Advances and Challenges in TinyML-Based Water Trace Element Monitoring

Machine learning (ML) deployments on microcontroller-class hardware, commonly referred to as TinyML, have emerged as a promising approach for trace element monitoring in environmental, agricultural, biomedical, and industrial applications. However, the extent of technological maturity, deployment feasibility, and real-world performance remains underexplored.This systematic review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. A comprehensive search of SCOPUS, Web of Science, and Google Scholar (2015–2025) identified 1,160 candidate articles. After removing duplicates and applying inclusion criteria focused on ML models deployed on microcontroller-class devices for trace element or environmental monitoring, 46 studies were included. Data were extracted on study type, application domain, ML framework, algorithm, hardware platform, dataset source, and reported constraints. The included studies comprised experimental (52.17%), applied research (28.26%), and case study (2.17%) designs. Application domains were dominated by water quality monitoring and prediction (26.09%), agriculture and smart farming (19.57%), and waste/environmental management (25.00%). TensorFlow (13.04%) and scikit-learn (6.52%) were the most frequently used ML frameworks. ESP32 (26.47%) and Arduino (23.53%) platforms were the predominant hardware choices, with XGBoost (33.33% of implementations) emerging as the most common algorithm. Reported classification accuracy ranged from 75–99.8% in laboratory settings; however, only 31% of studies included field validation. Memory limitations (< 100 KB RAM) were reported in 51.96% of cases, and power-related constraints in 27.45%. Sensor drift and environmental variability issues were noted in 68% of studies, while ultra-low-power optimisation was addressed in only 4.35%. TinyML-based trace element monitoring demonstrates high potential in controlled environments but faces persistent challenges in real-world deployment, including hardware memory constraints, environmental adaptation, and energy optimisation. Addressing these gaps—particularly through standardised model–hardware co-design, improved sensor robustness, and power-efficient architectures—will be essential for translating laboratory advances into scalable field solutions.