| # Terms in plain language | |
| - **Prototype (three training-feature cluster centres per crop class):** K-means centres learned from balanced training-pixel embeddings of the first temporal block. Standardized, normalized embeddings are compared with the centres; temperature-scaled similarities and log-sum-exp form class scores, then softmax gives probabilities. It adds no labeled patches and shares the SAM encoder. | |
| - **SAM optimizer (sharpness-aware minimization):** an optimization method; this project does not use Segment Anything foundation features. | |
| - **New CNN control (a separately trained plain per-pixel temporal CNN):** the same 28,306-parameter architecture as the old temporal CNN, trained under the newer matched-control RNG protocol. A control for attention/Transformer experiments, not a new large backbone. | |
| - **Blend (probability-weighted prediction ensemble):** combine class probabilities, then take argmax. No checkpoint-weight averaging or majority vote. | |
| - **Highest observed test configuration:** attention-CNN addition at seed3407, 49.5298% test. It is kept separately because its validation is below the recommendation. | |
| - **Validation-selected recommendation:** new plain-CNN addition at seed3407, 52.5640% validation and 49.4555% test. | |