A Integrated Classifier for Large Regional-scale Land-cover Classifying

With a lot of successful applications of neural network-based classification, it has been recognized the classification can produce more accurate results than conventional approaches for remotely-sensed data. Although its training procedure is sensitive to the choice of initial network parameters and to over-fitting, the multilayer feed-forward networks trained by the back-propagation algorithm is adopted by many study cases. To overcome the shortages the genetic algorithm is involved, for one has the advantage during searching the infinitesimal point in local space and the other in global space. The integrated method can avoid the risk of premature convergence while BP network was training, moreover the end total mean square error of the network is steadier even if you redo the whole course several times, and turn the instability of classification result away. In the article, the algorithm mentioned above was used to work Chinese surface land cover classification at coarse spatial scales with input parameter dataset from MODIS data,at the same time two other algorithms: Fuzzy ARTMAP ANN and Maximum likelihood were involved for a comparison purpose. The result shows that the error back propagation neural network and genetic algorithm integrated method not only run with better efficiency, but also produce the best result than other two classifiers.