Abstract: With the continuous expansion of land development and construction scale, land use conditions are also changing year by year. Accurate forecasting of future land use development trends can provide a basis for regional land use planning and improve the efficiency of regional land use. Traditionally,the methods of CA_Markov, ANN, and CA_ANN models are usually used for prediction; however,there are problems such as long training time, poor prediction accuracy, and lack of persuasiveness. Aiming at the above problems, the authors established an adaptive variable filter network model in combination with the cellular automaton and neural network models, and created multiple data sets based on the number of land use categories within a certain area for training of multiple neural networks with different parameters. This model can predict the future land change situation, thus avoid the cancellation of network weights when training a single network. Compared with the best model CA_ANN out of the traditional ones, the overall accuracy of this model is improved by 1%-3%, the accuracy of land conversion is improved by 12.82%-33.33%, and the model predicting time is reduced by 49.47%.

Key words: remote sensing image, land use forecast, artificial neural network, cellular automata, adaptive variable filter