An automatic river target extraction method based on reciprocal cross entropy thresholding and improved graph cuts was proposed in order to overcome the arbitrariness of choosing seed points in interacting graph cuts method and the subsequent incorrect segmentation result. First, the river image was pre?segmented based on the re?ciprocal cross entropy threshold selection criterion and the seed points were automatically chosen from the pre?seg?mentation result. Next, the river remote sensing image was segmented by the improved graph cuts. In improved graph cuts, the seed point region was modeled by Gaussian mixture model and the smoothing terms were calculated by the structure tensor matrix. Finally, the small regions were eliminated using connected region detection method and the result was thus obtained. Comparing with the other four methods including interactive graph cuts and the method based on reciprocal cross entropy and graph cuts, the experimental results showed that the proposed method could obtain the best segmentation result.