Estimation of Soil Water Content Based on Hyperspectral Features and the ANN Model

Soil water content (θ) is an important factor for the crop growth and crop production. The objectives of this study were to (i) test various regression models for estimating θ based on spectral feature parameters, and (ii) compare the performance of the proposed models by using artificial neural networks (ANN) and spectral feature parameters. The θ data of sand and loam and concurrent spectral parameters were acquired at the laboratory experiment in 2014. The results showed that: (1) the maximum reflectance with 900～970 nm and the sum reflectance within 900～970 nm estimate θ had the significant, when sand bulk density was 1.40 g·cm-3; the maximum reflectance with blue edge and the sum reflectance within 900～970 nm had the best correlation (R2>0.70) when sand bulk density was 1.50 g·cm-3; while soil bulk density was 1.60 g·cm-3, the sum reflectance within 780～970 nm and normalized absorption depth in 560～760 nm reached a significant (R2>0.90); when soil bulk density was 1.70 g·cm-3, the maximum reflectance with 900～970 nm and the sum reflectance within 900～970 nm had the best correlation estimate θ (R2>0.88). 2) When the soil type was loam, the maximum reflectance with 900～970 nm and the sum reflectance within 900～970 nm had a best correlation estimate θ. The spectral feature parameters the sum reflectance within blue edge (R2=0.26 and RMSE=0.09 m3·m-3) and 780～970 nm absorption depth (R2=0.32 and RMSE=0.10 m3·m-3) were best correlated with θ in the sand. The θ model based on maximum reflectance with 900～970 nm (R2=0.92 and RMSE=0.05 m3·m-3) and the sum reflectance within 900～970 nm had a high correlation (R2=0.92 and RMSE=0.04 m3·m-3) in the loam. The BP-ANN model presented a better estimation accuracy of θ (R2=0.87 and RMSE=0.05 m3·m-3) in two soils. Thus, the ANN model has great potential for estimating θ. Thus, the BP-ANN model has great potential for θ estimation.