Reading Water Quality from the Haze: Feed-Forward Underwater Reconstruction with a Spatially Varying Medium

Cameras on underwater robots see the world through a murky medium. Water absorbs light, and suspended particles scatter it back toward the lens as a veil of haze. Every underwater 3D reconstruction method treats that haze as a nuisance: it estimates the haze and removes it so the seafloor geometry comes out clean. In eutrophic lakes, though, the haze is the thing ecologists want to measure. Eutrophic lakes are waters overloaded with nitrogen and phosphorus, where algal blooms and organic particles grow. The same algae and particles that blur the image are the signature of the degradation. We propose a reconstruction model that runs in a single forward pass, in seconds, with no per-scene training. It outputs two things from survey video: clean 3D geometry, and a coarse 3D map of how strongly the water attenuates and scatters light at each location. We then test whether that map, calibrated against onboard sensors, tracks turbidity and chlorophyll-a (the standard pigment proxy for algal biomass) away from the vehicle's own path.

The opportunity exists because two lines of work stop just short of it. Physics-aware underwater splatting methods already separate scene from medium, and one of them models the medium as spatially varying. All of them need minutes of optimization per scene. The one feed-forward underwater model assumes the water is roughly uniform and was trained on clear marine scenes. Shallow hypereutrophic lakes break both assumptions: the water is far murkier, and the murk is patchy.

Core Claim

A feed-forward Gaussian-splatting reconstructor can be trained purely on synthetic data where the water's optical properties vary in space. Gaussian splatting represents a scene as many small colored 3D blobs that are rendered by projecting them onto the image. In the synthetic training data, the variation in water properties is generated from a small set of water-constituent parameters: algal pigment, dissolved organic matter, and suspended sediment, each arranged in plumes and gradients. Trained this way, the reconstructor should do two things. First, it should keep reconstruction quality in turbid, patchy water where uniform-medium models degrade. Second, it should recover a medium field whose structure correlates with in-situ turbidity and chlorophyll measurements. The new element is not speed, and not physics-aware rendering, since both exist. The new element is a regressed, spatially varying medium that serves as a scientific output. The design choice we must defend is that varying the medium in space, rather than simply randomizing a uniform medium over a wider range, is what produces both gains.

Relation to Prior Work

NemoSplat is the closest work and our starting point. It uses a streaming transformer to predict camera poses, depth, clean "vacuum" Gaussians, and global attenuation and backscatter parameters from uncalibrated video in under ten seconds. It trains on Unreal Engine 5 sequences plus 256 real sequences. Its medium is assumed roughly uniform. It evaluates only on rendering quality, and its authors name GPU memory as their main bottleneck. We change three things: the medium output becomes a low-resolution 3D field, the synthetic medium follows a constituent-based generator with plumes, and the evaluation adds ecological validity.

The per-scene optimization methods supply the rendering physics but motivate the speed requirement. SeaThru-NeRF splits the scene into separate object and medium networks but trains for about ten hours per scene. WaterSplatting and Aquatic-GS replace the object network with Gaussians and keep a small network for the medium. They reach real-time rendering but still need minutes of fitting per scene. UW-3DGS is the most relevant constraint: it already fits a spatially varying medium on a low-rank voxel grid, per scene. Our medium head regresses the same kind of low-rank grid in a single pass, which keeps memory small relative to NemoSplat's per-pixel attention.

R-Splatting supplies counter-evidence we must take seriously. It argues that idealized image formation models break under caustics and organic matter. It shows that naive 2D restoration before 3D fitting causes flicker and floaters, the spurious blobs that float in empty space in a bad reconstruction. Clear-air feed-forward models, GS-LRM, DepthSplat, and ReSplat, supply the architecture and the error-feedback refinement idea. All assume vacuum light transport. In ReSplat's case, the mismatch would push haze errors into the geometry.

Two sim-to-real papers supply competing recipes for crossing from synthetic to real data. Adversarially Masking Synthetic to Mimic Real learns to degrade synthetic data adversarially, a model-free route. Synthetic vs. Real Training Data for Visual Navigation shows that a frozen pretrained encoder lets a simulation-trained policy transfer, reaching 90% real success. Ecology motivates the target regime. Lake Taihu chlorophyll-a spans 0.48 to 347.85 mg/m³ across stations, and surface biomass decouples from local nutrients because of stratification (Lake Taihu). Satellites lose more than half of observations to cloud and see only the surface (STIMP).

Research Program

Synthetic data. We extend a UE5-style pipeline with shallow lakebed scenes containing submerged vegetation, sediment, and debris. We replace uniform water with a 3D constituent field. Smooth random plumes of pigment, dissolved organic matter, and sediment are mapped through a fixed per-channel absorption and scattering model to attenuation and backscatter at each point. Light is then integrated along each ray. Training ranges extend into hypereutrophic murk, so distant surfaces fully vanish into the haze.

Model. The model is NemoSplat-style: a streaming geometry backbone predicts poses, depth, and clean Gaussians. A new medium head regresses a low-rank 3D grid of attenuation and backscatter in a world-aligned volume around the camera trajectory. Rendering composites Gaussians and medium along each ray, as in WaterSplatting. Training uses synthetic ground truth for geometry and medium, plus a self-supervised re-rendering loss on unlabeled real lake video.

Sim-to-real crossing. We compare three recipes on otherwise identical training runs: the physics-based constituent generator, ASM-style adversarial degradation of clean synthetic frames, and both combined. This directly tests the open tension between explicit physics and model-free degradation.

Field data. We collect a held-out set of shallow eutrophic-lake transects. Each run pairs camera video with a co-mounted fluorometer and turbidity sensor. Each site adds a sparse grid of diver or second-vehicle profiles that lie off the camera vehicle's path.

Baselines

Clear-air reference. DepthSplat applied directly to lake video is the naive reference. It shows how badly clear-air feed-forward reconstruction fails here. It runs with released weights, using the poses our backbone provides so pose failure is not confounded with rendering failure.

Strongest feed-forward competitor. NemoSplat is the strongest feed-forward competitor, because it is the only method that is fast, aware of the water, and trained across simulation and real data. On its own synthetic UE5 benchmark it reports 23.98 dB PSNR (peak signal-to-noise ratio), 0.47 dB above Droid-W. On 14 real degraded sequences it reports a 36.6% relative LPIPS reduction over WildGS-SLAM. LPIPS is a learned perceptual distance where lower is better. For fairness, we fine-tune NemoSplat on our synthetic lake data and the same unlabeled real lake video, with the same GPU-hour budget and hyperparameter search as our model. Beating it means two things on held-out turbid lake scenes. Our model must have higher PSNR and lower LPIPS. Its medium field must also correlate better with off-path sensor readings, since NemoSplat's single global medium value can only be compared as a scene average.

Strongest per-scene quality reference. UW-3DGS fits a spatially varying medium per scene and represents best achievable quality. WaterSplatting is a second per-scene reference; it reports 32.67 PSNR on the Curaçao scene of the SeaThru-NeRF dataset and 24.45 PSNR on BlueCoral3D, where R-Splatting reports 26.31. Both get their default per-scene schedules. We do not expect to beat them on PSNR. Success means coming within a stated margin at roughly 100× lower latency, and matching UW-3DGS's medium-field correlation with sensors.

Ablation of the key ingredient. This ablation is our own model with the medium head reduced to one global attenuation and backscatter value. It is trained on the same synthetic scenes, with uniform-medium randomization over the full turbidity range. It isolates whether spatial variation, rather than wider training ranges, drives the result.

No reported result was found on Lacuna for any method on turbid inland or eutrophic-lake reconstruction, or for medium-to-chlorophyll correlation, so these comparisons are new measurements.

Why not just…?

Why not just fine-tune NemoSplat on turbid lake video? This is the most direct threat. If NemoSplat's global medium, fine-tuned on our data, matches our rendering quality, then the spatially varying head has to earn its keep through the medium map alone. The fine-tuned-NemoSplat baseline and the global-medium ablation settle this separately for image quality and for ecological signal. We will report the result even if only the ecological signal favors our model.

Why not just randomize a uniform medium over a wider range? This is the cheaper version of our idea. Physically, a single attenuation value per scene cannot explain a plume that is clear on one side of the frame and opaque on the other, so its errors should appear as depth-dependent floaters. The global-medium ablation tests this directly. If the ablation matches on patchy scenes, the core claim fails.

Why not just run UW-3DGS offline? For one transect it is fine and probably higher quality. The argument against it is throughput over hours of survey footage and use onboard the vehicle. We will report the quality gap honestly rather than claim parity.

Why not just interpolate the fluorometer track? The onboard sensor already measures water quality along the vehicle's path. The image-derived map is only useful if it predicts readings off that path better than spatial interpolation of the track does. That is our decisive ecological check.

Why not restore frames in 2D first? R-Splatting shows that per-frame restoration introduces multi-view inconsistency and floaters in subsequent splatting. We include a restore-then-DepthSplat run only as a sanity check.

Proposed Experiments

Synthetic. On held-out synthetic lake scenes with known geometry and medium, we measure depth error, medium-field error, and novel-view PSNR, SSIM (structural similarity), and LPIPS. Results are binned by turbidity level and by how patchy the medium is. The prediction is that our model's margin over the global-medium ablation grows with patchiness.

Marine generalization. On the SeaThru-NeRF dataset, BlueCoral3D, and the NemoSplat benchmark, we check that the lake-oriented model does not lose ground on clear marine scenes.

Field. On our real lake transects, we measure rendering quality, latency, and peak GPU memory. We then run the ecological test: rank correlation between the medium field sampled at off-path profile locations and measured turbidity and chlorophyll-a, compared against interpolation from the onboard track. The sim-to-real recipe comparison (physics generator, adversarial degradation, both) is run on this field set.

What counts as success. The idea succeeds if three conditions hold. The spatially varying model beats both fine-tuned NemoSplat and the global-medium ablation on patchy turbid scenes. It stays within reach of UW-3DGS at seconds per scene. Its off-path medium estimates beat track interpolation. If only the first two hold, the contribution is a faster reconstructor rather than a water-quality instrument, and we will report it as such.