Coupling an Inverse Gaussian Model with Artificial Neural Networks to Predict Soil Moisture from Hyperspectral Imagery

Abstract:

Soil moisture is one of the most crucial properties for monitoring and modeling landscape processes. For this study hyperspectral imagery and soil physical properties were collected in both in situ and controlled laboratory experiments to establish predictive capabilities for soil moisture in saline soils. An inverse Gaussian model was first applied to fit the spectral reflectance curves and to derive three curve-specific parameters, namely the inverted amplitude, the distance from the center to the inflection point, and the area under the Gaussian curve. Then both linear regression analysis and artificial neural networks (ANN) were applied to develop soil moisture prediction models. Results indicate that soil salinity greatly affects surface reflectance and thereby prediction of soil moisture. The linear regression model failed to predict soil moisture for all in situ field samples as well as for controlled laboratory samples with moderate salinity levels. It was only able to predict moisture reasonably well when salinity levels were extremely high. Application of ANNs significantly improved prediction accuracy as evidenced by a substantial increase of the correlation coefficient and Nash – Sutcliffe efficiency. Based on obtained results, the coupling of an inverse Gaussian model with artificial neural networks provides practical and accurate means for prediction of soil moisture of saline soils and shows great potential for large-scale soil moisture mapping based on hyperspectral imagery.