It seems like you could take advantage of some background subtraction. Detecting the horizon seems trivial enough (hough transform? or a robust edge detector).

It seems like you could take advantage of some background subtraction. Detecting the horizon seems trivial enough (hough transform? or a robust edge detector). If you know where that line is then you can take series of gradients normal to it and form an average temperature model as a function of distance from the horizon. If you take enough of these gradients across the image, you should be able to get a robust average. You then use this model to predict the temperature at each pixel and take a difference with the observed data. What should happen is you get a ton of ~zeros, a bit of noise due to waves and any non-sea object will stick out like a sore thumb. The rule here is that absolute measurements will kill you when it comes to dynamic range. You're much better off taking relative measurements.