From Machine Learning to Transformers in Temperature and Moisture Sensing

This manuscript systematically reviews artificial intelligence (AI) methodologies for temperature and moisture sensing systems from 2010 to 2025. PRISMA-guided analysis of 76 papers delineates three phases: traditional machine learning (2010–2014), deep learning integration (2015–2019), and advanced neural architectures (2020–2025). Quantitative results show temperature prediction RMSE falling from 1.82–2.64 °C to 0.38–0.82 °C and soil moisture RMSE improving from 0.042–0.054 m3/m3 to 0.009–0.024 m3/m3. Contemporary innovations include transformer architectures, attention mechanisms, physics-informed neural networks, and hybrid models that enable multi-scale feature extraction, real-time processing, and uncertainty quantification. Applications demonstrate efficacy across precision agriculture, climatological modeling, and building management systems. Persistent challenges include data quality assurance, computational optimization, and model interpretability requirements. Future research targets explainable AI, edge computing, and physics-constrained paradigms.