Deep Temporal Ensemble for Millet Yield Prediction in Senegal: TimesNet, FEDformerLite, TimeMixer and Optimized Hybrid Stacking with Leakage-Corrected Temporal Validation and External India Dataset Confirmation

Accurate millet yield prediction is essential for food-security planning in the Sahelian zone of West Africa, yet practical deployments face limited remote sensing coverage, heterogeneous tabular records, and the ever-present risk of data leakage in small-sample temporal settings. We present a tabular temporal deep learning pipeline that deploys three complementary neural base learners—TimesNet, FEDformerLite, and TimeMixer—combined with an optimized Hybrid Stacking meta-learner (six diverse base regressors plus a three-hidden-layer MLP) and compares it to a standard XGBoost flat-feature baseline. A post-hoc leakage audit revealed that five target-derived features (yield_kg_ha, yield_t_ha, anomaly, anomaly_pct, yield_ma3) were inadvertently included in the last step of each input window, inflating earlier results; the audit and correction procedure are documented in Appendix A. After fixing the leakage by zeroing all contaminated features at the prediction timestep, corrected results on the Senegal millet dataset (2000–2020, 14 regions) show that XGBoost operating on leakage-free flat tabular features achieves R2 = 0.821 (MAE = 0.126 t/ha), outperforming the corrected MLP stacking ensemble (R2 = 0.504). This reversal highlights that sequential temporal structure alone does not guarantee superiority over well-tuned gradient boosting when legitimate lag features already encode most of the temporal signal. To test the hybrid stacking architecture under genuinely non-linear agronomic relationships, we validate on an external India crop yield dataset (32 states, 1995–2021). There, the optimized Hybrid Stacking achieves R2 = 0.743 (MAE = 0.682 t/ha), surpassing XGBoost (R2 = 0.721, ∆R2 = +0.022), confirming that multi-learner stacking captures non-linear interactions that XGBoost alone misses. An optional, decoupled blockchain inspired provenance layer certifies each zone–year transaction via SHA-256 hashing and HMAC signing for transparency only (Appendix B).