Change detection of the Karukinka Nature Park peatlands ecosystem using multimodal foundation models

The Karukinka Nature Park, in Southern Chile, is home to some of the world’s most ecosystems. None more so than the extensive peatlands, that are responsible for storing hundreds of millions of tons of carbon. However, the increasing prevalence and threat of wildfires, and human activity amongst other, have highlighted the need for increased monitoring. The exact areal extent of the peatlands are not well-defined, across the region. Here, we present an effort to map the peatlands in the Karukinka Nature Park, and determine the change in peatlands extent, if any. We demonstrate the use of several geospatial foundation models, such as Terramind and Prithvi. Relying on open-source earth observation data, such as Sentinel 1 and 2, we fine tune these geospatial foundation models to detect peatlands land cover. Our efforts highlight challenges with defining high quality ground-truth for peatlands landcover type, at high resolution. From a modelling perspective, multimodal models outperform unimodal models, demonstrating the value of data such as Sentinel 1. While model performances are generally satisfactory against the ground-truth, determining and validating true change at a high resolution remains a challenge.