Beyond open water: mapping the spectral and hydrological limits of Pampean wetlands classification with Landsat time series

Abstract

Conventional wetland mapping approaches using satellite remote sensing often exhibit a pronounced bias towards detecting open water bodies, while systematically underestimating vegetated wetlands and those with intermittent inundation regimes. This limitation becomes particularly acute in agricultural landscapes like Argentina’s Pampa region, where altered hydrology and vegetation patterns further complicate detection. Our study evaluates how wetland landscape configuration mediates classification performance across different sampling strategies, and assesses the capacity of the optimal approach to characterize wetland hydroperiods when compared against expert-derived maps. We applied Random Forest classification to the complete Landsat 5 and 8 archives (1984–2024) across six pilot sites representing distinct wetland patterns in Buenos Aires Province. A per-scene classification strategy was implemented to preserve ephemeral hydrological signals. The selected model (EG) demonstrated robust performance (AUC-ROC: 0.87) while revealing a critical limitation: wetlands with low wetness frequency ( < 0.2) or lacking surface water were less detectable than permanent open water bodies. Key findings highlight that while SWIR bands and modified water indices (MNDWI) were the most effective features, they still failed to detect non-inundated wetlands, while significant spectral confusion persisted in vegetated wetlands. When validated against manually delineated reference data, the wetness frequency maps proved valuable for identifying vulnerable wetlands in agricultural settings, though with noted constraints for ephemeral systems. Frequency analysis of 38-year results showed landscapes with isolated wetlands had 82% of expert-mapped wetlands below 0.2 wetness frequency, reflecting either methodological detection gaps or anthropogenic drying. In contrast, landscapes composed of mosaics of wetlands and terrestrial environments maintained higher frequencies, demonstrating greater hydrological resilience. This work provides critical insights for managing expectations about what satellitebased monitoring can achieve, particularly for ecosystems that rarely show free surface water, thereby providing a critical evidence base for more effective conservation policies and management strategies.