K-Means algorithm is the most commonly chosen technique for color image segmentation task. Although this algorithm is famous for its low complexity and easy implementation, but usually it is seen that the segmentation results are suffering from noises and over-segmentation, which mislead the final image analysis process. This is because of the inappropriate selection of the number of clusters in K-Means. Also, choosing a proper distance metric for this algorithm is very important as it impacts on the final segmentation results. This paper deals with these problems and presents a novel approach to solving the same. As per color image segmentation concerned, so choosing a suitable color space is the first mandatory condition. HSV color space is selected for the proposed research work. The input RGB image is first converted to HSV one. Here, the number of clusters is determined from the input image beforehand. For this, it is considered that number of regions in an image is equivalent to the number of clusters that can be formed through clustering the pixels of the image. The number of regions of an image is determined with Meyer’s Watershed algorithm with a proper preprocessing technique. Generally, mere applying Watershed algorithm results in over segmentation. So, we have introduced an improved Sobel filter based on multiple directional edge detection to deal with this problem. The V channel of the HSV converted image is filtered by the proposed improved Sobel filter first and then the filtered image is sent as input to the watershed algorithm. The watershed algorithm analyses the regions of the image through local minima calculations and the total number of regions hereby found is assigned to K. Then with the predetermined K, the pixels of the HSV converted image are clustered with K-Means Algorithm. “Cosine Distance Metric” is chosen for the distance based calculations involved in the K-Means algorithm. By properly labeling the different clusters, the final segmented image is obtained. The experimental results proved the better performance of the proposed approach in comparison to K-Means algorithm. Also, when applied to satellite color images, it is found that the proposed approach succeeds to form clear and distinct segments of the same and hence establishes a good framework for satellite color image segmentation.