Site-Dependent Performance of Spectral Indices and Machine Learning for Flood Detection in Mediterranean Karst Poljes
2026 IEEE Mediterranean and Middle-East Geoscience and Remote Sensing Symposium, M2GARSS 2026, Hybrid, Marrakech, Morocco, 22 - 24 April 2026, pp.36-40, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/m2garss67833.2026.11582324
- City: Hybrid, Marrakech
- Country: Morocco
- Page Numbers: pp.36-40
- Keywords: flood detection, karst polje, machine learning, Sentinel-2, spectral indices, transferability
- Istanbul University Affiliated: Yes
Abstract
Karst poljes experience recurrent flooding, yet optimal remote sensing detection methods remain poorly documented across diverse environments. This study evaluates six approaches-three spectral indices (Modified Normalized Difference Water Index [MNDWI], Sentinel Water Index [SWI], and Automated Water Extraction Index-shadow [AWEI-sh]) and three machine learning (ML) classifiers (Random Forest [RF], XGBoost, and Support Vector Machine [SVM])-across five morphometrically diverse poljes in southwestern Turkey using Sentinel-2 imagery (2016-2024). Results reveal pronounced site-dependent performance, with ML advantages ranging from negligible (0-1%, simple agriculture) to substantial (18.6%, greenhouse sites). MNDWI achieved 98% accuracy in urban areas but degraded to 56.9% where greenhouse plastic created spectral confusion, while ML methods maintained 95-96% accuracy through multi-feature integration. No single method universally outperforms others-site-specific land use complexity determines method selection, contradicting the common assumption that machine learning always performs best.