In this paper, we propose to estimate soil moisture in bare soils directly from hyperspectral imagery using support vector regression (nu-SVR). nu-SVR is a supervised non-parametric learning technique, e.g. making no assumption on the underlying data distribution, which shows good generalization properties even when only a limited number of training samples is available (which is often the case in soil moisture estimation). Estimation in six tilled bare soil fields shows the potential of using non-linear nu-SVR for the prediction of gravimetric soil moisture. Dependence to the origin of training samples, as well as their number, is thoroughly considered.