Landsat time series clustering under modified Dynamic Time Warping
Yao Zhao, Lin Lei, Wei Lu, Yu Meng · 2016
Compared with the single remote sensing image, the time series images provide more information of ground objects, which can greatly improve the clustering accuracy. But time series clustering also has many difficulties, such as the impacts of cloud and other sharp noise. The sharp noise impacts time series clustering by affecting the calculation of the similarity measure of sequences. Therefore, this paper proposed a CD-DTW method (Canberra Distance-Dynamic Time Warping) based on the classic dynamic time warping distance and canberradistance. The CD-DTW method helps to construct a more reliable distance measure for remote sensing time series clustering. The CD-DTW is applied to Landsat time series clustering, and the overall performance of CD-DTW distance (overall accuracy 91.52%; kappa coefficient 0.89) was considerably better than that of classic DTW distance.