Mechanistic Interpretability for Mapping Ecosystem Functioning

Mapping ecosystem properties and functioning from Earth observation data remains fundamentally an extrapolation problem. Ground-based measurements of ecosystem processes, such as carbon fluxes, are sparse and geographically biased towards the Global North. Machine-learning models are therefore trained on limited labeled data and subsequently upscaled globally using environmental covariates derived from satellite remote sensing and reanalysis products. A central challenge is ensuring that such models generalize robustly beyond their training domain, rather than exhibiting spurious confidence or biased predictions in poorly observed regions.

In this contribution, we explore how recent advances in mechanistic interpretability and self-supervised representation learning from AI safety research can help address these challenges. In particular, sparse autoencoders (SAEs), and more specifically Top-K sparse autoencoders, have recently been used to disentangle deep neural representations into interpretable and steerable concepts in large language models. We propose to adapt these methods to Earth system science, with the goal of learning sparse, disentangled, and spatially meaningful latent representations of ecosystem-relevant variables.

We first evaluate this approach on a self-supervised proxy task: compressing and reconstructing mean seasonal cycles derived from MODIS remote sensing products and ERA5 climate reanalysis data. Using a Matryoshka BatchTopK SAE, we obtain latent features that are highly localized in space, with individual features activating only over specific regions of the Earth. In contrast to dense embeddings, e.g. from variational auto-encoders, our approach offers a control on the average sparsity level. In other words, this intrinsic, data-driven partitioning of geographic space can be interpreted as emergent climate regimes or ecosystem types, without relying on predefined biome maps or expert labels. 

Building on these results, we apply the SAE framework to the mapping of ecosystem carbon fluxes, using FluxNet tower observations as ground truth. The sparse and disentangled latent structure provides a transparent link between remote sensing inputs and predicted ecosystem functioning. Simultaneous training on a self-supervised reconstruction task and on predicting net ecosystem exchange provides competitive performance, with the sparsity of the features offering a promising avenue to enhance robustness by controlling the extrapolation behavior of the neural network. Beyond predictive performance, we introduce an interpretability workflow that enables systematic inspection of learned features, supporting model diagnostics and scientific analysis.

Overall, we argue that self-supervised, interpretable representation learning offers a promising pathway toward robust global ecosystem mapping from both labeled and unlabeled satellite data. This approach leverages the full scale of Earth observation archives while improving trust and insight in mapping ecosystem properties and functioning. In addition, it sheds insight into geographical partitioning, offering a novel perspective on decade-old maps of plant functional types.