Tiny ML based crop recommendation system for precision agriculture 5.0

This article presents a crop recommendation system (CRS) for precision agriculture 5.0 using tiny machine learning (TinyML). An extensive statistical analysis is conducted to identify the most relevant features, and data cleaning is performed to ensure the effectiveness of the machine learning (ML) models while retaining essential information. This work conducts a first-level experimentation on a computer-based system for model selection and validation. The results of the first level demonstrate an accuracy of 94% for both logistic regression (LR) and decision tree classifier (DTC), 96% for support vector machine (SVM), and 99% for the random forest classifier (RFC). Therefore, in the second level RFC was selected to be incorporated in a resource-constraint microcontroller, ATMega328P to develop a cost effective, real-time, stand alone CRS. In doing so the model size is reduced to 98.78% by dint of hyper-parameter tuning maintaining an accuracy of 98.4%. Finally, the model is downsized to 99.26% for hardware implementation, occupying only about 28% of the available space. This makes the system ultra-light, highly efficient, and optimized, allowing it to produce crop recommendations instantaneously and locally without cloud support or internet connectivity.