Research On Regional Land Cover Classification In China Using MODIS Data

Land cover types recognition plays a chief and basic role in land cover/use research. By remote sensing technology, acquiring more detailed, exact and real-time land cover fact in regional and global scale area has becoming a hotspot in recent global change research. Internationally, the research on acquisition of large area land cover in China is few . This paper, with the support of MODIS data, carries out a relative systemic study on Chinese land cover classification in 21-century using advanced remote sensing theory and technology and methods.From this research, the following conclusions have been reached:1, Classification characters selection and extraction based on MODIS can improve regional land cover classification accuracy. And acquisition of geographic background studying and professional knowledge helps to valid classification character selection.Addition of the characters such as vegetation index, water and humidity, texture and temperature extracted from MODIS data can increase classification accuracy. However, addition of classification characters can't always enhance some types classification accuracy.2, Selection of good classifiers and advanced technology help to increase accuracy. In our experiment, Parzen window which is based on statistical theory performs best in six classifiers and RBF neural net also has good performance. And CART decision tree, BP neural net and C4.5 perform better than MLC. However, Fuzzy ARTMAP known as good behaviour in the past has unexpected bad accuracy in comparison with MLC. The result also shows that adding boosting technique introduced in machine learning area to decision tree can evidently increase classification accuracy for the poorly separable classes in MLC.3% Parzen window can truly approach the class probability density to solve complicated problems under various sampling schemes. And CART decision tree has better flexibility and robustness, however, it pursues high accuracy at the cost of sample size. RBF and BP neural network has high accuracy but requiring high-quality samples and it is hard to define its net structure parameters. Fuzzy ARTMAP neural network can't perform what is expected for some possible reasons.4, How to define training sample size and therefore select classifiers is a problem to solve in actual classification considering the cost of acquisition of samples. The experiment result shows that classification effect caused by the size of training...