A new preprint reports that satellite water-quality estimates can vary sharply depending on the feature-engineering steps used before machine-learning models see the data. In Norwegian coastal and fjord waters, optimized configurations produced the highest reported correlations for both chlorophyll-a concentration and Secchi disk depth, a measure of water clarity, but no one configuration was optimal across all targets and model types.
The researchers tested Sentinel-3 OLCI reflectance against 625 chlorophyll-a match-ups and 1,378 Secchi disk depth measurements. The models included multilayer perceptrons (MLPs), support-vector machines (SVMs) and XGBoost, so the feature-engineering search could be examined across different model families.
The model was only part of the experiment
Feature engineering here means the sequence of choices used to turn satellite reflectance into model inputs. The proposed search space had seven sequential levels: band selection, log scaling, spectral-shape normalization, index extraction, principal-component analysis, feature scaling and a final transformation that placed values between 0 and 1.
A guided search called TPE selected feature-engineering settings and machine-learning hyperparameters together. It ran through 500 trials and treated the best-scoring configuration as optimal. Each experiment repeated the search 30 times, allowing the researchers to examine accuracy uncertainty and how often particular feature choices were selected.
For pipeline ranking, the training data were randomly divided into three splits, with three folds in each split, and configurations were ranked using an averaged R-squared score. The temporal test sets held out 2021–2023 for chlorophyll-a and 2022–2023 for Secchi depth; those periods represented 38% and 45% of the respective match-ups.
The methods specified Pearson R, a score for how closely estimates track measurements, together with RMSE and MAPE. R was calculated on base-10 log-scaled targets, while RMSE and MAPE used the original target scale. The paper's abstract and results also report MAE, however, and the supplied text does not define that measure, making the error comparisons less straightforward to interpret.
The reported gap depended on the comparison
Across the feature-engineering choices, Pearson R ranged from 0.01 to 0.55 for chlorophyll-a and from 0.15 to 0.68 for Secchi disk depth. Optimized feature engineering was at the top of those reported ranges, illustrating the spread in scores among the alternative input pipelines.
Against the standard CHL_OC4ME and CHL_NN algorithms, the abstract reports that optimized-feature machine-learning models could improve R by two times and reduce mean absolute error by up to 63%. Those are reported predictive comparisons within the study; they do not establish a causal improvement in water-quality monitoring or a direct ecological benefit.
In a chlorophyll-a benchmark, the average of 30 SVM models labelled NOR_SVM reportedly increased R by 2.4 times and reduced MAE by 52% compared with OC4ME. For Secchi disk depth, NOR_SVM reportedly increased R by 1.9 times and reduced MAE by 63% compared with CHL_OC4ME.
The result was not one universal recipe. The feature-engineering choices that ranked best depended on both the target and the machine-learning model, leading the authors to describe the optimization as application-specific.
The boundary of the evidence
The study's geographic and optical reach is narrow. It examines Norwegian coastal and fjord waters, and no independent external validation across other regions or optical regimes is reported. The open question is whether the framework would hold in waters with different optical conditions, including eutrophic waters where near-infrared signals may be useful.
The chlorophyll-a sample was also mixed: 506 records followed the stated protocol, 12 were ferry-box measurements and 107 had no measurement-method description. The Secchi disk depth dataset contained 1,378 measurements from multiple sources and monitoring programmes. The supplied analysis flags those differences as possible measurement heterogeneity.
The optimization may not have fully settled every model choice. The authors note that some MLP configurations could remain unoptimized after the 500-trial search, while variability in MLP optimization was wider than for SVM and XGBoost.
Future work proposed in the analysis includes expanding the search to atmospheric correction, glint removal, adjacency correction and semi-analytical models. It also calls for validation on independent spatial and temporal datasets and for consistent definitions of the error metrics.
This is an arXiv:2608.19899v1 preprint dated 20 August 2026; its supplied front matter still contains placeholder journal fields.
The funding statement lists support from the BCCR/SKD NorHAB-ML project and INSTSTIP through Norwegian Research Council funding (#318085), Nansen Center basic funds grant 342624, NFR Climate Futures grant 309562 and a NORSTORE storage grant.
Paper data and sources
Original title: The impact of feature engineering and an optimisation framework for ocean colour machine learning
Authors: Edson Silva, Julien Brajard, Simon Cappe et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text