RadMoist: Physics-Guided Features and Hybrid Edge Inference for mmWave Soil-Moisture Sensing in IoT Systems

Abstract

Reliable and fine-grained knowledge of near-surface soil moisture is a foundation for precision irrigation, drought forecasting and hydrological modelling. However, existing probes, such as TDR and gravimetric sampling, remain labour-intensive and sparse. This study presents RadMoist, a fully contactless framework that estimates volumetric water content with a single low-cost 77 GHz FMCW radar (TI IWR1443BOOST, 4 GHz sweep and 3.8 cm range resolution). Beyond conventional surface amplitude and time-of-flight (ToF), RadMoist introduces three physics-guided indices: Microangular backscatter slope (μABS) for angular roughness, phase coherence decorrelation index (PCDI) for subsecond stability and diffusion gradient index (DGI) for high-band attenuation, yielding a five-element feature vector. These cues contribute to a custom-designed novel Hybrid-MoistureNet, an ensemble that averages a 200-tree XGBoost regressor with a compact three-layer MLP, enabling sub-millisecond inference on edge CPUs. The system was evaluated on loam, sandy-loam and clay soils spanning 0%–40% volumetric water content. RadMoist achieved a mean absolute error of 6.9% VWC, improving on an amplitude + ToF baseline (9.8%) by 29% and matching laboratory TDR measurements within 2%. Performance remained stable under ± 8°C temperature drift and ± 10° sensor tilt demonstrating robustness to field conditions. Therefore, RadMoist offers a practical calibration-light alternative for dense real-time soil-moisture mapping in precision agriculture and environmental monitoring.