[Discussion] sample in thermal time

[robmarkcole]

For many ag applications, we want to sample in thermal time, not calendar time, see https://arxiv.org/abs/2506.12885
I've scoped the approach below but am interested to hear if (1) this is a feature that could be considered to incorp in rslearn and (2) if there are any simpler alternatives.
Thanks

Possible approach: Compute GDD stage boundaries externally, create one rslearn window per stage interval, and attach metadata (e.g., field_id, year, stage_idx) so windows can later be grouped. During training, group and sort these stage windows by a shared key (e.g., (field_id, year)) and stack them to form a single stage-aware temporal sample without modifying rslearn’s core pipeline.

[robmarkcole]

@favyen2 any thoughts? I expect this is too niche to be implemented directly in rslearn, but good to hear your thoughts

[robmarkcole]

An approach is outlined in https://arxiv.org/pdf/2604.19217

The main innovation of this work is the incorporation of an Attention Layer in order to process multi-season data. The attention mechanism does not assign time steps the same weight, instead it assigns each time step in the growing season a score of importance which is represented as the alpha t... This allows the model to prioritize critical phenological windows, such as the anthesis stage in cereals, which are more sensitive to climate fluctuations

This sounds more general than the approach I suggested, as many time series tasks will have more critical phases than other