Abstract: SIFT descriptor is widely used for local feature extraction. However，some problems such as large numbers of extracted key points and its high dimension appear when using SIFT to extract local features from remote sensing imagery directly. To solve these problems and improve the retrieval results，we use a visual attention model to extract objects using their saliency from remote sensing images. The visual attention model is used to extract salient objects through their saliency from remote sensing images firstly，then we use a K-means algorithm to cluster local features，these results are then used as feature vectors for similarity measures. Some experimental results show that our method not only decreases the number of key points and the dimension of local features，but also improves retrieval results at the same time. It also accords with the human visual system.