Toward sustainable hydroponic farming: An AI-driven IoT framework for virtual pH sensing and sensor lifespan extension

Soilless cultivation systems such as hydroponics rely on continuous and accurate monitoring of physicochemical parameters to maintain optimal plant growth conditions. In practice, however, hydroponic operations are constrained by sensor degradation, calibration drift, and delayed human intervention, which collectively reduce measurement reliability and disrupt timely nutrient regulation. Although existing IoT-based hydroponic systems enable real-time data acquisition, they largely depend on manual decision-making and offer limited support for sensor fault tolerance or predictive control. In this study, we propose an AI-driven IoT framework that integrates real-time sensing, automated environmental regulation, and data-driven prediction to enhance the robustness of hydroponic cultivation. Multivariate time-series data collected from a sensor-enabled hydroponic setup — including pH, nutrient concentration, and environmental parameters — are modeled using a gated recurrent unit (GRU)-based deep learning architecture optimized for temporal dependency learning. The GRU model functions as a virtual sensor, predicting critical parameter values and thereby reducing reliance on low-precision physical sensors and extending their effective operational lifespan. Experimental evaluations conducted under controlled cultivation conditions demonstrate that the proposed framework improves system stability and plant growth performance. The GRU-based virtual pH sensor exhibits high predictive performance, as evidenced by a coefficient of determination of R 2 = 0 . 9824 and low error metrics ( R M S E = 0 . 0181 , M A E = 0 . 0116 ). These results substantiate the efficacy of temporal deep learning approaches for accurate and reliable monitoring in hydroponic systems. Moreover, the proposed model exhibits low computational complexity, with an average inference latency of 0.049 s (49 ms), a compact architecture comprising 24,351 parameters, and a total model size of 95.12 KB. These characteristics make the framework well suited for deployment on resource-constrained edge devices, such as a Raspberry Pi, enabling real-time, on-device hydroponic monitoring without reliance on cloud-based computation.