Underwater, the ocean doesn’t just surround you — it deceives you. Light bends, scatters, and loses its color long before it reaches a camera lens, wrapping everything in a blue-green haze that obscures as much as it reveals.
A coral reef that looks like a uniform, faded blur from a few meters away could be bleached and dying — or it could be thriving. From a distance, through water, it’s nearly impossible to tell. For scientists trying to document and protect marine ecosystems, that perceptual gap has long been one of the most stubborn problems in ocean research.
When water becomes the enemy of vision
Light does not travel through water the way it travels through air. It bends, scatters off suspended particles, and loses entire wavelengths before reaching a camera — leaving a distorted, color-drained image that misrepresents nearly everything it captures.
Two optical phenomena account for most of this distortion. Backscatter occurs when light bounces off tiny particles suspended in the water, casting a haze across the image. Attenuation, meanwhile, is the gradual fading of certain wavelengths with distance — red objects lose their color far faster than blue ones. Together, these effects make underwater images unreliable as scientific records.
Researchers have attempted fixes before. An algorithm called Sea-Thru can accurately restore true colors in underwater images, but it demands significant computational power, making it impractical for real-time use or for building 3D models of ocean environments.
Three-dimensional Gaussian Splatting, or 3DGS, offered another promising path — above water, it stitches images into seamless, explorable 3D scenes. Applied to underwater footage, however, backscatter and attenuation warp colors so inconsistently across angles and distances that the method cannot produce a coherent result.
How SeaSplat works: stripping the ocean away, pixel by pixel
At the core of SeaSplat is a color-correcting algorithm that addresses backscatter and attenuation at the same time. For every pixel in an underwater image, it calculates the precise degree of optical distortion introduced by the water, then mathematically removes those effects to recover the object’s true color.
MIT graduate student Daniel Yang embedded that algorithm directly into a 3DGS model. The result is a system that takes underwater images as input and generates a true-color, fully explorable 3D scene as output — quickly, without sacrificing accuracy.
What makes the finished model genuinely useful is how it handles color. In real water, color shifts depending on viewing distance and angle. Inside a SeaSplat model, that no longer applies — objects retain their true color at every distance and perspective. “Once it generates a 3D model, a scientist can just ‘swim’ through the model as though they are scuba-diving, and look at things in high detail, with real color,” Yang says.
Tested across oceans: from the Red Sea to the Caribbean
To validate SeaSplat, the team tested it across a wide range of environments. They drew on a pre-existing dataset of images captured in the Red Sea, off the Caribbean coast of Curaçao, and in the Pacific Ocean near Panama — each location representing different water conditions and lighting.
The system was also tested on live footage captured by a remote-controlled underwater robot operating in the U.S. Virgin Islands. Across all locations, SeaSplat produced 3D models with more vivid, varied, and accurate colors than previous methods had delivered.
One practical limitation remains. SeaSplat currently requires the computing power of a desktop machine — too large to fit aboard an autonomous underwater robot. For now, it is best suited to tethered vehicle operations, where a robot connected to a ship sends images to an onboard computer for processing.
A new lens for coral reefs and marine biodiversity
The most immediate application is coral reef monitoring. As an underwater robot moves through a reef and captures images, SeaSplat can process that footage and build a true-color 3D model in near real time. Scientists can then navigate the model at their own pace, inspecting features that would be invisible or ambiguous in raw footage.
Yogesh Girdhar, an associate scientist at the Woods Hole Oceanographic Institution, points to coral bleaching as a direct example. Bleaching appears white up close, but from a distance through water it looks blue and hazy — easy to miss or misread entirely. SeaSplat could close that gap.
Beyond bleaching, the technology may help researchers quantify biodiversity and assess the overall health of marine communities with a level of visual accuracy not previously achievable. The paper will be presented at the IEEE International Conference on Robotics and Automation (ICRA), with support from the Investment in Science Fund at WHOI and the U.S. National Science Foundation. As computing hardware continues to shrink, the gap between SeaSplat’s current tethered setup and fully autonomous underwater deployment may not hold for long.