent classification using transformer-based models, optimized for efficiency with LoRA and quantization. Collectively, these contributions push the state-of-the-art in both applied and theoretical NLP by providing new methods to boost performance, adaptability, and efficiency in domain-specific and large-scale applications.

Committee:
     PRESIDENT: PEREZ NEIRA, ANA ISABEL
     SECRETARI: IBARS CASAS, CHRISTIAN
     VOCAL: CLERCKX, BRUNO
Thesis abstract: As wireless networks evolve toward sixth-generation (6G) systems, the integration of advanced technologies like massive MIMO, millimeter-wave communication, intelligent reflecting surfaces, and massive multiple access schemes has led to significant increases in system complexity. Traditional signal processing and optimization techniques are grounded in explicit modeling and mathematical tractability. They increasingly fall short in addressing the high-dimensional, dynamic, and often non-linear behaviors of these systems. To bridge this gap, Artificial Intelligence and Machine Learning (AI/ML) methods have emerged as powerful alternatives, capable of learning directly from data and adapting to complex and uncertain environments.Rather than relying solely on conventional models or pure data-driven solutions, this dissertation advocates for hybrid approaches that combine model-based structures with learning-based flexibility. Hybrid methods mitigate the high data demands and limited interpretability of black-box models while overcoming the rigidity and assumptions of traditional schemes. This balance makes them particularly well-suited for the complex, imperfectly modeled conditions encountered in wireless systems.This PhD dissertation focuses on applying such hybrid methods to four key tasks in the wireless signal processing chain: channel estimation, symbol detection, digital precoding, and hybrid beamforming. Though each task poses unique technical challenges, they collectively contribute to the overarching goal of enabling robust, efficient, and scalable communication in next-generation systems.The first objective addresses channel estimation, a foundational task in massive MIMO systems where performance hinges on accurate knowledge of the channel state. In highly dynamic or non-linear environments, traditional estimators like LS or MMSE are often inaccurate. This work proposes low-complexity neural network architectures, specifically 1D-CNNs and Graph Neural Networks (GNNs), that serve as hybrid estimators by embedding structural insights while learning to compensate for model mismatches, showing strong performance across varying MIMO configurations.The second objective explores symbol detection in Rate-Splitting Multiple Access (RSMA), where receivers must decode a common message followed by private streams. The conventional RSMA receivers use successive interference cancellation (SIC) or joint detection schemes, that are capable of detecting all streams simultaneously. These are theoretically superior, but computationally infeasible in large-scale systems. To balance performance and complexity, this work introduces deep receiver architectures such as RS-Net+, which combine the interpretability of SIC with the adaptability of neural networks. These hybrid designs exhibit robustness to channel impairments and outperform both purely model-based and black-box neural schemes under linear and non-linear channel conditions.The third and fourth objectives focus on designing precoding techniques for RSMA, both in fully digital and hybrid analog-digital beamforming settings. Standard approaches like WMMSE, though effective, are iterative and computationally intensive. This thesis proposes self-supervised learning frameworks, based on GNN and MLP architectures, that learn precoding policies directly from channel data while respecting underlying model structures. These methods are then extended to hybrid beamforming scenarios, where learned digital precoders work in tandem with fixed analog beamformers to provide scalable and energy-efficient solutions.In summary, this dissertation advances the design of hybrid AI/ML solutions for core signal processing tasks in wireless communication systems. By combining domain knowledge with learning, it delivers architectures that are robust, interpretable, and well-suited to the challenges of future 6G networks.

Committee:
     PRESIDENT: HARO ORTEGA, GLORIA
     SECRETARI: RUIZ HIDALGO, JAVIER
     VOCAL: VENTURA ROYO, CARLES
Thesis abstract: Image understanding is a fundamental task in computer vision as it aims to enable machines to visually comprehend the real world. In this thesis, the task of image understanding is addressed through the combination of instance-level recognition systems and Large Multimodal Models (LMMs). To do so, the thesis is structured in three parts, starting with a study of deep metric learning techniques for creating rich image embeddings spaces. Then, the implementation of instance-level recognition systems specifically focused on landmark recognition using content-based image retrieval. And finally, enhance the capabilities of LMMs by incorporating instance-level recognition results in order to generate improved image descriptions with instances. The first part of this thesis focuses on advancing deep metric learning methods. A novel Smooth Proxy-Anchor Loss is proposed to address the problem of having noisy labels in training data. The proposed loss introduces confidence-based sample weighting, soft proxy assignment, and adaptive margins to mitigate the negative impact of mislabeled samples, demonstrating superior performance on real-world data. Building on top of this foundation, the thesis extends the selective weighting of samples to a selective weighting of features, through the Generalized Local Attention Pooling (GLAP) method. GLAP dynamically weights different regions of feature maps according to their informational content, enabling better representation learning while reducing computational requirements. Evaluations on multiple benchmark datasets show the significant improvements of using GLAP over state-of-the-art approaches.The second part introduces the Multi-Scale Transformer-based Feature Combination (MSTFC) method for instance-level landmark recognition. This approach applies a transformer-based attention mechanisms to select and combine relevant information from feature maps. These feature maps are extracted from multiple scales and have different spatial resolutions. The combination of these multi-scale feature maps results in a compact global representation that is used for image retrieval. The MSTFC method achieves a superior performance on the challenging landmark retrieval datasets, outperforming existing approaches on the Google Landmarks Dataset v2 and the Revisiting Oxford and Paris benchmarks.Finally, the thesis explores the enhancement of Large Multimodal Models by incorporating instance-level recognition results to generate more accurate and detailed image descriptions. By bridging specialized instance-level recognition systems with the generic understanding capabilities of LMMs, this integration demonstrates how domain-specific computer vision techniques can be combined with multimodal techniques for comprehensive image understanding.Through these contributions, the thesis advances the field of image understanding across multiple levels, from robust feature learning in real-world noisy dataset, to effective techniques for combining the information of feature maps in deep metric learning, then expanded to multi-scale representation for instance-level recognition, and ultimately to the integration of specialized instance-level systems with modern multimodal frameworks.

Committee:
     PRESIDENT: VENTURA ROYO, CARLES
     SECRETARI: MORROS RUBIO, JOSEP RAMON
     VOCAL: MOSELLA MONTORO, ALBERT
Thesis abstract: Spatio-temporal action localization is a field of computer vision that determines both the spatial and temporal locations of actions taking place within a video. This particular task is one of the cornerstones of video understanding due to its inherent complexity. Impressive scientific advances in the artificial intelligence and computer vision industry have led to a significant performance increase.Nevertheless, even state of the art systems struggle to recognize actions that involve some kind of interaction between people, objects, or with the scene. This limitation arises because current spatio-temporal action localization models neglect contextual information when identifying a person's actions. Specifically incorporating contextual cues in these models can help to address this challenge. Furthermore, the integration of spatio-temporal action localization systems into multimedia content applications presents several challenges. The primary difficulties lies in finding ethical and practical use cases and addressing the computational requirements to run these models in production.The goal of this industrial PhD is to solve the above problems. The lack of contextual information is addressed in different forms. This thesis, is focussed on two different approaches. First, a novel system to explicitly model the relations of the different actors and objects in a scene, which includes a novel structure to model long-term temporal information is proposed. Second, an innovative multi-modal method to combine task specific features to improve action recognition is introduced. These developed contextualization systems are evaluated on different datasets. The results demonstrate the effectiveness of the proposed systems. Regarding the applications, this PhD is focused on finding, implementing and exploiting specific applications of spatio-temporal localization system for media content. The developed applications consists of an improved system for search and recommendations on video content, an smart cropping method to generate image and videos of any desired aspect ratio from another video, and the improvement of a system to generate highlight keyframes of football matches.

Committee:
     PRESIDENT: CARPINTERO DEL BARRIO, GUILLERMO
     SECRETARI: ROMEU ROBERT, JORDI
     VOCAL: PRUNERI, VALERIO
Thesis abstract: This thesis investigates the design, development, and characterization of optical-enabled wireless systems operating in the millimeter wave (mmWave) and Terahertz (THz) frequency ranges, addressing the need for high-capacity, low-latency networks in next-generation communication systems.In the context of optical networks, analog radio-over-fiber (A-RoF) systems are optimized for 5G frequencies wireless communication. Key advancements include the demonstration of cascaded optical modulators capable of performing optical frequency conversion (OFC), and optical power budget (OPB) enhancement.The study follows with the integration of photodetectors (PDs) and antennas, where reconfigurable photonically-aided antennas are designed and experimentally validated for operation at 30GHz and 50 GHz. These antennas showcase optimal impedance matching and efficient performance, leveraging flip-chip bonding techniques to ensure compact and scalable designs.For beam steering, a photonic-assisted lens array system is developed, showing excellent control of directional beams through theoretical analysis and experimental validations, with potential applications in satellite communications and advanced sensor networks.In the THz domain, photoconductive antennas (PCAs) are evaluated for time domain spectroscopy (TDS). Alignment procedures are improved, and a novel cross-polar beam pattern analysis method is introduced to identify fabrication tolerances and improve performance. These contributions enhance the precision and reliability of THz imaging and spectroscopy setups. The results present in this thesis highlight the potential of photonically integrated systems to contribute on bridging the gap between optical and wireless domains, offering robust solutions for next-generation communication networks and high-frequency sensing applications.

Committee:
     PRESIDENT: FARRUS CABECERAN, MIREIA
     SECRETARI: RODRIGUEZ FONOLLOSA, JOSE ADRIAN
     VOCAL: SEGURA PERALES, CARLOS
Thesis abstract: Although Automatic Speech Recognition (ASR) technology has achieved remarkable improvements in transcription accuracy in recent years, it still struggles to correctly transcribe certain words. In particular, proper nouns often exhibit lower accuracy due to their unique pronunciations and alternative spellings. To address these challenges, contextualisation is commonly integrated into ASR models to improve the transcription of rare proper nouns and disambiguate between similar-sounding proper nouns.This industrial PhD thesis focuses on the development of contextualisation systems for ASR models. The ASR technology employed in this thesis is used to generate automatic transcripts for podcast content. The objective of the contextualisation system is to improve the accuracy of proper nouns; for example, the names of podcast shows, hosts, and guests.The contextualisation system developed in this thesis is composed of two parts. First, an ingestion pipeline gathers proper nouns relevant to the podcast episode to be transcribed. This pipeline has an automated component that extracts proper nouns from the metadata of each podcast episode, such as the title and description, using tools like a named entity recogniser. Additionally, a module that allows the manual addition of proper nouns to specific podcast shows has been developed. The second part, which represents the most relevant contribution of this thesis, is a novel ASR contextualisation algorithm based on deep neural networks. The contextualized ASR model utilises the gathered proper nouns, resulting in an improvement in accuracy when compared to the same model without contextualization capabilities.The accuracy of the ASR system developed in this thesis is evaluated and analysed using episodes from popular public podcast shows. A human evaluation of the word error rate was employed during this assessment.This evaluation compares the quality of the ASR model system's transcripts to those provided publicly by the podcast content creators. The results show that the system developed in this thesis produces transcripts that contain four times fewer errors than the transcripts offered by podcast hosting providers.The contextualization method is also evaluated on a public dataset and compared to state-of-the-art methods. The results show that the contextualization method proposed in this thesis significantly outperforms the existing systems.