Akademik Veri Yönetim Sistemi
ECOLOGICAL INFORMATICS, cilt.0, sa.96, ss.1-25, 2026 (SCI-Expanded, Scopus)
Accurate surface water detection in heterogeneous wetlands remains a significant challenge due to spectral confusion arising from mixed pixels, dense aquatic vegetation, and fluctuating turbidity. High-precision hydrological mapping is indispensable for evaluating the success of wetland restoration initiatives and managing complex hydroperiods. Although the Normalized Difference Water Index (NDWI) is a standard tool, selecting a robust, unsupervised thresholding method capable of adapting to diverse environmental conditions is essential for effective operational monitoring. To address the disconnect between image segmentation and operational ecological management, this study evaluates the comparative performance of 18 thresholding algorithms applied to high-resolution unmanned aerial vehicle (UAV)-derived NDWI data. To overcome the limitations of single-site evaluations and ensure practical applicability, a dual-site validation strategy was implemented. A primary analysis was conducted at a reference site (Site A) using spectrally pure pixels, followed by rigorous validation at an independent test site (Site B) that explicitly incorporated mixed pixels, transition zones, and shadowed areas. Binary water/non-water masks were assessed through nine quantitative metrics: overall accuracy (OA), Kappa, F1-score, Matthews correlation coefficient (MCC), intersection over union (IoU), balanced accuracy (Bal. Acc.), specificity, precision, and recall. Percentile (60), K-Means, and Mean thresholding delivered the highest and most consistent performance across both sites. Notably, the Percentile (60) method showed superior robustness against mixed pixels characteristic of wetlands undergoing active restoration, achieving a strong precision–recall balance (F1-score > 0.93) at the independent test site. Although the K-Means and Mean methods attained high recall, they exhibited slightly lower specificity within complex transition zones. Statistical analyses, including analysis of variance (ANOVA) and bootstrap resampling, confirmed significant performance differences. Consequently, this study establishes a robust methodological framework for selecting automated, computationally efficient thresholding algorithms that enhance the precision of UAV-based wetland monitoring, providing an essential tool for adaptive management and the rigorous assessment of ecohydrological restoration outcomes.