Hu TG (Hu Tan-gao), Pan YZ (Pan Yao-zhong), Zhang JS (Zhang Jin-shui), Li LL (Li Ling-ling), Li L (Li Le)

Beijing Normal Univ, State Key Lab Earth Surface Proc & Resource Ecol, Beijing 100875, Peoples R China

Abstract: This paper presents a new soft and hard classification. By analyzing the target objects in the image distribution, and calculating the adaptive threshold automatically, the image is divided into three regions: pure regions, non-target objects regions and mixed regions. For pure regions and non-target objects regions, hard classification method (support vector machine) is used to quickly extract classified results; For mixed regions, soft classification method (selective endmember for linear spectral mixture model) is used to extract the abundance of target objects. Finally, it generates an integrated soft and hard classification map. In order to evaluate the accuracy of this new method, it is compared with SVM and LSMM using ALOS image. The RMSE value of new method is 0. 203, and total accuracy is 95. 48%. Both overall accuracies and RMSE show that integration of hard and soft classification has a higher accuracy than single hard or soft classification. Experimental results prove that the new method can effectively solve the problem of mixed pixels, and can obviously improve image classification accuracy.

Keyword: Hard/soft classification; Linear spectral mixture model (LSMM); Support vector machines (SVM); Adaptive threshold