Over 80% of the Mediterranean coastline is rocky, consisting of complex erosional landforms where marine and subaerial processes interact. These environments have been classified into seven distinct geomorphotypes, each with unique topographic, lithostructural, and geomorphological properties (Biolchi et al., 2016a). The primary objective of this thesis is to use an extensive archive of high-resolution images to study these environments by reconstructing detailed 3D models, combining the geomorphological survey with Deep Learning techniques. The research focuses specifically on the intertidal zone of rocky coasts in the Mediterranean Sea, between high and low tide levels, where distinctive tidal notches develop within limestone formations. Antonioli et al. (2015) demonstrated that the Mediterranean region contains numerous tidal notches that exhibit considerable variation in morphological characteristics, dimensions, and elevational positions relative to mean sea level. Through the swim survey, an extensive dataset was collected using a custom-built raft equipped with comprehensive instrumentation. The methodology involves maintaining a constant swimming speed of 2 km/h while positioning the data collection platform as close as possible to the shoreline, keeping it parallel to the coast. This enables the systematic acquisition of time-lapse images and 360-degree videos above and below the waterline. Subsequently, a comprehensive organizational framework was established for the extensive dataset comprising 9 terabytes of coastal images, creating "cancelled" folders dedicated to each group of images. where unsuitable images for analysis were placed. This structured classification system provides a robust foundation for subsequent analytical procedures. To enhance data accessibility and collaborative research capabilities, the Geoswim Image Viewer platform was developed, able to display syncronyzed above and below waterline image with integrated geolocation metadata. Leveraging this organized dataset structure, two Convolutional Neural Networks (CNNs) were developed. The first network achieved exceptional performance with 98.85% accuracy in discriminating between useful and non-useful coastal images, while the second network demonstrated 80.6% accuracy in identifying five distinct coastal geomorphotypes including plunging cliffs, sloping coasts, shore platforms, screes, and pocket beaches, following the classification system established by Biolchi et al. (2016a). The most significant contribution of this research involves the generation of integrated three-dimensional models of the intertidal zone through the fusion of above and below waterline photogrammetric techniques. The methodology was tested through three representative case studies across the Mediterranean Sea, including Cala Uccello in Lampedusa, Mġarr ix-Xini in Gozo, and Halq it-Tafal in Malta. The 3D models enabled quantitative and semi-qualitative analyzes of tidal notch morphology across the study sites. At Mġarr ix-Xini, statistical analysis of twenty measurement points revealed mean notch dimensions of 2.34 ± 0.074 m in width and 1.26 ± 0.060 m in depth. The estimated eroded volume was 35.87 m³ along 15.5 m of coastline. In contrast, Halq it-Tafal exhibited morphological variability, with depth measurements ranging from 0.14 to 0.95 m and a mean depth of 0.47 ± 0.052 m. However, the study revealed some methodological limitations when applied to certain coastal morphologies, especially in the area of Cala Uccello where the coast is sloping coast type. The systematic progress through data organization, the implementation of AI, and the development of visualization and three-dimensional modeling creates a comprehensive picture that combines traditional analytical methods with the latest Deep Learning techniques.
Over 80% of the Mediterranean coastline is rocky, consisting of complex erosional landforms where marine and subaerial processes interact. These environments have been classified into seven distinct geomorphotypes, each with unique topographic, lithostructural, and geomorphological properties (Biolchi et al., 2016a). The primary objective of this thesis is to use an extensive archive of high-resolution images to study these environments by reconstructing detailed 3D models, combining the geomorphological survey with Deep Learning techniques. The research focuses specifically on the intertidal zone of rocky coasts in the Mediterranean Sea, between high and low tide levels, where distinctive tidal notches develop within limestone formations. Antonioli et al. (2015) demonstrated that the Mediterranean region contains numerous tidal notches that exhibit considerable variation in morphological characteristics, dimensions, and elevational positions relative to mean sea level. Through the swim survey, an extensive dataset was collected using a custom-built raft equipped with comprehensive instrumentation. The methodology involves maintaining a constant swimming speed of 2 km/h while positioning the data collection platform as close as possible to the shoreline, keeping it parallel to the coast. This enables the systematic acquisition of time-lapse images and 360-degree videos above and below the waterline. Subsequently, a comprehensive organizational framework was established for the extensive dataset comprising 9 terabytes of coastal images, creating "cancelled" folders dedicated to each group of images. where unsuitable images for analysis were placed. This structured classification system provides a robust foundation for subsequent analytical procedures. To enhance data accessibility and collaborative research capabilities, the Geoswim Image Viewer platform was developed, able to display syncronyzed above and below waterline image with integrated geolocation metadata. Leveraging this organized dataset structure, two Convolutional Neural Networks (CNNs) were developed. The first network achieved exceptional performance with 98.85% accuracy in discriminating between useful and non-useful coastal images, while the second network demonstrated 80.6% accuracy in identifying five distinct coastal geomorphotypes including plunging cliffs, sloping coasts, shore platforms, screes, and pocket beaches, following the classification system established by Biolchi et al. (2016a). The most significant contribution of this research involves the generation of integrated three-dimensional models of the intertidal zone through the fusion of above and below waterline photogrammetric techniques. The methodology was tested through three representative case studies across the Mediterranean Sea, including Cala Uccello in Lampedusa, Mġarr ix-Xini in Gozo, and Halq it-Tafal in Malta. The 3D models enabled quantitative and semi-qualitative analyzes of tidal notch morphology across the study sites. At Mġarr ix-Xini, statistical analysis of twenty measurement points revealed mean notch dimensions of 2.34 ± 0.074 m in width and 1.26 ± 0.060 m in depth. The estimated eroded volume was 35.87 m³ along 15.5 m of coastline. In contrast, Halq it-Tafal exhibited morphological variability, with depth measurements ranging from 0.14 to 0.95 m and a mean depth of 0.47 ± 0.052 m. However, the study revealed some methodological limitations when applied to certain coastal morphologies, especially in the area of Cala Uccello where the coast is sloping coast type. The systematic progress through data organization, the implementation of AI, and the development of visualization and three-dimensional modeling creates a comprehensive picture that combines traditional analytical methods with the latest Deep Learning techniques.
3D Models of Rocky Coasts using Structure-from-Motion (SfM) Technology Enhanced by Deep Learning
VACCHER, VALERIA
2025
Abstract
Over 80% of the Mediterranean coastline is rocky, consisting of complex erosional landforms where marine and subaerial processes interact. These environments have been classified into seven distinct geomorphotypes, each with unique topographic, lithostructural, and geomorphological properties (Biolchi et al., 2016a). The primary objective of this thesis is to use an extensive archive of high-resolution images to study these environments by reconstructing detailed 3D models, combining the geomorphological survey with Deep Learning techniques. The research focuses specifically on the intertidal zone of rocky coasts in the Mediterranean Sea, between high and low tide levels, where distinctive tidal notches develop within limestone formations. Antonioli et al. (2015) demonstrated that the Mediterranean region contains numerous tidal notches that exhibit considerable variation in morphological characteristics, dimensions, and elevational positions relative to mean sea level. Through the swim survey, an extensive dataset was collected using a custom-built raft equipped with comprehensive instrumentation. The methodology involves maintaining a constant swimming speed of 2 km/h while positioning the data collection platform as close as possible to the shoreline, keeping it parallel to the coast. This enables the systematic acquisition of time-lapse images and 360-degree videos above and below the waterline. Subsequently, a comprehensive organizational framework was established for the extensive dataset comprising 9 terabytes of coastal images, creating "cancelled" folders dedicated to each group of images. where unsuitable images for analysis were placed. This structured classification system provides a robust foundation for subsequent analytical procedures. To enhance data accessibility and collaborative research capabilities, the Geoswim Image Viewer platform was developed, able to display syncronyzed above and below waterline image with integrated geolocation metadata. Leveraging this organized dataset structure, two Convolutional Neural Networks (CNNs) were developed. The first network achieved exceptional performance with 98.85% accuracy in discriminating between useful and non-useful coastal images, while the second network demonstrated 80.6% accuracy in identifying five distinct coastal geomorphotypes including plunging cliffs, sloping coasts, shore platforms, screes, and pocket beaches, following the classification system established by Biolchi et al. (2016a). The most significant contribution of this research involves the generation of integrated three-dimensional models of the intertidal zone through the fusion of above and below waterline photogrammetric techniques. The methodology was tested through three representative case studies across the Mediterranean Sea, including Cala Uccello in Lampedusa, Mġarr ix-Xini in Gozo, and Halq it-Tafal in Malta. The 3D models enabled quantitative and semi-qualitative analyzes of tidal notch morphology across the study sites. At Mġarr ix-Xini, statistical analysis of twenty measurement points revealed mean notch dimensions of 2.34 ± 0.074 m in width and 1.26 ± 0.060 m in depth. The estimated eroded volume was 35.87 m³ along 15.5 m of coastline. In contrast, Halq it-Tafal exhibited morphological variability, with depth measurements ranging from 0.14 to 0.95 m and a mean depth of 0.47 ± 0.052 m. However, the study revealed some methodological limitations when applied to certain coastal morphologies, especially in the area of Cala Uccello where the coast is sloping coast type. The systematic progress through data organization, the implementation of AI, and the development of visualization and three-dimensional modeling creates a comprehensive picture that combines traditional analytical methods with the latest Deep Learning techniques.File | Dimensione | Formato | |
---|---|---|---|
Tesi PhD_Vaccher_20.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
19.46 MB
Formato
Adobe PDF
|
19.46 MB | Adobe PDF | Visualizza/Apri |
Tesi PhD_Vaccher_20_1.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
19.46 MB
Formato
Adobe PDF
|
19.46 MB | Adobe PDF | Visualizza/Apri |
Tesi PhD_Vaccher_20_2.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
19.46 MB
Formato
Adobe PDF
|
19.46 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/303793
URN:NBN:IT:UNITS-303793