Local cloud and fog can cause low contrast and poor visibility in optical remote sensing images of certain ocean regions

which hinders ship surveillance. To overcome this

a multi-scale phase spectrum is used to reconstruct the low observable image to form a saliency map in the first step. Then

a global threshold is used to extract the regions of interest (ROI)

which has higher saliency. The order statistic of mean intensities from the sub regions of a circular window around each ROI is used to estimate the local threshold for target pixel segmentation. The mean saliency

shape complexity

and spatial extent are extracted from the target pixels to form a feature vector. Then a minimum distance classifier on the extracted feature vector is trained to discard the false alarms. Results on many cloudy SPOT-4 panchromatic images show the effectiveness of the proposed algorithm.