Analyzing the potential of geospatial foundation models and earth embeddings for assessing dynamic riverine processes

Satellite remote sensing has gained significant relevance in the monitoring of riverine processes over the past years. Time series of satellite imagery, e.g. from the Landsat or Sentinel-1 and -2 constellations, are of particular interest as they allow us to assess not only distinct landcover classes but also dynamic processes such as inundation duration. Over the recent years, the rise of AI supports the analysis of these big earth observation archives. The increasing volume of earth observation data and computational burden of large AI models, however, lead to challenges in data processing. In the remote sensing community, the use of foundation models, large neural networks trained on massive amounts of data, promise to reduce the computational burden for end users. In addition, raw data encoded by these models, so-called earth embeddings, provide a ready-to-use dataset which transforms complex, multitemporal data of hundreds of satellite scenes from different sensors into annual images with abstract numerical representations of the original data. While in the remote sensing community, there is an increasing number of studies evaluating the application of geospatial foundation models and embeddings for various applications, there is to the best of our knowledge no comprehensive evaluation of the use of these recent developments for assessing dynamic hydro-geomorphic patterns and processes of river corridors.

In our study, we assess the potential of two openly available embedding databases for their potential to represent typical riverine habitat types and river dynamics such as the inundation duration or erosion/formation features related to channel shifting. Specifically, we evaluate the recently released AlphaEarth embeddings (available on Google Earth Engine) and University of Cambridges TESSERA embeddings database (retrievable from their online database through a dedicated python package). Primary case study is the Naryn River in Kyrgyzstan. This is still a near-natural river on a length of more than 600 km. Along with an average active channel width of 400 m and average river corridor width of 1200 m, this makes the Naryn an ideal example for remote sensing applications in river science. We use a total number of 1873 ground truth points representing different geomorphic features and typical riparian habitat types as reference to evaluate how well embeddings are capable to distinguish these classes. To test the predictive capability of embeddings, we train supervised classification models based on the embeddings. In addition, we use satellite derived time series of inundation duration and inter-annual change derived from Planet Scope images to evaluate the potential of embeddings to represent dynamic characteristics and enable change detection. To analyze how this approach generalizes to other river systems, we apply the trained models to selected European rivers and validate the outcomes. Our initial results show a high potential of embeddings for analyzing riverscapes and their dynamics. We discuss how geospatial foundation models and embeddings as novel, AI driven tools in earth observation can contribute to generalizing remote sensing models across different river systems and how this can path the way towards global monitoring of riverscapes and their dynamics.