F Taghavi; A Ahmadi; Z Zargaran
Volume 8, Issue 2 , November 2016, , Pages 53-72
Abstract
In this study, a combined model of modular networks and satellite image processing and optimization algorithms to forecast land surface temperature in an area including city of Tehran is presented. Calculating the LST has been done based on brightness temperature features in 31 and 32 MODIS channels. ...
Read More
In this study, a combined model of modular networks and satellite image processing and optimization algorithms to forecast land surface temperature in an area including city of Tehran is presented. Calculating the LST has been done based on brightness temperature features in 31 and 32 MODIS channels. Thus, brightness temperature data related to these images is fed to neural network and values of land surface temperature are recovered as the output of the network. In this way,after obtaining the optimal structure obtained for networks they are trained and their weights are extracted. Then by applying a neural network with a modular structure and clustering algorithms, training will be also modular. Decomposition of the networks and after that combining the results to get the final forecast makes the performance of the modular network more effective. As a result , a new approach based on the combination of neural network or self-organizing map and particle swarm optimization algorithms is proposed. The results showed that using PSO algorithm causes appropriate distribution of cluster of SOM method and using satellite images improved performance of the proposed model. Finally, results are compared with training neural network models and non-modular structure. The results of this comparison show that model-training time in predicting the land surface temperature is decreased and the accuracy of model increased. The little difference between the predicted values and actual (real) values of temperature in the region shows that this model could predict the temperature accuraetly, so that, in this hybrid model Mean Square Errors (MSE) and Mean Absolute Percentage Error (MAPE) are 0.0081 and 10.59 respectively.
