Estimation of Near-Surface Soil Moisture Through GPR Signal Processing Based on Multi-Scaled Spectral Feature Weighting

The measurement of near-surface soil water content is very important for many applications. However, the current measurement technology for this task is not sufficient, especially for a large observation mission. Ground penetrating radar (GPR) become a promising non-invasive technique which can solve this issue. To address the building of an automatic measurement system which straightforwardly estimates the soil moisture content, this paper proposes a new signal processing technique of ultra-wideband impulse GPR signature based on multi-scaled spectral features which are weighted by neural network approach. In this method, the GPR response in the frequency domain is processed by multi-scaled triangular filter bank producing feature constant set. Using feature selection algorithm, twenty most significant features are selected to be multiplied by corresponding weight optimized by the neural network to directly estimate the soil moisture. The range of volumetric water content considered in this study is between 0.001 cm ³ / cm ³ to 0.300 cm ³ / cm ³ while the depth of moist soil is between 10 cm to 50 cm under the homogeneous dry-sand soil layer. According to the numerical simulation results, the proposed method provides a good agreement with the simulated water content with a variation of depth.