This paper presents a semi-automatic procedure to discriminate seasonally flooded areas in the shallow temporary marshes of Doñana National Park (SW Spain) by using a radiommetrically normalized long time series of Landsat MSS, TM, and ETM+ images (1974–2014). Extensive field campaigns for ground truth data retrieval were carried out simultaneous to Landsat overpasses. Ground truth was used as training and testing areas to check the performance of the method. Simple thresholds on TM and ETM band 5 (1.55–1.75 μm) worked significantly better than other empirical modeling techniques and supervised classification methods to delineate flooded areas at Doñana marshes. A classification tree was applied to band 5 reflectance values to classify flooded versus non-flooded pixels for every scene. Inter-scene cross-validation identified the most accurate threshold on band 5 reflectance (ρ