High precision photosynthetically active radiation estimation via a sensor-near AI architecture and low-cost sensors

In recent years, a crucial demand for more effi-cient natural resource exploitation has emerged, being one of the pivotal research fields in precision agriculture. Measuring Photosynthetically Active Radiation (PAR) assumes particular importance when evaluating the irrigation condition of plants and define the optimal water management policies. However, the current electronics available for PAR sensing are typically based on expensive and high-performance sensors that prevent large scale applicability. In this context, the present work aims at proposing a tightly embedded and low-cost alternative which can achieve performances similar to those of industrial references thanks to the integration of tiny Artificial Intelligence (AI) architectures capable to run on resource-constrained devices. More in detail, a prototype environmental sensor for PAR acquisition via inexpensive RGB sensors coupled with a dedicated polycarbonate diffuser is proposed and combined with a small multi-layer perceptron (MLP) model to predict high-quality PAR data. Results demonstrated that the solution can replicate PAR sensing with an accuracy 43 % better than standard regression methods, while offering significant energy autonomy (more than 430 days) when supplied by a small lithium battery.