We can automatically classify river geomorphic features from Sentinel 2 images - but what about their uncertainty?

Machine learning models that automatically delineate river geomorphic features on Sentinel 2 (S2) images have the potential to provide a weekly monitoring of their dynamics and a better understanding of the underlying river channel processes. The accuracy (e.g. 95%) of these feature delineations is generally assessed by quantifying the percentage of pixels of known nature correctly classified by the model. However, the pixels used for such calculations are often sampled within the classified satellite image (with a resolution of 10m of larger) laying shadow on the real, relative extent of the misclassified pixels (e.g. the remaining 5%) usually located at the borders between features, which unavoidably lead to the under or overestimation of one feature for another. This issue raises questions on the real extent of the geomorphic features measured or on the true geomorphic temporal change that can be detected. In this work, we identified the nature and extent of the misclassified pixels on S2 images of a section of the river Sesia (North Italy) by comparing the classes of water, sediment and vegetation automatically delineated by a machine learning model with those manually delineated in higher resolution images: Planet at 3m, and aerial orthophotos at 0.3m resolution. Assuming the orthophoto as error-free, we found that: (1) in both S2-based automatic classification and Planet-based manual classification, water is underclassified and that (2) the error of the misclassified area is insensitive to the spatial resolution, with the water class ~20% underestimated in both the S2 (10m) and the Planet (3m). By considering the period between 2018 and 2022, we also demonstrated that the active channel (water + sediment) trajectory assessed by using the S2 images on a weekly basis is comparable to the trajectory determined using the Planet or aerial orthophotos on a yearly basis. However, the frequent image acquisition of the S2 was able to capture the river corridor abrupt response and prompt recovery to a major flood in 2019, overlooked in the other two image sources. This work therefore shows that once the spatial uncertainty is quantified (e.g. 20% for the water class), the frequent image acquisition of the S2 provides a robust reconstruction of the river geomorphic trajectories as well as a better interpretation of the river processes, in particular recognising transient states in between significant events.