This paper addresses the inference of spatial dependence in the context of a recently proposed framework. More specifically, the paper focuses on the estimation of model parameters for a class of generalized Gibbs random fields, i.e., Spartan Spatial Random Fields (SSRFs). The problem of parameter inference is based on the minimization of a distance metric. The latter involves a specifically designed distance between sample constraints (variance, generalized ``gradient'' and ``curvature'') and their ensemble counterparts. The general principles used in the construction of the metric are discussed and intuitively motivated. In order to enable calculation of the metric from sample data, estimators for generalized ``gradient'' and ``curvature'' constraints are constructed. These estimators, which are not restricted to SSRFs, are formulated using compactly supported kernel functions. An intuitive method for kernel bandwidth selection is proposed. It is proved that the estimators are asymptotically unbiased and consistent for differentiable random fields, under specified regularity conditions. For continuous but non-differentiable random fields, it is shown that the estimators are asymptotically consistent. The bias is calculated explicitly for different kernel functions. The performance of the sample constraint estimators and the SSRF inference process are investigated by means of numerical simulations.