Modelling Carbon and Groundwater in Peatlands using Alpha Earth Embeddings

Spatially explicit knowledge of soil organic carbon and groundwater variability in peatlands is essential to support climate action, such as planning and implementing restoration projects. Geospatial machine learning is a key tool to obtain such knowledge at high spatial resolution, based on linking maps of explanatory variables with point information to train regression or classification models. Explanatory variables are usually identified through expert knowledge and derived from satellite remote sensing, digital elevation models or maps of soil types or land use. Identifying and processing relevant explanatory variables at large scale is non-trivial and cumbersome.

Geospatial foundation models, such as Google’s Alpha Earth change how satellite data and other geospatial data can be utilized in downstream machine learning tasks. Such models provide analysis-ready unified layers, i.e. embeddings, that are semantically rich representations capturing the underlying input data. In the case of Alpha Earth, input data cover archives of Sentinel-1 and Sentinel-2 as well as other geospatial data sources.

In the present study we introduce Alpha Earth embeddings into the modelling of soil organic carbon and groundwater across Danish peatlands at 10 m resolution. We use existing datasets and models of the two variables to benchmark the potential of foundation models for low-barrier large-scale modelling. The models trained against solely Alpha Earth embeddings are contrasted with models trained against explanatory variables selected through expert knowledge as well as with hybrid models combining basic topographical variables with Alpha Earth embeddings.

The Alpha Earth model of soil organic carbon produces meaningful spatial patterns while having a 6% decrease in performance (RMSE) with respect to the expert model. The true positive rate to predict peaty and peat soils is 0.68 and 0.65 for the expert and Alpha Earth model, respectively. The hybrid model increases the performance slightly with respect to the Alpha Earth model and all models achieve very comparable result of mapping the overall peat extent.   

The Alpha Earth model predicting groundwater has a 3% performance decrease with respect to the expert model (RMSE). When introducing synthetic training data for drained and wet conditions to the groundwater model, the Alpha Earth model shows limited performance. However, the hybrid model can utilize the synthetic data in a more meaningful way and achieves satisfactory results with respect to performance and spatial patterns.

In addition, we carry out feature importance analysis to explain the Alpha Earth embeddings, which is clear limitation in the usage of foundation models, where explainability is typically not provided.