Congratulations to our MSc student, Nikos Koroniadis, on the successful completion of his Master’s thesis!
His research explores how multispectral, thermal, and geometric information can significantly improve the automated classification of urban land cover from UAV imagery using deep learning.
The study compares a conventional RGB semantic segmentation model with an enhanced 7-band DeepLabV3 framework that integrates RGB, RedEdge, Near-Infrared, Long-Wave Thermal imagery, and a normalized Digital Surface Model (nDSM) derived from MicaSense Altum-PT data.
Using a challenging urban dataset acquired over Pamfila, Lesvos, the multidimensional model achieved remarkable improvements:
• Overall Accuracy: 93.90% (vs. 83.29% using RGB)
• Macro-F1 Score: 93.99% (vs. 83.71%)
• Cohen’s Kappa: 92.89% (vs. 82.09%)
Beyond the impressive performance gains, the research provides important scientific insights:
🔹 The nDSM proved to be the most influential feature for accurate building detection.
🔹 The thermal infrared channel substantially improved vehicle detection and information recovery in shadowed areas.
🔹 The greatest benefits of multidimensional imagery extended well beyond vegetation discrimination.
To ensure robust conclusions, the comparison employed rigorous statistical validation, including paired sampling, the McNemar exact test, Holm–Bonferroni correction, and bootstrap uncertainty estimation.
An additional outcome of the thesis is an open and freely available ArcGIS Pro toolbox that automatically converts semantic segmentation outputs into structured, topologically consistent, and attribute-enriched vector datasets, helping bridge the gap between AI-based image analysis and operational GIS workflows.
The research outputs are openly available at Outputs section. Congratulations, Nikos, on this excellent contribution to UAV remote sensing, deep learning, and geospatial data processing! We wish you every success in your future research and professional career.
