Semi-automatic decay detection from hyperspectral images for historical bridge assessment

Hyperspectral imaging (HSI) is an advanced non-destructive evaluation (NDE) technique with significant potential for the Structural Health Monitoring (SHM) of civil infrastructure, particularly aging bridges. By capturing detailed spectral signatures, HSI can identify material degradation, such as efflorescence and moisture, which are often precursors to structural decay and invisible to standard visual inspection. However, a critical barrier to the practical application of HSI is the requisite for robust data pre-processing. Raw sensor data must be corrected for geometric distortions and radiometrically calibrated to produce physically meaningful surface reflectance. This thesis confronts a primary challenge encountered in field-based SHM: a catastrophic failure of the standard Empirical Line Calibration (ELC) workflow, which, upon investigation, was traced to a combination of band-to-band geometric misregistration and metadata-handling errors in the reference spectral libraries. The primary aim of this thesis is to develop, validate, and optimize a complete and repeatable processing pipeline for the semi-automatic detection of material degradation on historical masonry bridges from ground-based HSI data. This work systematically diagnoses the pre-processing failures and establishes a methodology that proceeds from data correction to final material classification. This study utilizes a dataset of 96-band hyperspectral images acquired from a Rikola frame camera at the Cavour Canal bridge in Saluggia (TO). The developed workflow first addresses data integrity, beginning with a comparative analysis of geometric coregistration methods to correct inter-band spatial misalignment. It then details the diagnosis and resolution of the critical radiometric calibration failure by correcting spectral library metadata errors within the ENVI software environment. Using the corrected data, an Object-Based Image Analysis (OBIA) approach is implemented in eCognition. This research systematically investigates the impact of feature engineering (spectral, geometric, and textural features), segmentation scale, and classifier selection (Support Vector Machine, k-Nearest Neighbors, Decision Tree, Random Trees, and Bayes) on classification accuracy. The results demonstrate that meticulous pre-processing is paramount. A comparative assessment of machine learning models found that a Bayes classifier and a Support Vector Machine provided the most stable and accurate classifications when using an independent validation set. The study confirms that a combined feature space integrating mean spectral values with object-based geometric features (e.g., Asymmetry, Compactness) significantly outperforms models based on spectral features alone. Furthermore, a segmentation scale sensitivity analysis identified a critical trade-off between object detail and classification accuracy, while model transferability tests confirmed that "zero-shot" classification between scenes is unreliable due to illumination and domain shift. This thesis establishes a validated, end-to-end OBIA workflow for processing ground-based hyperspectral data for bridge monitoring. It provides a novel diagnostic framework for overcoming common geometric and radiometric pre-processing errors. The findings provide powerful empirical evidence that classification accuracy is not only dependent on the choice of algorithm but is critically sensitive to segmentation scale and the integration of geometric features to resolve spectral ambiguities in complex