We propose an unsupervised method for slight change extraction and detection in multitemporal hyperspectral image sequence. To exploit the spectral signatures in hyperspectral images, autoregressive integrated moving average and fitting models are employed to create a prediction of single-band and multiband time series. Minimum mean absolute error index is then applied to obtain the preliminary change information image (PCII), which contains slight change information. After that, feature vectors are created for each pixel in the PCII using block processing and locally linear embedding. The final change detection (CD) mask is obtained by clustering the extracted feature vectors into changed and unchanged classes using k-means clustering algorithm with k=2. Experimental results demonstrate that the proposed method extracts the slight change information efficiently in the hyperspectral image sequence and outperforms the state-of-the-art CD methods quantitatively and qualitatively.