Abstract: Abstract: There are inherent variations of color features in images of different natural field environments, which lead to a big challenge when areas of all kinds of green plants need to be extracted from soil background according to the color information in some applications, such as weeds recognition and agricultural robot navigation. So an adaptive image segmentation method was developed. Firstly, the initial segmentation model was set up based on support vector machine (SVM) with the labeled training samples. Then the unlabeled training samples were obtained automatically from the images to be segmented based on the K-means clustering. Both the unlabeled and labeled samples constituted the set of hybrid training samples. Secondly, the transductive support vector machine (TSVM) was trained based on these hybrid training samples. The unlabeled samples were labeled by the TSVM, and the distribution information of color features in the images which provided unlabeled samples was mined out. Finally, the new set of labeled training samples was obtained by discarding some inappropriate ones in the hybrid training samples according to their distance to the optimal classification plane. The initial segmentation model was updated with new labeled training samples, and in this way the updated model was more pertinent to the particular field images for segmentation. The experimental results showed that the pertinence between the segmentation model and the specific natural field image had been improved, and the adaptability of segmentation processing had been enhanced significantly.