In this paper, we present an unsupervised automatic target detection algorithm for multitemporal SAR images. The proposed two-fold method is expected to reduce processing time for large scene sizes with sparse targets while still improving detection performance. Firstly, pixel blocks are extracted from an initial change map to reduce the algorithm's search space and favor target detection. Secondly, an adaptive k-means algorithm selects the number of clusters that better separates targets from false alarms, which are discarded. Preliminary results show the advantages of the proposed method in processing time and detection performance over a recently proposed supervised method for the CARABAS-II dataset.