Fishes of the Bay of Biscay — Community Modelling

A portfolio project combining marine ecology and spatial machine learning (K-Means + spatially blocked Random Forest) to understand fish community patterns in the Bay of Biscay.

Ecology × ML portfolio project. I built spatial–temporal–depth “cells”, derived a Hellinger-transformed community matrix, clustered 5 fish community types (C1–C5) with K-Means, then trained a spatially blocked Random Forest to predict community type from environment.

Performance: accuracy ≈ 0.39, macro-F1 0.357 (↑ from 0.317 baseline).
Top drivers: sst_actual, ocean_depth, coast_km, chl.

Highlights

• 5 communities (C1–C5) with clear spatial/seasonal structure.
• Spatial cross-validation (GroupKFold) to reduce leakage across space.
• Interpretability: permutation importances, partial dependence, and MDS of predictor space.

Workflow (Ecology × ML)

1. Cells & community matrix — Built spatial–temporal–depth cells; computed species-level community matrix and applied Hellinger transform.
2. Unsupervised structure — K-Means → 5 communities (C1–C5) to capture dominant assemblage patterns.
3. Supervised prediction — Random Forest predicts community_type from environment with spatial-block CV (GroupKFold) to reduce spatial leakage.
4. Model tuning & baseline — Grid search around RF hyperparams; compared to a stratified baseline.
5. Interpretation — Permutation importances (drivers: sst_actual, month_sin/cos, ocean_depth, coast_km, chl) and class-wise PDPs (not all shown here to keep the README lean).

Key Results

Item	Value
Communities discovered	5 (C1–C5)
CV scheme	GroupKFold (spatial blocks)
Final model	Random Forest (tuned)
Accuracy	≈ 0.39
Macro-F1	0.357 (↑ from 0.317 baseline)
Top predictors	sst_actual, month_sin, month_cos, ocean_depth, coast_km, chl

Visuals above: community map (ecological pattern) and permutation importance (drivers).

Data Sources

All environmental and biological data used in this project are open-access.

• Fish occurrence data:
  OBIS – Ocean Biodiversity Information System (accessed via the OBIS API).
  Records were filtered for Actinopterygii (ray-finned fishes) within the Bay of Biscay polygon.

• Sea surface temperature (SST):
  HadISST1 dataset (Hadley Centre, UK Met Office).
  Monthly SST fields extracted for 2015–2020.

• Chlorophyll concentration (CHL):
  Copernicus Marine Environment Monitoring Service (CMEMS).
  Near-surface chlorophyll-a data, aggregated to monthly means.

• Bathymetry:
  GEBCO 2020 Gridded Bathymetry.

• Distance to coast:
  Computed internally from the GEBCO coastline using Euclidean distance in projected space.

All raw datasets are excluded from the repository (data/ folder is git-ignored) but can be retrieved from the original sources above.

Reproducibility

• Environment

  conda env create -f environment.yml
  conda activate github-fish-biscay
  jupyter lab
  

Author

Theo Murphy