|Paper title||Application of Sentinel-2 and Landsat time series for sub-pixel river morphology monitoring|
|Form of presentation||Poster|
Fluvial and riparian ecosystems have many important ecological, social, and economic functions. Several EO-based tools and products have been developed for their monitoring at a global scale. However, the spatial resolution of these global-level products is often too coarse for monitoring narrow rivers and especially their upper sections. These are the areas where rivers are most dynamic and where frequent and accurate monitoring is particularly pressing. To overcome spatial resolution limitations, we developed a method for river monitoring using fraction maps produced with linear spectral signal unmixing.
We developed and tested the method on the Soča and Sava rivers in Slovenia and the Vjosa river in Albania. We mapped three land cover classes of interest – surface water, vegetation, and gravel. The use of spectral bands in combination with the NDVI, MSAVI2, NDWI, and MNDWI indices produced the best results. We achieved similar accuracies with endmembers selected manually and endmembers selected automatically with the N-FINDR algorithm. The optimal total number of endmembers used for spectral signal mixture analysis was found to be between three and five. A larger number of endmembers led to clustering of spectral signatures and thus redundant information. Tests showed that the inclusion of shade as a separate endmember did not improve fraction map accuracy. Furthermore, we found that endmembers selected manually or automatically on one satellite image can be successfully transferred to analyse another image acquired in a comparable geographic region and at a similar phenophase.
Results of the soft classification were compared to hard classification using the Spectral Angle Mapper with the same endmembers. Fraction maps were more accurate than maps based on hard classification both for Sentinel-2 and Landsat images. Water presence detected on the fraction maps was correlated with in situ measured water level and river discharge with Pearson’s r > 0.6 (p < 0.0001). We examined the ability of fraction maps to detect changes in river morphology. By looking at three different timestamps (13 October 2017, 3 July 2019, 5 September 2020), the results showed that fraction map differencing could distinguish changes in gravel deposition down to 400 m2 in extent. We found that change detection accuracy was best on pixel level when changes amounted to at least 30%. Finally, we tested the possibility of detecting river morphology changes from a time series of land cover presence based on fraction maps. The extents of water and gravel can vary following changes in water level. However, we found that a decrease of gravel bar size within two standard deviations of the mean indicated regular variations while a larger decrease pointed to gravel bar removal.
The developed method can be used for monitoring fluvial and riparian environments in highly heterogeneous areas. The main limitations of the method are associated with cloud obstruction and terrain shadow that are known problems of optical images. An interesting line of future investigations is to test the possible contribution of using SAR data for fluvial morphology monitoring.