SINGLE-IMAGE SUPER-RESOLUTION OF S5P RADIANCE DATA THROUGH A PSF-AWARE AND SCALABLE SPECTRAL DEEP-LEARNING-BASED APPROACH

ALTHOUGH THE SENTINEL-5 PRECURSOR (S5P) SATELLITE PROVIDES VALUABLE GLOBAL INFORMATION ON KEY AIR POLLUTANTS, ITS SPATIAL RESOLUTION REMAINS INSUFFICIENT TO CAPTURE THEIR FINE-SCALE DISTRIBUTION IN CONFINED AREAS SUCH AS CITIES, WHERE CONCENTRATIONS ARE TYPICALLY HIGHEST. GROUND-BASED MONITORING NETWORKS, WHILE MORE DETAILED, ARE SPARSE OR EVEN ABSENT IN MANY REGIONS, MAKING SATELLITE OBSERVATIONS INDISPENSABLE. AMONG CURRENT PLATFORMS, S5P OFFERS THE BEST COMPROMISE BETWEEN SPECTRAL RESOLUTION — ALLOWING THE MONITORING OF MULTIPLE POLLUTANTS — SPATIAL RESOLUTION, AND GLOBAL COVERAGE. NEVERTHELESS, PHYSICAL CONSTRAINTS PREVENT IMPROVEMENTS BEYOND ITS NOMINAL SPATIAL RESOLUTION, UNDERSCORING THE RELEVANCE OF IMAGE ENHANCEMENT TECHNIQUES. THIS THESIS PRESENTS THE FIRST APPLICATION OF A SUPER-RESOLUTION APPROACH — SINGLE-IMAGE SUPER-RESOLUTION (SISR) — TO S5P RADIANCE DATA. WE INTRODUCE S5NET, A LIGHTWEIGHT DEEP-LEARNING MODEL THAT DELIVERS EXCELLENT PERFORMANCE WHILE REMAINING SIMPLER THAN STATE-OF-THE-ART ARCHITECTURES. IN PARTICULAR, WE CHARACTERISE S5P'S PAYLOAD BY MODELLING ITS IMAGING ACQUISITION PROCESS AND GENERATING REALISTIC LOW-HIGH RESOLUTION PAIRS, MAKING THE FRAMEWORK EXPLICITLY PSF-AWARE — AN ASPECT OFTEN OVERLOOKED IN REMOTE SENSING STUDIES. ADDITIONALLY, WE PROPOSE A DYNAMIC MULTI-DIRECTIONAL CASCADE FINE-TUNING SCHEME THAT ADAPTIVELY DETERMINES THE NUMBER OF ITERATIONS PER CHANNEL FROM INTER-BAND CORRELATIONS. THIS APPROACH PRESERVES SPECTRAL FIDELITY WHILE OPTIMISING BOTH SPATIAL RECONSTRUCTION AND COMPUTATIONAL EFFICIENCY. OUR FRAMEWORK ACHIEVES EXCELLENT RESULTS. ACROSS ALL ORBITS, PSF-AWARE SISR CONSISTENTLY OUTPERFORMS EXISTING METHODS, SHOWING THAT EXPLICITLY MODELLING THE SENSOR’S DEGRADATION SUBSTANTIALLY IMPROVES SPATIAL RECONSTRUCTION. S5NET ACHIEVES THE OPTIMAL BALANCE BETWEEN SPATIAL DETAIL AND SPECTRAL FIDELITY, ALIGNING QUANTITATIVE METRICS WITH VISUAL QUALITY. FURTHERMORE, THE PROPOSED DYNAMIC FINE-TUNING SURPASSES PCA-BASED AND STATIC CASCADE APPROACHES BY ENHANCING SPATIAL PERFORMANCE, PRESERVING SPECTRAL CONSISTENCY, AND ACCOUNTING FOR INTER-SPECTROMETER VARIABILITY, LEADING TO IMPROVED RESULTS ESPECIALLY ON SPECTROMETER-PARTITIONED SUB-IMAGES. TO DEMONSTRATE PRACTICAL BENEFITS, WE ADDITIONALLY APPLY THE FRAMEWORK TO S5P IMAGERY OVER A POLLUTED URBAN AREA, EXTRACTING SURFACE FEATURES FROM BOTH ORIGINAL AND SUPER-RESOLVED IMAGES TO PREDICT THE OXIDATIVE POTENTIAL (OP) OF PARTICULATE MATTER. PREDICTIONS FROM SUPER-RESOLVED IMAGERY CONSISTENTLY OUTPERFORM THOSE FROM NOMINAL-RESOLUTION DATA. SINCE OP IS A KEY INDICATOR OF PARTICULATE MATTER'S TOXICITY, CLOSELY LINKED TO ADVERSE HEALTH EFFECTS, THESE RESULTS HIGHLIGHT THE FRAMEWORK’S POTENTIAL FOR SUPPORTING URBAN EXPOSURE ASSESSMENT. BY INTEGRATING SENSOR-SPECIFIC CHARACTERISTICS AND ADDRESSING THE CHALLENGES OF HYPERSPECTRAL REMOTE SENSING, OUR PSF-AWARE AND SPECTRALLY SCALABLE SISR FRAMEWORK ENHANCES S5P IMAGERY IN A PROBLEM-TAILORED WAY, ENABLING MORE ACCURATE URBAN EXPOSURE ASSESSMENTS AND SUPPORTING LARGE-SCALE AIR QUALITY MONITORING.