Monitoring Freshwater Bodies over the Past 40 Years Using Synthetic Monthly Sentinel-2 MSI Imagery

In the last decade rapid advancements in remote sensing have opened new frontiers in our ability to monitor freshwater bodies dynamics at the global scale. Most works have taken advantage of the long time series of Landsat constellations (30 m resolution) relying on spectral indices to identify water. Recently, much progress has also been made in the development and use of deep learning models capable of explicit semantic classification of river water, lake water and sediment bars, based on Sentinel-2 (S2) MSI imagery (10 m resolution). In this work, we present an approach that seeks to extend these existing, trained, fluvial landscape classification models to Landsat data in order to observe long-term water and morphological shifts in rivers and lakes. Rather than explicitly re-training the models with Landsat data and labour-intensive manual label data, we apply a domain transfer approach to generate synthetic S2 MSI imagery from Landsat inputs. This approach has the advantage that the training of deep learning domain transfer models only requires synchronous Landsat and Sentinel data and thus obviates the need for manual labels.

The results show that, when using these synthetic images, river water, lake water and sediment bars are classified with an F1 score of 0.8, 0.94, 0.65 respectively, which represents a decrease of ca. 10% for river water and 20% for sediment with respect to real S2 imagery. By adopting this integrated approach, we are therefore able to monitor, for the first time, lake water, river water and sediment bars at 10 m resolution, over a 40-year period, integrating both synthetic S2 and real S2 acquisitions through a single, fluvial landscape segmentation model. Classification obtained from median monthly images can then be aggregated at the yearly or multi-yearly scale to delineate river or lake water fluctuations, and active channels (river water plus sediment bars) trajectories, from specific freshwater bodies to the global scale.