Main content area

Estimation of spatially enhanced soil moisture combining remote sensing and artificial intelligence approaches

Moosavi, Vahid, Talebi, Ali, Mokhtari, Mohammad Hossein, Hadian, Mohammad Reza
International journal of remote sensing 2016 v.37 no.23 pp. 5605-5631
Landsat, algorithms, artificial intelligence, data collection, irrigation, models, moderate resolution imaging spectroradiometer, remote sensing, soil water, temperature, vegetation cover, vegetation index
The main objective of this study is to combine remote-sensing and artificial intelligence (AI) approaches to estimate surface soil moisture (SM) at 100 m spatial and daily temporal resolution. The two main variables used in the Triangle method, that is, land-surface temperature (LST) and vegetation cover, were downscaled and calculated at 100 m spatial resolution. LSTs were downscaled applying the Wavelet-Artificial Intelligence Fusion Approach (WAIFA) on Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat imageries. Vegetation fractions were also estimated at 100 m spatial resolution using linear spectral un-mixing and Wavelet–AI models. Vegetation indices (VIs) were replaced with the vegetation fractions obtained from sub-pixel classification in the T ₛ–VI triangle space. The downscaled data were then used for calculating the evaporative fraction (EF), temperature-vegetation-dryness index (TVDI), vegetation temperature condition index (VTCI), and temperature-vegetation index (TVX) at 100 m spatial resolution. Thereafter, surface SM modelling was performed using a combination of Particle Swarm Optimization with Adaptive Neuro Fuzzy Inference System (PSO-ANFIS) and Support Vector Regression (PSO-SVR) modelling approaches. Results showed that the best input data set to estimate SM includes EF, TVDI, T ₛ, F ᵥₑgₑₜₐₜᵢₒₙ, F ₛₒᵢₗ, temperature (T), precipitation at time t (P ₜ, P ₜ –₁, P ₜ –₂), and irrigation (I). It was also confirmed that PSO-SVR outperformed the PSO-ANFIS modelling approach and could estimate SM with a coefficient of determination (R ²) of 0.93 and a root mean square error (RMSE) of 1.29 at 100 spatial resolution. Range of error was limited between −2.64% and 2.8%. It was also shown that the method proposed by Tang et al., (2010) improved the final SM estimations.