Biodiversity and forest fragmentation assessment based on aboveground biomass prediction produced by fine-tuned geospatial foundation models

Abstract

Global biodiversity is rapidly declining and human impacts influence the distribution and abundance of wild species, the distinctness of ecological communities, and the integrity of ecosystems. Satellite imagery has huge potential for tracking progress towards conservation targets, but satellite imagery based products are only a proxy for biodiversity. In this context, we propose to use satellite image based indices , Rao-index and Shannon Entropy index, calculated over a certain area and track the change in index values over time to identify areas that remain stable or experience drastic changes. Moreover we use an adjacent reference area, that is affected by deforestation or wildfire, to show the relative difference between the two regions. Geospatial foundation models are designed to replace task-specific models and therefore are trained in a self-supervised manner on large datasets of unlabelled satellite imagery which can then be fine-tuned to a wide range of geospatial downstream tasks. In this work, we aim to use the heterogeneity measures of spatial aboveground biomass estimates as a proxy of species diversity and forest fragmentation for the Karukinka region in Chile and Water Towers in Kenya over the years.  To predict aboveground biomass in this region, we fine-tune the geospatial foundation models on space-borne (i.e., GEDI measurements) collected across similar eco-regions. The yearly generated above ground biomass is used as a base layer for biodiversity index calculations and assess the quality of biomass using forest fragmentation, forest connectivity and biodiversity indices. In this approach major landscape changes like wildfire, deforestation or tree health can be monitored in near real time tracking progress of global biodiversity framework goals.