The impact of feature engineering and an optimisation framework for ocean colour machine learning

The 30-second take

• What: The authors define a seven-level ocean-colour feature-engineering sequence and show that MLP, SVM, and XGBoost accuracy on Norwegian Sentinel-3 OLCI data swings wildly with those choices.
• Why it matters: Coastal water-quality numbers from space are still scarce and noisy. Optimizing the inputs is a step toward cheaper, more default environmental monitoring.
• Who should care: Ocean-colour algorithm developers, coastal water-quality agencies, and applied-ML teams who usually only tune hyperparameters.

What the paper actually did

Machine learning is common in ocean-colour algorithms, but most papers tune model parameters and hyperparameters, not the data transformations that feed the model. This work measures feature engineering (FE) and proposes a seven-level optimisation sequence: band choice, log scaling, spectral-shape normalisation, index extraction, principal component analysis, feature scaling, and zero-to-one scaling.

They demonstrate the framework with multilayer perceptrons, support-vector machines, and XGBoost on Sentinel-3 OLCI scenes of Norwegian coastal waters, targeting chlorophyll-a concentration and Secchi-disk depth (Zsd). Accuracy varies sharply across FE recipes from six prior OLCI studies versus their optimised recipes: R ranges from 0.01 to 0.55 for chlorophyll-a and from 0.15 to 0.68 for Zsd, with optimised FE at the top.

Optimised-FE models could also double R and cut mean absolute error by up to 63% versus the CHL_OC4ME and CHL_NN standard algorithms. No single FE recipe won for every target and every model, so they argue FE must be optimised per application—and that the framework is a key to better coastal water-quality monitoring.

What makes this disruptive

The scarce capability is trustworthy coastal chlorophyll and clarity from a satellite, where standard algorithms struggle. If most of the lift is in a sequenced FE stack rather than a new architecture, the field’s default workflow is mis-aimed. That is a methods disruption with operational consequences.

Why it matters (outside the lab)

Abundance lens: accurate monitoring of coastal waters is scarce. Cheaper sensing and modeling that more people can use is the path—if the FE optimisation is done per task.

Horizon is mid: measurement and policy both matter. Near-term: a Norwegian OLCI case study, not a global operational switch.

Limitations & open questions

Results are for Norwegian coastal waters and two targets (Chl-a, Zsd) on three model families. R still tops out at 0.55 / 0.68 even when optimised—useful, not solved. No universal FE recipe exists in their results. Comparison to CHL_OC4ME and CHL_NN is encouraging but region-specific. Preprint ≠ global satellite product. Abundance is not automatic.