A Multiple Level Set Model for Multispectral Image Unsupervised Classification

Satellite imagery especially with high spatial resolution often shows spectral variations and details disturbances in a class. These characteristics bring difficulties to people who are working at automatic classification in the remote sensing fields. To seek more effective method, this paper presents a new multiple level set model to implement unsupervised classification for multispectral images. Firstly, medium filtering technique oriented from image processing is introduced into a traditional level set model to improve the performance of classification. Then, to alleviate classification errors caused mainly by spectral in homogeneity, a novel class constraint energy term is constructed. By reducing energy among similar classes and punishing those pixels with wrong class label, the class constraint term can effectively improve classification result from basic model. Comparative experiments on real data have demonstrated effectiveness and robustness of our proposed model.