With the significantly improved data availability in remote sensing technology

mid-resolution images have become the primary data source for crop sown area estimation in large scale.However

it is still difficult to solve the problems of spectrum heterogeneity in one field and spectra similarity between fields

especially in transitional region by using mid-resolution images.In order to maximally avoid above motioned problems and accurately measure the sown area of winter wheat

this paper developed per-field classification method and tested the method in an urban agriculture region with complex planting structure through several steps:first

digitalizing field boundary from QuickBird image;second

extracting characteristic index including spectrum and texture information as well as vegetation index for each field from the multi-temporal TM images;third

operating support vector machine（SVM） and maximum likelihood classification（MLC） with different field characteristic index;finally

estimating the accuracy of our method.Results show that the per-field classification method has a higher accuracy than per-pixel classification both in amount（estimated sown area of winter wheat divide by reference sown area of winter wheat

Kr） and position（equal to product accuracy

Kp）.Although both SVM and MLC could get very high amount and position accuracy（97% and 90% respectively）

the estimations of SVM are more stable.The errors of per-field classification mainly happened at the fragmentized parcels.Additionally

characteristic information could enhance the performance of per-field classification.Our method also has an outstanding advantage that no optimum period requires on satellite imagery which could enhance practicability and operationality of our method.