Remote sensing image classification based on improved watershed segmentation and Fuzzy Support vector machine

Traditional classification methods only based on spectrum features of pixels are not suitable for high-resolution remote sensing image. In this paper, we proposed a new object-oriented classification method. Our work included three aspects: improved image segmentation, features selection and improved Fuzzy Support Machine classifying combined with ISODATA. We made some improvements in these aspects respectively. Firstly, we use morphological reconstruction filter with a suitable scale to alleviate the conflict between noise reduction and boundary protection. Secondly, in order to extract markers precisely, we designed a new index, which we called gradient flatness index, and proposed a new method of marker extraction based on it. Thirdly, considering that land covers are very complex, we used spectrum features, texture features, and fractal dimension as classification features. Fourthly, in order to obtain high-quality training samples, we proposed an improved FSVM classifying algorithm combined with ISODATA. In order to reduce the impact of non-critical samples, we designed a new algorithm to compute fuzzy memberships of samples taking into account sample scale, distance and position synthetically. From the experiments, our improvements can not only improve the classification results, but also make objects classified automatically.