When classifying imagery of reality that seldom presents itself with hand boundaries but transitions and gradual interfaces between biophysical phenomena, remote sensing researchers have proposed a fuzzy partition matrix that is more representative of the real situation than the conventional hard classification. In achieving a complete fuzzy approach in supervised classification, this paper discusses the fuzzification of the training stage to improve the classification performance. “Fuzzy training” is proposed to cope with data uncertainty in the early stage of image classification in order to derive statistical parameters that more closely resemble reality. A fuzzy parameter estimator (FPE) was developed with a modified Bezdek's Fuzzyc-Means engine and empirically evaluated. Classification results based on the fuzzy spectral signatures were superior to results obtained using conventional training methods.