River boundaries extraction from SAR imagery is valuable for flood monitoring and damage assessment. Several rivers, parts of which include dammed lakes caused by landslides and rock avalanches triggered by the 2008 Wenchuan Earthquake, were taken as a case study for robust extraction. In this paper, a novel state-of-the-art approach for automated river boundaries extraction using high resolution synthetic aperture radar (SAR) intensity imagery is presented. The key of our approach lies in the combined usage of local connectivity feature of the river and a region-based active contours model (ACM) in a variational level set framework to differentiate between river and the background. First, sub-patched intensity thresholding segmentation is applied to SAR imagery. Pixels with intensities below the threshold are selected as potential river pixels while the others are potential background pixels. Second, potential river pixels are divided into several connected regions, considering that the river is a big connected region, only relatively bigger regions with similar contrast value are retained as the regions of interest (ROI) while others are noise due to pixel-level decision approach in the first step or shadows due to mountains terrain. Third, the ROI and their contours are regarded as local region and the initial contours to refine the river boundaries, which are used to reduce the scene complexity of ACM and its sensitivity to initial situation, respectively. A novel ACM driven by local image fitting (LIF) energy is presented and used for river boundaries extraction for the first time, which is not only robust against inhomogeneity widely spread in SAR imagery but also can work with efficiency without the need of re-initialization during iteration compared to traditional ACM. The proposed approach was tested on numerous high resolution airborne SAR images containing connected rivers or dammed lakes obtained by Chinese domestic radar system after Wenchuan Earthquake. For the overall dataset, the average commission error, omission error and root mean squared error were 6.5%, 3.3%, and 0.51, respectively. The average computational time for 4000 by 4000 image size was 21 min using a PC-based MATLAB platform. Our experimental results demonstrate that the proposed approach is robust and effective.