Extreme learning machine (ELM) has been applied in near infrared spectral analysis as a novel chemometric method which attracted the attentions of various researchers. However, the dimension of spectral data is usually very high while more hidden nodes should be incorporated in original ELM model for spectral data. Thus the problems of high dimension and high colinearity in the output matrix of hidden layer of ELM model are inevitable. The solutions obtained with the existing Moore-Penrose generalized inverse can be ill-conditional due to the high dimension and high colinearity in the hidden layer output matrix. This study aims to propose an improved ELM to build spectral regression model. The proposed method firstly uses extreme learning machine (ELM) to relate spectral variables to response variable; then the output of each hidden node are treated as new variables; VIP-SPLS ( improved stacked PLS based on variable importance in the projection) proposed by our group recently is used to build the regression model between those new variables and the response variable. In this paper, this method is called as improved ELM (iELM). VIP-SPLS model can fully utilize the output information of each hidden node and can effectively solve the problems of high dimension and high colineariy. At the same time, VIP-SPLS also has the advantage of model ensemble. Therefore, the performance of ELM model used for spectral data can be improved if the VIP-SPLS is incorporated to relate the hidden layer output matrix and response variable. The proposed method is applied to a commonly used benchmark NIR spectral data for evaluation. The results demonstrate that the precision improvement of iELM model is 29.06% to PLS model and 27.47% to original ELM model, respectively.