Determine which foundation-model design choices control wildfire-disturbance sensitivity

Determine which design choices—including spatial context, temporal sampling, input modalities, training targets, and loss functions—make annual Earth-observation embeddings more sensitive to abrupt wildfire disturbance, thereby explaining the differing disturbance sensitivity of Tessera and AlphaEarth.

Background

The paper finds that Tessera and AlphaEarth produce markedly different wildfire signals despite both being annual temporal embeddings incorporating Sentinel-1 and Sentinel-2 observations. Tessera makes burned and unburned pixels more directly separable, whereas AlphaEarth benefits more from spatial decoding and explicit multi-year temporal context.

The authors identify several confounded differences between the models, including training data, sensors, spatial context, architecture, loss functions, temporal sampling, and training targets. They state that systematic ablations are needed to determine which of these factors drives the observed difference in sensitivity to abrupt disturbance.

References

The source of this difference remains unclear, as the models differ in training data, input sensors, spatial context, architecture, loss functions, temporal sampling, and training targets. One notable pattern is that adding spatial context through U-Net models improved performance much more for AlphaEarth than for Tessera, despite AlphaEarth embeddings already containing patch-level spatial context. The pattern suggests that when a pixel's annual trajectory already contains a clear disturbance signal, strong temporal encoding reduces dependence on spatial decoding. AlphaEarth's latent projections also seem more strongly organized by land cover in the fire year, which may partly reflect differences in training objectives, including its use of land-cover information as a target. Without systematic ablations, however, we cannot attribute the observed differences to any single design choice. Future work should isolate the effects of spatial context, temporal sampling, input modalities, training targets, and loss functions to determine what makes annual embeddings more sensitive to abrupt disturbance.

— Annual Earth-observation embeddings encode wildfire disturbance and support simplified burned area mapping  (2609.25731 - Knezevic et al., 22 Sep 2026) in Section "Disturbance sensitivity differs across foundation models"