Rubber plantations are expanding rapidly in the tropical region of China because of their high profitability

while large areas of natural rainforests are deforested and replaced by artificial rubber trees. A series of ecological problems has occurred because of unreasonable and disordered plantation of rubber trees in some areas. Acquiring spatial distribution of rubber plantations rapidly and monitoring their dynamics can contribute to control and manage rubber plantations and to protect the local ecological environment. Remote sensing plays an important role in monitoring the dynamics of land use and cover. However

the high complexity of tropical ecosystems makes it difficult to extract information from a single-temporal image. This study aims to develop a method that extracts information about rubber plantations effectively and to improve classification accuracy of a tropical ecosystem with multi-temporal remote sensing images. Based on Moderate Resolution Imaging Spectroradiometer（ MODIS） Normalized Difference Vegetation Index（ NDVI） time series data and multi-temporal Landsat TM data

the seasonal characteristics and spectral characteristics of different objects

including rubber plantation

broadleaved evergreen forest

dry land

and paddy field in the study area

are analyzed and compared in this study. A decision tree model that integrates object-oriented technology is developed to extract rubber plantation in Hainan province. The model choose appropriate characteristic parameters and remote sensing data to separate rubber plantation and other land cover types. Rubber plantation demonstrates obvious seasonal characteristics according to time series data. During the dormant period of rubber trees

an obvious difference in spectral characteristics among different land covers is observed. NDVI and Land Surface Water Index（ LSWI） can be selected as the two main parameters to distinguish rubber plantation and other land covers in the study area. In addition

remote sensing data in winter are optimal for extracting rubber plantation. Rubber plantation and four other objects are extracted by using the five-level classification decision tree model based on LSWI

NDVI

MNDWI

and slope data. The total accuracy in the study area is 94. 5% according to the test result obtained with field observation data. Rubber plantation and broadleaved evergreen forest have a high classification accuracy

whereas dry land has a low classification accuracy. For the rubber plantation

the producer accuracy obtained with two-temporal TM images（ February and December） is 96. 2%

whereas the results obtained with a single-temporal image are 78. 1%（ December） and 86. 1%（ February）. Similarly

the user accuracy of rubber plantation obtained with two-temporal images is higher than that obtained with a single-temporal image. The presented decision tree model with multiple temporal images effectively integrates spectral and phonological information on vegetation

and some uncertainties of single temporal image are avoided. The use of this model ensures that rubber plantation can be extracted with high accuracy. The key and premise of this study are to identify the phonological and spectral characteristics of specific vegetation. Moreover

appropriate images that show specific characteristics of object must be available. Furthermore

if additional images in the key phonological phase of vegetation can be acquired

the classification accuracy can be improved further.