Abstract: The training,verification and testing data set enough for neural network training and verifying was generated according to the soil quality classification criteria The optimal neural network topology parameters were determined by step-elimination or step-increase criteria The verification data set was used to keep the training process out of over-learning,thus the comprechensive model for evaluation and prediction of soil quality with more generalized and robust was established in this paper The changing trends of soil quality after large area reclamation in the Sanjiang Plain were quantitatively evaluated and shown that the soil quality of the top soils of the main cultivated soils were decreased at various speed with different kind of soil after large area reclamation Furthermore,case study shown that the neural network-based model was more suitable for the soil quality evaluation and prediction than weighted comprehensive index method

Key words: artificial neural network, soil quality, evaluation index, comprehensive evaluation and prediction