🔬 physics

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

This paper proposes a seven-level feature engineering optimization framework that significantly enhances the accuracy of machine learning models for estimating ocean color parameters like Chlorophyll-a and Secchi disk depth in Norwegian coastal waters, demonstrating that tailored data transformation is crucial for outperforming standard algorithms.

Original authors: Edson Silva, Julien Brajard, Simon Cappe, Lasse H. Pettersson, François Counillon

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Edson Silva, Julien Brajard, Simon Cappe, Lasse H. Pettersson, François Counillon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Satellites orbiting high above the Earth act as powerful eyes, scanning the oceans to reveal the invisible life and chemistry hidden beneath the waves. By measuring the color of the water reflected back into space, scientists can estimate how much algae is growing and how clear the water is, which are vital signs for the health of marine ecosystems. However, looking at the ocean from space is not like taking a simple photograph; the water near the coast is a chaotic mix of mud, dissolved organic matter from rivers, and microscopic plants, all of which twist and distort the light in ways that confuse standard computer programs. For decades, researchers have tried to build better computer models to interpret these signals, often focusing on tweaking the internal settings of the algorithms themselves. But a new study suggests that the real key to unlocking the secrets of coastal waters lies not in the computer's brain, but in how the data is prepared before it ever reaches the machine.

In the complex, often murky waters along the coast of Norway, standard satellite tools frequently fail. These algorithms were originally designed for the open ocean, where the water is clearer and the ingredients are more predictable. When applied to the Norwegian fjords and coastal currents, where dark, tea-colored water from rivers mixes with bright, chalky blooms of microscopic algae, the standard tools often produce wildly inaccurate results. They might see a harmless swirl of dark water and mistake it for a massive, toxic algal bloom, or they might miss a real bloom entirely. To solve this, a team of researchers turned to machine learning, a type of computer science where programs learn from examples rather than following rigid rules. While many studies have tried to improve these programs by adjusting their internal knobs and dials, this team asked a different question: what if the way we feed the data to the computer is the real problem?

The researchers focused on a process called feature engineering, which is essentially the art of cleaning and reshaping raw data so that a computer can understand it better. Imagine you are trying to teach a child to recognize different types of leaves. You could hand them a pile of leaves in various states—some wet, some dry, some torn, some whole—and expect them to learn. Or, you could first sort them by color, dry them out, and arrange them by size to make the patterns obvious. In the same way, the researchers took the raw light measurements from the Sentinel-3 satellite and subjected them to a series of seven distinct transformations. These steps included choosing which specific colors of light to use, adjusting the brightness levels to handle extreme differences, and creating new mathematical combinations that highlight specific water properties. They built a massive search space where every possible combination of these steps could be tested, treating the preparation of the data with the same rigorous attention usually reserved for tuning the computer model itself.

To test their idea, the team gathered thousands of real-world measurements from monitoring stations along the Norwegian coast, matching them up with satellite images taken on the same days. They trained their machine learning models to predict two critical things: the concentration of chlorophyll-a, which indicates the amount of algae, and the Secchi depth, a measure of how deep a white disk can be seen underwater, which tells us how clear the water is. They compared their new, optimized approach against the standard methods currently used by satellite agencies. The results were striking. The models that used the carefully optimized data preparation were significantly more accurate than the standard tools. In fact, the new approach improved the correlation between the satellite estimates and the real measurements by up to two times and reduced the average error by as much as 63 percent. The standard algorithms often struggled to distinguish between different types of water, frequently overestimating the amount of algae in dark, organic-rich waters, while the new models correctly identified these areas as having lower algae levels.

Perhaps the most surprising discovery was that there is no single "perfect" way to prepare the data for every situation. The best combination of data transformations depended entirely on which specific computer model was being used and what it was trying to predict. For example, the best way to prepare data for estimating algae clarity was different from the best way to prepare data for estimating water depth. This finding challenges the idea that there is a universal recipe for satellite ocean monitoring. Instead, it suggests that for every new region or every new goal, scientists must carefully search for the specific data preparation method that works best. The study demonstrates that by treating the data itself as something that can be optimized, rather than just a fixed input, we can build much sharper, more reliable tools for watching over our coastal waters. This approach offers a promising path forward for environmental monitoring, ensuring that the satellites watching our oceans see the truth, even in the most complex and challenging waters.

Technical Summary: The Impact of Feature Engineering and an Optimisation Framework for Ocean Colour Machine Learning

Problem Statement
Machine learning (ML) has become a prevalent tool for estimating water quality parameters, such as Chlorophyll-a concentration ([Chl-a]) and Secchi disk depth ( $Z_{sd}$ ), from satellite ocean colour data. However, current research predominantly focuses on optimizing model hyperparameters and architecture, while the optimization of the input data itself—Feature Engineering (FE)—remains largely ad-hoc and under-explored. In optically complex coastal waters (Case-2), standard algorithms often fail due to the non-covarying nature of optically active constituents. While FE techniques like band selection, scaling, and index extraction are commonly employed, the vast number of possible combinations suggests that manual selection is suboptimal. The authors hypothesize that treating FE as an optimization problem, similar to hyperparameter tuning, could significantly improve model accuracy in challenging environments like Norwegian coastal and fjord waters.

Methodology
The study proposes a systematic Feature Engineering (FE) optimisation framework applied to Sentinel-3 Ocean and Land Colour Instrument (OLCI) observations. The methodology involves the following components:

Study Region and Data: The research focuses on Norwegian coastal waters, characterized by high Colored Dissolved Organic Matter (CDOM) and coccolithophore blooms. The dataset comprises in-situ measurements of [Chl-a] (625 samples) and $Z_{sd}$ (1,378 samples) matched with Sentinel-3 OLCI reflectance data from 2016 to 2023. The data was pre-processed using the Polymer atmospheric correction algorithm.
The FE Space (Pipeline): The authors define a structured FE space consisting of seven sequenced levels of data transformation:
- Band Choice (BC): Selection of specific spectral bands (e.g., Blue, Red, Visible, or combinations based on previous literature).
- Log Scaling (LS): Logarithmic transformation of reflectance.
- Spectral Shape (SS) Normalisation: Normalizing reflectance by the sum of bands.
- Index Extraction (IE): Computation of band ratios, Normalized Difference Chlorophyll Index (NDCI), slopes, and Maximum Band Ratio (MBR).
- Principal Component Analysis (PCA): Dimensionality reduction.
- Feature Scaling (FS): Standardization using Standard Deviation (STD) or Inter-Quartile Range (IQR).
- Zero-to-One Scaling (ZOS): Min-max scaling.
Optimisation Algorithm: The authors employ the Tree of Parzen Estimators (TPE) algorithm to search the combined "pipeline space" of FE configurations and ML hyperparameters. TPE uses Gaussian Mixture Models to iteratively select configurations that maximize the coefficient of determination ( $R^2$ ) via cross-validation.
Machine Learning Models: Three distinct ML architectures were tested: Multi-Layer Perceptron (MLP), Support Vector Machines (SVM), and eXtreme Gradient Boosting (XGBoost).
Experimental Design: The study compares the "Optimised FE" approach against FE configurations adapted from six previous studies (Brando et al., Joshi et al., Lapucci et al., Pahlevan et al., Shen et al., Zoffoli et al.) and standard Sentinel-3 products (CHL_OC4ME and CHL_NN).

Key Results

Variability of Existing FE: The study found that accuracy varies significantly depending on the FE configuration chosen from existing literature. For [Chl-a], $R$ values ranged from 0.01 to 0.55, and for $Z_{sd}$ , from 0.15 to 0.68, depending on the FE used.
Performance of Optimised FE: The framework successfully identified FE configurations that outperformed both literature-adapted FE and standard algorithms.
- Compared to standard CHL_OC4ME and CHL_NN algorithms, the ML models with optimised FE improved the correlation coefficient ( $R$ ) by up to two times and reduced the Mean Absolute Error (MAE) by up to 63%.
- Specifically, the optimised SVM model (NOR_SVM) increased $R$ by 2.4 times for [Chl-a] and 1.9 times for $Z_{sd}$ compared to CHL_OC4ME.
Model and Target Dependency: A critical finding is that no single optimal FE configuration exists for all target variables or ML models.
- For [Chl-a], most models converged on visible band choices and, for XGBoost, log scaling.
- For $Z_{sd}$ , models often converged on visible bands including Near-Infrared (NIR) and log scaling.
- The optimal FE choices were highly specific to the combination of the target variable and the ML architecture.
Seasonal and Spatial Accuracy: The optimised models better captured seasonal bloom patterns (spring and autumn peaks) and geographical variability (e.g., distinguishing between dark CDOM-rich waters and cyan coccolithophore blooms) compared to standard algorithms, which tended to overestimate [Chl-a] in CDOM-dominated areas.

Significance and Claims
The paper claims that Feature Engineering is a critical, yet often overlooked, factor in the accuracy of ocean colour machine learning models, potentially more impactful than the choice of the ML architecture itself. By proposing a framework that treats FE selection as an optimization problem using search algorithms like TPE, the authors demonstrate that:

Optimisation is Necessary: Manual selection of FE based on previous studies is insufficient for complex coastal environments; systematic optimization yields superior results.
Context-Specificity: FE optimization must be performed individually for each application (specific target variable and ML model), as a "one-size-fits-all" solution does not exist.
Improved Monitoring: The proposed framework offers a pathway to significantly enhance the reliability of water quality monitoring in optically complex coastal and fjord waters, addressing issues like CDOM-induced overestimation that plague standard bio-optical algorithms.

The authors conclude that while the "black box" nature of transformed features may limit physical interpretability, the substantial gains in retrieval accuracy justify the approach, particularly as observational databases grow and the need for accurate regional models increases.

More like this