Floodplain areas belong to the most diverse, dynamic and complex ecological habitats of the terrestrial portion of the Earth. Spatial and temporal quantification of floodplain dynamics is needed for assessing the impacts of hydromorphological controls on river ecosystems. However, estimation of land cover dynamics in a post-classification setting is hindered by a high contribution of classification errors. A possible solution relies on the selection of specific information of the change map, instead of increasing the overall classification accuracy. In this study, we analyze the capabilities of Unmanned Aerial Systems (UAS), the associated classification processes and their respective accuracies to extract a robust estimate of floodplain dynamics. We show that an estimation of dynamics should be built on specific land cover interfaces to be robust against classification errors and should include specific features depending on the season-sensor coupling. We use five different sets of features and determine the optimal combination to use information largely based on blue and infrared bands with the support of texture and point cloud metrics at leaf-off conditions. In this post-classification setting, the best observation of dynamics can be achieved by focusing on the gravel-water interface. The semi-supervised approach generated error of 10% of observed changes along highly dynamic reaches using these two land cover classes. The results show that a robust quantification of floodplain land cover dynamics can be achieved by high-resolution remote sensing.