Establish the geographic and seasonal generality of late-season encoding failures

Establish whether the late-season wildfire-detection failure of calendar-year annual embeddings is specific to United States fire regimes, characterize its behavior in Southern Hemisphere fire seasons, and determine whether shifting the summary window or pre-training on temporally diverse disturbance data mitigates the failure.

Background

The study reports that burned-area recall declines sharply for fires ignited in November and December, particularly for Tessera, and that this relationship persists after accounting for fire characteristics and geography. The authors discuss limited post-fire observations, sparse temporal sampling, and possible under-representation of late-season disturbances in pre-training data as potential explanations.

Because calendar-year alignment may correspond poorly to fire seasons in regions such as Southern Australia, the authors identify unresolved questions about whether the limitation generalizes beyond the United States and whether alternative summary windows or more diverse pre-training data can alleviate it.