Scholarlink

<p>Satellite-derived water indices such as the Normalized Difference Water Index (NDWI) and the Modified NDWI (MNDWI) are widely used for surface-water monitoring, yet their direct forecasting as regional-scale time series remains underexplored. This study benchmarks machine-learning (ML) models and an Adaptive Neuro-Fuzzy Inference System (ANFIS) for forecasting regionally aggregated NDWI and MNDWI over Thessaly, Greece, across five geomorphological units (entire region, lowlands, highlands, lakes, and rivers) using multi-year Sentinel-2 time series processed in Google Earth Engine. Two experimental configurations are evaluated: (A) models driven only by contemporaneous hydroclimatic predictors (NDVI, precipitation, and thermal variables), and (B) models augmented with lagged index terms to capture temporal persistence. In Experiment A, ANFIS achieves the lowest errors across most subregions, demonstrating strong ability to model nonlinear environmental relationships without temporal memory. In Experiment B, the inclusion of lagged predictors yields substantial accuracy gains for both indices under the benchmarked imputed daily time-series setting, and learning-based models consistently outperform the corresponding persistence and mean baselines. Random Forest provides the most stable NDWI forecasts, particularly in lakes and rivers, while MNDWI performance is more region-dependent, with XGBoost and Polynomial Regression excelling in different contexts. Moving block bootstrap confidence intervals and Diebold–Mariano tests confirm statistically significant improvements within this benchmarked setting. However, observed-only one-step-ahead experiments showed substantially weaker skill, highlighting the dependence of operational forecasting performance on temporal continuity and missing-data treatment. Overall, NDWI exhibits stronger spatial robustness, whereas MNDWI shows greater sensitivity to land-cover heterogeneity and extreme hydrological conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Benchmarking Machine Learning and ANFIS Models for Forecasting NDWI/MNDWI Time-Series as Spectral Proxies of Surface Water in Region of Thessaly, Greece

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Maria Drogkoula,
  • Nicholas Samaras,
  • Omiros Iatrellis,
  • Konstantinos Kokkinos,
  • Eftihia Nathanail

摘要

Satellite-derived water indices such as the Normalized Difference Water Index (NDWI) and the Modified NDWI (MNDWI) are widely used for surface-water monitoring, yet their direct forecasting as regional-scale time series remains underexplored. This study benchmarks machine-learning (ML) models and an Adaptive Neuro-Fuzzy Inference System (ANFIS) for forecasting regionally aggregated NDWI and MNDWI over Thessaly, Greece, across five geomorphological units (entire region, lowlands, highlands, lakes, and rivers) using multi-year Sentinel-2 time series processed in Google Earth Engine. Two experimental configurations are evaluated: (A) models driven only by contemporaneous hydroclimatic predictors (NDVI, precipitation, and thermal variables), and (B) models augmented with lagged index terms to capture temporal persistence. In Experiment A, ANFIS achieves the lowest errors across most subregions, demonstrating strong ability to model nonlinear environmental relationships without temporal memory. In Experiment B, the inclusion of lagged predictors yields substantial accuracy gains for both indices under the benchmarked imputed daily time-series setting, and learning-based models consistently outperform the corresponding persistence and mean baselines. Random Forest provides the most stable NDWI forecasts, particularly in lakes and rivers, while MNDWI performance is more region-dependent, with XGBoost and Polynomial Regression excelling in different contexts. Moving block bootstrap confidence intervals and Diebold–Mariano tests confirm statistically significant improvements within this benchmarked setting. However, observed-only one-step-ahead experiments showed substantially weaker skill, highlighting the dependence of operational forecasting performance on temporal continuity and missing-data treatment. Overall, NDWI exhibits stronger spatial robustness, whereas MNDWI shows greater sensitivity to land-cover heterogeneity and extreme hydrological conditions.