The Support Vector Data Description（SVDD） method has been recently applied to specific land-cover classifications.However

previous studies acquired pure pixels to construct the training set

which does not conform to the SVDD principle.In the present research

the study area was located in Tongzhou

Beijing.We first constructed several training sets with different spectral responses based on QuickBird image and ground truth data derived through visual interpretation.Then

two sampling schemes were adopted

selecting training samples in a given training sets and manual interpretation.Results showed that when the penalty parameter C increased or the kernel width s decreased

the volume of a hypersphere declined

thus resulting in lower producer accuracy and higher user accuracy.Based on the optimal parameters

the SVDD classifications were conducted using training samples.In the simulated experiments

when training data moved from the center to the margin of wheat distribution

the volume of a hypersphere increased

thus classification accuracy was improved.By contrast

as the outermost samples approach the boundary of wheat spectral distribution

a hyper sphere fitted by these cases can easily accept pixels of other classes.In the actual experiment

center training samples acquired according to previous literature yielded poor classification

with accuracies of 80.45% and 45.15%

from the overall and the producer perspective

respectively.However

classification using marginal training samples was more accurate

with 92.71% overall accuracy and 95.81% producer accuracy.The results demonstrated that using marginal training data could efficiently improve classification accuracy of the target class.