Atmospheric corrections were conducted with the MODTRAN4+ model for 49 TM data from 1990 to 2010 in Fuzhou. Multi-temporal trajectories of major land cover type were derived from NDVI images. The trends of Mean NDVI were analyzed. To investigate the influence of different data combination on the classification and detection accuraly of different methods,induding maximum likelihood classification, support vector machine, artificial neural network, and object-oriented methods, and compared the deteetion methods before and after adding a sample, the areas converted from cropland to built-up land were added to the learning sample. The results show that the object-oriented method is the most accurate method compared with other methods for a small sample. By using the method, the classification accuracy improves up to 3.53% and 4.24% for different data combination and different season respectively.Adding NDVI data and the sample of changing features improves 3.54% of the whole classification accuracy and 4.24% of drawing accuracy of the buid-up land.