🔬 applied physics

Estimation of Fish Catch Using Sentinel-2, 3 and XGBoost-Kernel-Based Kernel Ridge Regression

This study proposes an XGBoost-kernelized Kernel Ridge Regression framework using Sentinel-2 and Sentinel-3 satellite imagery to accurately estimate fish catch by capturing the complex, nonlinear relationships between oceanographic factors and fish distribution.

Original authors: Kanu Mohammed, Vaishnavi Joshi, Pranjali Diliprao Patil, Sandipan Mondal, Ming-An Lee, Subhadip Dey

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kanu Mohammed, Vaishnavi Joshi, Pranjali Diliprao Patil, Sandipan Mondal, Ming-An Lee, Subhadip Dey

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

The "Satellite Fishing Guide": Finding Fish from Space

Imagine you are a chef trying to find the freshest ingredients for a massive feast. Instead of walking through a crowded market, you decide to use a high-tech drone to scan the ocean from miles above. You aren't looking for the fish themselves (they are too small and hide under the waves), but you are looking for the "oceanic fingerprints"—the colors of the water and the temperature—that tell you exactly where the fish are congregating.

That is essentially what this scientific paper is doing. Here is the breakdown of how they pulled it off.

1. The Problem: The Ocean is a Giant, Shifting Puzzle

Finding fish is hard because the ocean is constantly changing. Fish don't just hang out anywhere; they follow "buffets" of nutrients (like plankton) that are stirred up by temperature changes and currents.

Scientists have tried to predict fish locations before, but they usually use one of two methods:

The "Simple Rule" Method (Linear Regression): Like saying, "If it's sunny, people go to the beach." It’s too simple and misses the nuances.
The "Decision Tree" Method (XGBoost): Like a game of 20 Questions. "Is the water warm? Yes. Is it green? No. Then the fish are here." This is better, but it can be a bit "clunky" and jumpy in its predictions.

2. The Secret Sauce: The "XGBoost-KRR" Hybrid

The researchers decided to combine these two worlds to create a "Super-Brain" model.

Think of it like making a gourmet sauce.

The XGBoost part is like the heavy, flavorful ingredients (the meat and spices) that capture the complex, jagged patterns of the ocean.
The Kernel Ridge Regression (KRR) is like the cream or oil that smooths everything out, ensuring the sauce isn't too chunky and flows perfectly.

By using the "logic" of the decision trees to power the "smoothness" of the regression, they created a mathematical tool that can see the subtle, non-linear relationships between water color and fish populations.

3. The Eyes in the Sky: Sentinel-2 vs. Sentinel-3

The team used two different types of "eyes" (satellites) to look at the water near Taiwan:

Sentinel-2 (The Microscope): This satellite has incredible detail. It’s like looking at the ocean through a high-powered magnifying glass. It can see tiny, local changes in the water, but it doesn't cover a huge area at once.
Sentinel-3 (The Wide-Angle Lens): This satellite sees the "big picture." It’s like looking through a wide-angle lens at a landscape. It’s a bit blurrier, but it gives you a smooth, sweeping view of the entire ocean's health.

4. The Results: Did it work?

Yes! It worked brilliantly.

When they tested their "Super-Brain" model against the actual logs from real fishing boats, the results were stunning:

Accuracy Boost: Their hybrid model was much more accurate than the old methods, reducing errors by up to 61% when using the high-detail Sentinel-2 data.
Mapping the Feast: They created a "heat map" of the ocean. The map showed that fish weren't just scattered randomly; they were clustered in specific areas where ocean currents stir up nutrients—exactly where nature intended them to be.

Why does this matter to you?

This isn't just about math; it's about feeding the world.

By using satellites to predict where fish are, we can manage our oceans more sustainably. It helps prevent overfishing in certain areas and ensures that we can protect "Life Below Water" (one of the UN's Global Goals) while still making sure there is enough food for everyone on land. It’s like having a weather forecast, but for the world's most important food source.

Technical Summary: Estimation of Fish Catch Using Sentinel-2, 3 and XGBoost-Kernel-Based Kernel Ridge Regression

1. Problem Statement

Quantifying fish distribution and catch is critical for global food security and sustainable ocean management (supporting UN SDGs 2 and 14). However, establishing a reliable correlation between fish abundance and environmental oceanographic factors (such as Sea Surface Temperature, chlorophyll concentration, and upwelling) is complex due to the highly non-linear nature of marine ecosystems. Traditional remote sensing approaches often rely on simple regression or standard tree-based models, which suffer from limitations like piecewise constant predictions and heuristic biases, making them less effective at capturing the intricate, non-linear relationships between spectral data and biological productivity.

2. Methodology

The authors propose a novel hybrid machine learning framework: XGBoost-Kernelized Kernel Ridge Regression (KRR).

The Core Innovation: Instead of using standard mathematical kernels (like Linear or Radial Basis Function), the researchers use a trained Extreme Gradient Boosting (XGBoost) model to generate a custom kernel.
- They extract the leaf index functions from the $P$ trees of an XGBoost model.
- These indices are converted into a one-hot embedding matrix ( $Z$ ).
- The kernel $K$ is then defined based on these embeddings ( $K = \frac{1}{P} ZZ^T$ ), allowing the KRR to leverage the high-dimensional, non-linear feature extraction capabilities of gradient boosting within a ridge regression framework.
Data Sources: The study utilizes multispectral imagery from two different satellite sensors:
- Sentinel-2 MSI: High spatial resolution, useful for capturing fine-scale ecological interactions.
- Sentinel-3 OLCI: Coarser spatial resolution but provides smoother spectral responses suitable for bulk water-quality characterization.
Ground Truth: Monthly logbook data (catch in kg) from the Taiwanese fishery targeting seabreams (22°N, 117°E to 27°N, 126°E) was used for training and validation.

3. Key Contributions

Hybrid Model Development: The creation of the XGBoost-KRR framework, which bridges the gap between tree-based feature extraction and kernel-based regression.
Sensor-Agnostic Application: Demonstrating that the proposed method is effective across different satellite platforms (Sentinel-2 and Sentinel-3) despite their differing spatial and spectral resolutions.
Spectral Sensitivity Analysis: A detailed characterization of how specific wavelength bands (e.g., red-edge bands in Sentinel-2 and blue-green bands in Sentinel-3) respond to oceanographic conditions and biological productivity.

4. Results and Discussion

The performance of the XGBoost-KRR (KRR-XGB) was compared against Linear (KRR-Lin) and Radial Basis Function (KRR-RBF) kernels.

Sentinel-2 Performance: KRR-XGB significantly outperformed other models, reducing the Root Mean Square Error (RMSE) by 59% to 61%. It achieved a very high correlation coefficient ( $\rho = 0.924$ ) and a massive improvement in the D-value (a measure of predictive accuracy).
Sentinel-3 Performance: KRR-XGB again proved superior, reducing RMSE by 27.6% to 40% and achieving a strong correlation ( $\rho = 0.731$ ).
Spatial Mapping: The model generated spatial catch maps for the coast of Taiwan. The maps showed high-intensity catch areas in the North-East, which aligns with oceanographic knowledge regarding upwelling and nutrient-rich vertical mixing in that region, while the South-West showed lower intensity due to oligotrophic conditions.

5. Significance

This research advances the field of satellite-based fisheries assessment by providing a more robust, non-linear tool for estimating biomass. By successfully integrating high-resolution MSI data with bulk-water OLCI data, the methodology offers a scalable approach for monitoring marine ecosystems. This has direct implications for sustainable ecosystem management, helping authorities predict fish availability and manage marine resources more effectively to ensure long-term food security.