This paper presents a compact empirical study of county-level corn-yield forecasting under a deliberately small-sample setting. The experiment uses the Iowa 2022 corn slice from a CropNet-style USDA crop-label table: a single state, a single crop, a single year, and 97 county observations. The forecasting task estimates held-out county yield in bushels per acre from regional and textual tokens that are available before the target is revealed: agricultural statistics district, parsed direction words, and county-level numeric identifiers. A climate-aware multimodal token-attention model (CAMT) is evaluated against mean, empirical-Bayes, regularized linear, tree-ensemble, gradient-boosting, and neural-network baselines. Across five repeats of five stratified folds, the best purely numeric baseline, ElasticNet with regional tokens, obtained 8.819 bu/acre MAE, 11.479 bu/acre RMSE, and 0.698 mean R2. CAMT obtained 8.822 bu/acre MAE, 11.484 bu/acre RMSE, and the same 0.698 mean R2, while providing token-level evidence that is directly converted into agronomic explanations. The study shows that most predictive information in this state-year slice is carried by regional climate geography: district and direction tokens reduced MAE by about 49% relative to the state-mean baseline. The explanation layer produced concise county rationales grounded in district yield levels and model predictions, allowing the numerical results to be interpreted without using production as a predictor.

Keywords: Crop yield forecasting; small-sample learning; CropNet; USDA crop labels; climate-aware modeling; multimodal transformer; agronomic explanations; Iowa corn.