Abstract: In order to improve the robustness and imperceptibility of digital watermarking based on machine learning, considering the good learning ability and generalization ability of Support Vecor Machine(SVM) with limited training samples, and different relationship between the center pixel and neighborhood pixels in different areas, a new spatial digital image watermarking based on regional SVM model is proposed, which embed the watermarking based on Arnold transform. Due to the different degrees of releationship between the center pixel and neighborhood pixels, different kind of areas are built. SVM model is built in every kind of area, and then watermarking generated by Arnold transform is embedded and extracted. Experiments showed that the proposed scheme had good imperceptibility and better robustness against several attacks this algorithm attacks, such as cutting, pepper&salt, contrast enhancement and so on, than other watermarking algorithm based on machine learning.