Explainable ensembles and transformer‑centric learning in vision and beyond: Methods, validation and deployment assurance

Ensemble learning and transformer‑centric architectures now dominate vision and decision support, yet practitioners lack unified guidance on method selection, explainability and assurance. This title‑driven scoping review synthesises studies across healthcare, agriculture, cybersecurity, energy, finance and other sectors. Titles suggest a proliferation of ensemble techniques—from bagging, boosting and stacking to transformer‑based mixtures of experts—and increasing adoption of vision transformers (ViT, Swin, MaxViT) in imaging and multimodal tasks. The review infers a taxonomy of ensemble strategies, transformer backbones and fusion schemes, and surveys conceptual failure modes such as overfitting, spurious correlations and explanation instability. It characterises explainability families (saliency, concept‑based, example‑based, uncertainty‑based) and highlights unique challenges when interpreting ensembles and transformers. Domain‑specific synthesis across medical imaging, plant disease diagnostics, cybersecurity, energy forecasting and sentiment analysis reveals heterogeneous tasks and deployment contexts. A cross‑domain assurance checklist emphasises interpretability‑by‑design, robustness, privacy, monitoring and governance. A research agenda prioritises standardising explanation validation, agreement‑aware fusion, robustness calibration, privacy‑preserving explainable learning and multimodal XAI. This scoping synthesis maps current directions and outlines practical guidance, providing a foundation for evidence‑based updates once full texts become available.