Multi-variety rice yield prediction from heterogeneous UAV time-series data via a cross-period attention transformer

Harmonisation of UAV time-series of heterogenous length has long challenged crop yield prognostics in trials contrasting multiple varieties. Through derivation of an innovative Transformer model, Cross-Period Attention Transformer for Rice (riceCPAT), we invoke time-series data from over 500 varieties grown in the field (EXP1&2). We developed riceCPAT employing a hierarchical processing strategy where a standard Transformer encoder first extracted shallow features, retaining complete time-series information. After adaptive segmentation based on phenology, average pooling was conducted to compress segments of heterogeneous length into fixed-dimension vectors. RiceCPAT significantly reduced RMSE (67.96 g m −2 ) and MAE (53.52 g m −2 ) with the fewest parameters (72,561), lowering computational complexity and parameter count, outperforming other methods in prognostic accuracy and computational efficiency. Transfer validation across three independent experimental sets (EXP3-5) spanning diverse ecological sites achieved an average R 2 of 67%. Post-hoc model-agnostic feature analysis revealed that EVI2 and MCARI were key features in yield prediction, with grain filling being the most important prognostic. We contend that riceCPAT effectively eliminates information loss inherent in traditional alignment methods, offering an efficient, scalable solution for multi-variety yield prediction.