Geospatial Intelligence and Machine Learning Technique for Urban Mapping in Coastal Regions of South Aegean Volcanic Arc Islands

Pavlos Krassakis, Andreas Karavias, Paraskevi Nomikou, Κωνσταντίνος Καράντζαλος, Nikolaos Koukouzas, Stavroula Kazana, Issaak ParcharidisGeomatics 2022 · DOI

TL;DR

This paper introduces a novel geospatial intelligence (GEOINT) workflow using machine learning to map urban expansion and delineate coastal zones in the unique volcanic islands of the South Aegean. By combining satellite imagery with advanced techniques like Satellite-Derived Bathymetry (SDB) and Random Forest classification, the study achieves high accuracy in urban mapping and identifies significant urbanization trends, crucial for sustainable coastal management.

What Problem Does This Paper Address?

Coastal regions are vital economic hubs, particularly for tourism, but they are increasingly vulnerable to climate change impacts like sea-level rise and erosion, exacerbated by rapid urbanization. The South Aegean Volcanic Arc (SAVA) islands, with their distinctive architectural styles and volcanic landscapes, present unique challenges for accurate mapping using existing open-source datasets. These datasets often struggle with misclassification due to spectral similarities between urban areas, bare soil, and even clouds. This necessitates a more robust and tailored approach to monitor coastal zone boundaries, understand underwater topography, and accurately map urban development for effective planning and hazard mitigation.

Methods

The study employs a three-fold methodology:

  1. Coastal Zone Delineation: This involves defining the seaward boundary using a 10m isobath derived from Satellite-Derived Bathymetry (SDB) and the landward boundary using a 20m elevation contour. The coastline itself is delineated using the Normalized Difference Water Index (NDWI) from Sentinel-2 imagery.
  2. Satellite-Derived Bathymetry (SDB): An adapted GEBCO-like workflow is used to estimate water depths in shallow coastal areas. This involves processing Sentinel-2 multispectral imagery and using a log-band ratio method to derive bathymetric data, which is then calibrated against known depth points.
  3. Urban Mapping with Machine Learning: A Random Forest (RF) algorithm is implemented to classify urban areas. This approach is chosen for its high accuracy, especially with high-resolution data. The methodology includes pre-processing satellite images (atmospheric correction, cloud masking – though omitted in this specific case due to unique landscape features), resampling, and spectral/spatial subsetting. The RF model is trained and validated using ground truthing and statistical indicators like Overall Accuracy (OA) and Cohen's Kappa.

Datasets

The study utilizes a combination of open-access satellite and geospatial datasets:

  • Sentinel-2: Multispectral imagery (10m spatial resolution) from 2016-2021, used for coastline extraction, SDB, and urban mapping.
  • SRTM 1-sec (DEM): Shuttle Radar Topography Mission Digital Elevation Model (30m spatial resolution) from USGS, used for extracting contour lines.
  • Corine Land Cover (CLC) 2018: A Copernicus Land Monitoring Service dataset providing land cover information.
  • European Settlement Map (ESM) 2015: A Copernicus dataset mapping human settlements, used for comparison and validation.
  • OpenStreetMap (OSM): A global geodatabase providing road network data, used to refine urban area extraction.

Key Insights and Ideas

  • Addressing Unique Landscape Challenges: The study highlights how the distinctive Cycladic architecture (white buildings) and volcanic terrain can lead to misclassification in standard datasets like ESM, necessitating a tailored ML approach.
  • SDB for Coastal Management: The successful application of SDB provides crucial bathymetric data for shallow coastal zones, which is essential for understanding nearshore dynamics and potential flood risks.
  • Random Forest's Efficacy: The RF algorithm proves effective in accurately mapping urban areas, even in complex environments, outperforming existing datasets like ESM in certain aspects.
  • Urbanization Trends: The research quantifies significant urbanization expansion in specific settlements, providing critical data for future planning and climate change adaptation strategies.

Key Results

  • Urban Mapping Accuracy: The developed methodology achieved an average Kappa index of 80.15% for Thira and 88.35% for Milos.
  • Urbanization Expansion: An average urbanization expansion of over 22% was observed in specified settlements across both islands between 2016 and 2021.
  • SDB Performance: The SDB method showed good correlation with charted depths, with an R² of 0.76 for Thira and 0.67 for Milos, providing reliable depth estimates up to approximately 20m.
  • ESM vs. RF: The RF approach demonstrated a clearer distinction between urban environments and bare land compared to the ESM dataset, particularly in areas with challenging volcanic landforms.

Limitations and Future Work

  • SDB Optimization: While effective, the SDB methodology is described as empirical and could be further optimized using very high-resolution imagery, more extensive ground truth data, and more accurate control points. The correlation coefficients (R²) were noted as not being very high, potentially due to georeferencing errors in bathymetric maps, water turbidity, and satellite data resolution.
  • Spatial Resolution of Sentinel-2: The study acknowledges that Sentinel-2's 10m resolution can affect the detection of very small buildings, leading to potential underestimation of urban extent in some cases.
  • Future Work: The authors suggest future research could focus on:
  • Developing a Multiple Criteria Decision Making (MCDM) approach for identifying vulnerable areas.
  • Correlating urbanization expansion with existing land-use plans or predicting future population density.
  • Further optimizing SDB and ML techniques with higher resolution data.

Summary updated

Abstract

Coastal environments are globally recognized for their spectacular morphological characteristics as well as economic opportunities, such as fisheries and tourism industries. However, climate change, growth in tourism, and constant coastal urban sprawl in some places result in ever-increasing risk in the islands of the South Aegean Volcanic Arc (SAVA), necessitating thoughtful planning and decision making. GEOspatial INTelligence (GEOINT) can play a crucial role in the depiction and analysis of the natural and human surroundings, offering valuable information regarding the identification of vulnerable areas and the forecasting of urbanization rates. This work focuses on the delineation of the coastal zone boundaries, semi-automatization of Satellite-Derived Bathymetry (SDB), and urban mapping using a machine learning algorithm. The developed methodology has been implemented on the islands of Thira (Santorini island complex) and Milos. This study attempts to identify inaccuracies in existing open-source datasets, such as the European Settlement Map (ESM), as a result of the unique combination of the architectural style and bare-soil characteristics of the study areas. During the period 2016–2021, the average accuracy of the developed methodology for urban mapping in terms of the kappa index was 80.15% on Thira and 88.35% on Milos. The results showed that the average urbanization expansion on specified settlements was greater than 22% for both case studies. Ultimately, the findings of this study could contribute to the effective and holistic management of similar coastal regions in the context of climate change adaptation, mitigation strategies, and multi-hazard assessment.
BibTeX
@inproceedings{krassakis2022geospatial,
  author = {Pavlos Krassakis and Andreas Karavias and Paraskevi Nomikou and Κωνσταντίνος Καράντζαλος and Nikolaos Koukouzas and Stavroula Kazana and Issaak Parcharidis},
  title = {{Geospatial Intelligence and Machine Learning Technique for Urban Mapping in Coastal Regions of South Aegean Volcanic Arc Islands}},
  booktitle = {Geomatics},
  year = {2022},
  doi = {10.3390/geomatics2030017},
}
Figure