Image segmentation is subjective in nature. It will depend on what you want to observe in the image. If we talk about your given image, it seems that you may want to separate the brownish region from the bluish region. However, if your region of interest is the brownish region and you want to separate the blemishes in that region alone, the segmentation criterion will change.

Thank's dear Sanjiv

If we consider my ROI is the land(brownish region), how can I compare different algorithms effect on segmentation result? (statistically)

I mean what can I do for comparing the first image with this image?

If both the images are the same scale, you can do an image subtraction operation and take the average difference over the regions. It may be better to first segment the images. Assuming that the segments of interest are white (all 1s in the pixel representation) you can perform an AND operation on original image and segmented image to extract the original pixel values corresponding to the segment. Other than that, I'll need to know your application to give you specific ideas.

Segmentation of Multi-Spectral Remote Sensing Images Using Neural Network

There are several measures for the precision of segmentation. Direct subtraction might be "too precise". you should look at F-measure with error margin. It allows to determine which segmentation is closest to the target

You say target is a user made image(matrix) and measure is just comparing segmented image with target, right?

Usually in segmentation, both accuracy and variability (precision) are evaluated. For example, in medical imaging, the real truth boundary is not practical to obtain for algorithm evaluation, so the algorithm result is compared to the manual segmentation, which is a surrogate of ground truth.

Accuracy can be evaluated using Dice coefficient, which measures the overlap.

If you want to evaluate the precision, the algorithm segmentations need to be repeated multiple times and obtain measures such as coefficient of evaluation or ICC. Whether the obtained precision is good enough, depends on your application. For example, you can obtain statistical measurements such as minimum detectable difference, if you are measuring the change.

yes, the target is made by human - this implies a training set and such approach is well suited for Machine Learning

Thank you.

I made the target using region growing because of high resolution of satellite images and I used it for neural network training but I missed that it's usable for segmentation quality checking too.

I used MSE(Mean Square Error) to measure error between target and segmented image and it was a good criteria. I will try other measurments such as F-measure (thanks to Martin) and so on.

Thanks for all supports.

You may try Modified Hausdorff Distance, Dice Similarity Coefficient, Sensitivity and Specificity performance metrics for image segmentation