Phase-Informed Cross-Variable Transformer for County-Level Corn Yield Estimation

Minglin Zou

doi:10.1109/lgrs.2026.3743156

Phase-Informed Cross-Variable Transformer for County-Level Corn Yield Estimation

2026-10-14

0

PRE

AI

DOI：10.1109/lgrs.2026.3743156

Original

Original request for help

Save

Abstract

En 中文

Accurate county-scale crop yield estimation requires integrating heterogeneous environmental observations while preserving the distinct information carried by vegetation, meteorological, and soil-water variables. We propose PICVT, a cross-variable Transformer that combines static contextual routing, seasonal phase encoding, and variable-specific tokenization before cross-variable interaction. The framework is evaluated using U.S. corn yield records from 2008–2024, comprising 196,645 grid-cell-year samples corresponding to 1,929 county-year yield records. Under five-fold forward-chaining evaluation, PICVT achieves a county-level R² of 0.753±0.015 and an RMSE of 1.153 t/ha, outperforming the evaluated machine-learning and deep-learning baselines under identical inputs and temporal splits. Within-season performance increases to R²=0.746 by August, with little additional gain through October. The lower NDVI-only baseline indicates that vegetation dynamics carry substantial yield-related information while thermal, atmospheric, and soil-water observations provide complementary signals. Ablation and attention analyses further support distinct roles for static context, seasonal ordering, and cross-variable interaction. These results support structured multisource representation for county-scale yield estimation with progressive within-season updating.

Keywords:

Crop yield estimation

multisource remote sensing

within-season prediction

cross-variable attention

variable-centric Transformer

Phase-Informed Cross-Variable Transformer for County-Level Corn Yield Estimation

Abstract

Journal

Organization

Cited Papers