Kernel-Adaptive TinyRBF Networks for On-Device Calibration of Soil Moisture Sensors

We propose a Tiny Machine Learning (TinyML) approach for continuous in-field calibration of soil moisture sensors based on the periodic availability of a reference measurement provided by a high-quality sensor. Soil moisture sensors are essential components in precision agriculture to guide optimized irrigation management. However, most commercial-grade devices are known to be inaccurate, unless complex and seldom applied calibration procedures are adopted. To address the calibration issue, we investigate the effectiveness of TinyRBF, a kernel-adaptive RBF framework designed for on-device calibration and regression under sparse reference supervision. Based on an extended field experiment and power profiling of sensor nodes that perform inference and learning tasks, we report the impact of different kernel alternatives and reference frequency on the accuracy and power consumption of the model for three different soil moisture sensors. The analysis aims to support soil moisture monitoring in large agricultural areas using multiple low-cost sensing nodes that can operate autonomously for the whole cropping season.