Estimating soil moisture contents of farmland using UAV hyperspectral images of wheat canopy

Abstract: Accurate monitoring of soil moisture content (SMC) in agricultural fields can greatly contribute to the utilization of water resources and sustainable development. Low SMC can cause the soil to harden during crop growth, which in turn affects the absorption of the crops' water and nutrients. In this study, the field-scale wheat canopy spectra were gathered to quantitatively estimate and map the SMC using an unmanned aerial vehicle (UAV) hyperspectral sensor. The study area was selected as Fukang City, Xinjiang Uygur Autonomous Region, China (87°51'15′′E, 44°21'14′′N) at the transition zone between a desert and an oasis. The soil samples were also collected concurrently with the hyperspectral data. 70 sampling units of 0.5 m x 0.5 m were evenly selected in the target field, where the GPS locations were recorded. The wheat canopy spectra were smoothed using the savitzky-golay (SG). The spectral data were then transformed using the continuous wavelet transform (CWT) with seven different wavelet functions (bior4.4, coif4, db4, fk14, haar, rbio3.9, and sym 4). The wavelet coefficients were extracted by the genetic algorithm (GA), and finally combined with partial least squares (PLSR), Support Vector Machine (SVM), artificial neural network (ANN), radom forest (RF), and extreme gradient boosting (XGBoost) to predict the SMC. A neural Network was used to predict the SMC, and then map the SMC at a spatial scale. The results demonstrate that the accuracy of SMC estimation greatly increased using the GA-based feature band selection. The R2 of the feature band was 0.35-0.82 with the wavelet coefficients, while the R2 of the complete wavelet band ranged from 0.20 to 0.44. Furthermore, the R2 of SMC reached 0.82, 0.72, 0.79, 0.76, and 0.46, respectively, using PLSR, SVM, ANN, RF, and XGBoost models. The best estimation accuracy was achieved in the feature band with the db4 wavelet coefficients, compared with the rest wavelet functions. The predicted and measured values were remarkably similar. The GA significantly reduced the number of variables to maximize the feature factors in the remote sensing bands. The majority of the feature bands were retained to remove the redundant bands that frequently presented in the hyperspectral data. The more stable ANN model was achieved in the small number of sample points on the big sample, even though the accuracy was lower than that of PLSR. Consequently, the perfect estimation accuracy of SMC with excellent mapping was obtained in the continuous wavelet transform and GA feature extraction of the hyperspectral data acquired by the UAV. The improved inversion can be expected to achieve accurate SMC prediction using crop growth period spectra at the field scale. The finding can provide a strong reference for high-precision SMC monitoring using remote sensing.