Deep Temporal Ensemble for Millet Yield Prediction in Senegal: TimesNet, FEDformer, TimeMixer and MLP versus PCA--Logistic Regression Stacking with Blockchain-Certified Data

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 tab ular records, and trust concerns over data provenance. We present a tabular-only temporal deep learning pipeline that deploys three complementary neural base learners from the frequency domain and multi-resolution family, namely TimesNet, FEDformerLite, and TimeMixer, and compares two stacking meta-learners: a fully connected Multilayer Perceptron (MLP) and a dimensionality-reduction pipeline combining Principal Component Analysis (PCA) with Logis tic Regression (PCA–LogReg). A blockchain-inspired provenance layer certifies each zone–year transaction via SHA-256 hashing and HMAC signing. Experimental results on Senegalese mil let data (2000–2020, 14 regions) show that FEDformerLite, despite its theoretical limitation on short sequences, achieves the best individual performance (MAE = 0.080t/ha, R2 = 0.728). The tuned MLP stacking meta-learner further raises R2 to 0.930 (MAE = 0.088t/ha), a statis tically significant advantage over PCA–LogReg stacking confirmed by a paired Student t-test (p < 10−129). Dynamic zone-grouped cross-validation and comprehensive ablation, hyperpa rameter tuning, pruning, SHAP and gradient saliency analyses are also reported, confirming robustness and interpretability.