How Many Grab Samples Is a Simulated Spectrum Worth? Closed-Form Kernel Retrieval of Water Quality from Synthetic Pre-Fits

Water quality from the air is inferred by mapping a reflectance spectrum, the colour of the water across many wavelengths, to a concentration such as chlorophyll-a or suspended sediment. The learning problem is starved of labels. Each training example needs a cloud-free pixel matched to a lab assay taken at the same place and time. Published inland studies work with roughly 5 to 45 such match-ups per water body. Physics offers a way around this, because bio-optical forward models can simulate as many spectra as we like from assumed mixtures of algae, sediment and dissolved organic matter. Nobody has measured what those simulated spectra are actually worth in real labels, or which properties of a simulator make them worth more.

We propose a retrieval model built so that this question can be answered cleanly, and so that the model runs at sensor line rate. The non-linear part of the model, an explicit kernel feature map, is fixed using simulated spectra only. Real match-ups touch nothing but a linear output layer. That layer is solved in closed form and pulled toward the solution the simulator implies. With this design, every real sample's contribution can be counted exactly, a new lake can be calibrated from five points without gradient descent, and predicting each pixel costs one small matrix-vector product. The scientific output is a synthetic-to-real exchange rate for continuous environmental regression, and a test of which simulator ingredients drive it.

Core Claim

Three things push toward this design. Non-linear kernel retrieval beats empirical band formulas in optically complex Case-2 waters, where algae, sediment and dissolved organic matter vary independently. PSO-LSSVM fits a support-vector-style regressor with a Gaussian kernel to UAV hyperspectral data and beats exponential, logarithmic and power fits as well as Random Forest. However, it must search over kernel bandwidth and regularization with particle swarm optimization on the same 23 match-ups it learns from. Separately, adding synthetic data at scale breaks exact kernel methods, whose cost grows cubically with sample count (UniCon makes this point and turns to Nyström or random-feature approximations). Finally, how much synthetic data is worth has only been measured for vision tracking. There, On the Equivalency, Substitutability, and Flexibility of Synthetic Data finds about 30 synthetic frames are worth one real frame, and that 60–80% of real data can be replaced.

We claim that a single choice resolves all three. That choice is an explicit, finite feature map $\phi$ whose landmarks and bandwidth are set on simulated spectra. On top of it sits a linear head $w$. The head is fit to the $n$ real match-ups $(x_i, y_i)$ by ridge regression centred on the head $w_{\text{syn}}$ that was fit to the simulator output:

$$ w = \arg\min_{w} \sum_{i=1}^{n} \big(y_i - w^\top \phi(x_i)\big)^2 + \lambda \lVert w - w_{\text{syn}} \rVert^2 = (\Phi^\top \Phi + \lambda I)^{-1}(\Phi^\top y + \lambda w_{\text{syn}}), $$

Here $\Phi$ is the $n \times D$ matrix of real features and $\lambda$ sets how much the real data is trusted relative to the simulator. Two limits follow directly. At $n = 0$ the model is the pure simulator retrieval, which gives a zero-shot prediction for lakes nobody has sampled. As $n$ grows, the head approaches ordinary kernel ridge regression on the real data. Because each real sample enters only this $D$-dimensional solve, we can measure the exchange rate as a curve: the number of real match-ups a model with synthetic support needs in order to match a real-only model trained on more samples.

The second, falsifiable claim borrows from the tracking finding that behavioural complexity mattered more than photorealism. We predict that the diversity of simulated constituent mixtures, meaning decorrelated algae, sediment and dissolved organic matter, matters more to the exchange rate than the radiative-transfer accuracy of each simulated spectrum.

Research Program

The simulator is a semi-analytical bio-optical forward model. It maps sampled inherent optical properties of the constituents to remote-sensing reflectance. It also exposes two independent sets of controls. Fidelity controls cover the reflectance approximation used, illumination, and sensor noise and band response. Diversity controls cover the joint distribution of constituent concentrations and their spectral shapes. The research file contains no paper describing a specific water radiative-transfer package, so the forward model is a design component here, not a cited result.

From one to two hundred thousand simulated spectra, we choose $D$ Nyström landmarks by subsampling and set the kernel bandwidth by cross-validation on synthetic data alone. This removes the swarm search that PSO-LSSVM runs on scarce real points. We then fit $w_{\text{syn}}$ in closed form. Following UniCon's observation that closed-form kernel solutions require frozen representations, we deliberately freeze $\phi$. The simulator defines the representation, and reality adjusts only the readout.

We treat simulator evaluation as its own object of study. Is this Model for Real? compares goodness-of-fit landscapes over parameters, fitted on real versus simulated data, to judge whether a model reflects reality. We transfer that test to our setting. For each lake, we compute the validation-error landscape over $(\lambda, \text{bandwidth})$ on real match-ups and on matched synthetic draws. We then ask whether the distance between those landscapes predicts the exchange rate. If it does, practitioners get a check on simulator quality that needs no held-out real labels beyond the calibration set.

We also lesson-check the physics. PPIDM shows that strictly enforcing incomplete physics hurts. Its fully constrained chlorophyll model had infilling RMSE 0.538, against 0.410 for an unconstrained diffusion model and 0.270 for the partially constrained one. This is why the simulator enters only as a soft prior through $\lambda$, never as a hard constraint.

Speed is treated as a constraint rather than the contribution. Optimizing Inference Speed for Real Time Deep Learning notes that the field reports accuracy but rarely runtime, and singles out ensembles and stacked models as latency drivers. ESA-MDN and the FOD-DWT pipeline, with its 11 feature selectors and multi-order derivatives, are examples. Our predictor costs one $D$-dimensional kernel evaluation against fixed landmarks plus a dot product per pixel, independent of training-set size. We will profile throughput against push-broom line rate on an embedded GPU and on a CPU.

Scope is honest. Simulators produce optically active quantities, so chlorophyll-a, suspended solids and turbidity are in scope. Non-optically-active targets such as total phosphorus and chemical oxygen demand, which the Guangzhou urban-river study retrieves only through correlations specific to each site, are out of scope.

Baselines

The naive reference is a single-band or band-ratio empirical fit per lake. It is the least costly option for edge hardware and the method a skeptic would deploy. CoastalWQL reports such a single-band turbidity fit at 715 nm reaching $R^2 = 0.594$, RMSE 8.562 FNU on raw data and $R^2 = 0.884$, RMSE 4.579 FNU after radiometric, de-striping and glint correction. Every method in our study receives the same corrected reflectance, so preprocessing is not confounded with modelling. Beating this baseline means lower held-out-lake RMSE at every budget of $n$ real match-ups. The comparison matters most at $n \le 10$, where empirical fits are still easy to estimate.

The strongest existing method on the multi-lake question is MT-Hyper. It is a hypernetwork that generates a small per-lake regressor from lake metadata, trained across 88 GLORIA lakes with LakesATLAS attributes. It is the strongest competitor because it solves the same scarcity problem using real data pooled across lakes instead of simulation. The file reports only that its RMSE is "20% lower than all baseline methods" on the 88-lake dataset; no absolute RMSE for MT-Hyper was found on Lacuna. We will therefore re-run it on our exact lake splits with its published architecture, giving it the same tuning budget in GPU-hours as our bandwidth and $\lambda$ search. Beating it means lower RMSE on held-out lakes at $n \in \{0, 5\}$ without using metadata. The zero-shot case $n = 0$ is one MT-Hyper leaves untested.

The strongest per-site kernel method is PSO-LSSVM. On Beigong Reservoir suspended solids it reports test $R^2 = 0.95$, RMSE 0.75 mg/L, and on Shahu Port $R^2 = 0.964$, RMSE 28.56 mg/L. We run it per lake with its swarm search under the same evaluation budget.

Two ablations isolate the key ingredient. The first keeps the simulator-derived $\phi$ but sets $w_{\text{syn}} = 0$, isolating the head prior. The second builds $\phi$ from real spectra only, with landmarks and bandwidth chosen on real data, isolating the simulator-defined representation. A gain that survives both ablations is attributable to synthetic data. A gain that disappears in the second shows the simulator's value lies only in choosing the bandwidth.

Why not just…?

Why not just pre-train a neural network on simulated spectra and fine-tune it on real match-ups? This is the real threat, and the synthetic-data equivalency study used exactly that recipe. We run it as a full competitor. An MLP and a mixture density head in the style of ESA-MDN are pre-trained on the same synthetic set and fine-tuned with early stopping. If they match our model's held-out RMSE at $n = 5$, the contribution shrinks to the exchange-rate and simulator-diagnostic findings, which carry over to either learner. We would report that outcome plainly. Our prior is that SGD fine-tuning on five points is high-variance across random seeds while the closed-form solve is deterministic. We will therefore report the seed spread, not just the mean.

Why not just pool real data across lakes and skip simulation? MT-Hyper already shows that cross-lake pooling with metadata works. Our zero-shot and five-point lake-out experiments compare the two sources of prior knowledge directly. We also run a hybrid in which the ridge prior centre $w_{\text{syn}}$ is replaced by a head fit to pooled real lakes, and one in which both are combined. If pooled real data dominates simulation everywhere, that negative result is itself the substitutability answer the field lacks.

Why not just augment by sampling neighbouring pixels around each station? ESA-MDN expands 30 stations to 300 samples this way, but the extra pixels share one local water mass and add no new constituent mixtures. We include this augmentation in the real-only baselines. Beating it on held-out lakes, rather than held-out pixels of the same river, is the relevant test.

Proposed Experiments

The primary benchmark is GLORIA in situ reflectance with matched chlorophyll-a, using the 88-lake configuration of MT-Hyper. Suspended solids are added where GLORIA records them. Evaluation is leave-lakes-out, with real budgets $n \in \{0, 1, 2, 5, 10, 20, \text{all}\}$ drawn over 20 random draws per lake. The UAV datasets from Beigong Reservoir, Shahu Port, Lianqi River and Singapore coastal waters serve as secondary tests where the data can be obtained, with each simulator resampled to that sensor's bands. Metrics are RMSE on log concentration and $R^2$ on held-out lakes, the synthetic-to-real exchange curve (real samples saved per budget), and per-pixel latency and throughput against line rate.

Three checks decide the proposal's fate. The first is whether simulator support beats real-only kernel ridge and MT-Hyper on held-out lakes at $n \le 5$. If not, simulation is not worth a grab sample at the scales that matter. The second is a factorial sweep of simulator diversity against fidelity at fixed synthetic volume. Its shape tests the claim that mixture diversity drives the exchange rate. Failure here would mean better radiative transfer is the lever and the tracking-domain lesson does not transfer. The third is whether real-versus-synthetic error-landscape distance, computed lake by lake, correlates with the exchange rate across lakes. A strong rank correlation would give monitoring agencies a label-light check of whether to trust a simulator on a new water body.