Improving the remote sensing frameworks related to land cover mapping is necessary to make informed policy, development, planning, and natural resource management decisions. These efforts are especially important in tropical countries where technical capacity is limited. Land cover legend specification is a critical first step when mapping land cover, with consequences for its subsequent use and interpretation of results. We integrated the temporal metrics of SAR (Synthetic Aperture Radar) and multispectral data (Sentinel-1 and Sentienel-2) with visual pixel classifications and field surveys using five machine learning algorithms that apply different statistical methods to assess the prediction and mapping of two different land cover legends at a high spatial resolution (10 m) in a tropical region with seasonal flooding. The evaluated legends were CORINE (Coordination of Information on the Environment) and ECOSO, a legend that we defined based on the ecological and socio-economic conditions of the study area. Compared with previous studies, we obtained high accuracies for land cover modeling (kappa = 0.82) and land cover mapping (kappa = 0.76) when using ECOSO. We also found that the CORINE legend generated lower accuracies than the ECOSO legend (kappa = 0.79 for land cover modeling and kappa = 0.61 for the land cover mapping). Although CORINE was developed for European environments, it is the official land cover legend of Colombia, a South American country with tropical ecosystems not found in Europe. Therefore, some of the CORINE classes have ambiguous definitions for the study area, explaining the lower accuracy of its modeling and mapping. We used free and open-access data and software in this research; thus, our methods can be applied in other tropical regions.