Soil lead content is an important parameter in environmental and industrial applications. Chemical analysis, the most commonly method for studying soil samples, are costly, however application of soil spectroscopy presents a more viable alternative. The first step in the method is usually to extract some appropriate spectral features and then regression models are applied to these extracted features. The aim of this paper was to design an accurate and robust regression technique to estimate soil lead contents from laboratory observed spectra. Three appropriate spectral features were selected according to information from other research as well as the spectrum interpretation of field collected soil samples containing lead. These features were then applied to common Multiple Linear Regression (MLR), Partial Least Square Regression (PLSR) and Neural Network (NN) regression models. Results showed that although NN had adequate accuracy, it produced unstable results (i.e., variation of response in different runs). This problem was addressed with application of a Fuzzy Neural Network (FNN) with a least square training strategy. In addition to the stabilized and unique response, the capability of the proposed FNN was proved in terms of regression accuracy where a Ratio of Performance to Deviation (RPD) of 8.76 was achieved for test samples.