coral-reef-risk-fcm

A Hybrid Framework for Automated Coral Reef Risk Assessment by Zero-Shot Visual Perception and Fuzzy Cognitive Map Inference

Fabio Porcelli · Simone Cioffi · Emanuel Di Nardo · Angelo Ciaramella Department of Science and Technologies — University of Naples Parthenope

---

Overview

This repository contains the full reproducible pipeline described in the paper "A Hybrid Framework for Automated Coral Reef Risk Assessment by Zero-Shot Visual Perception and Fuzzy Cognitive Map Inference".

The system performs automated ecological risk assessment of coral reef ecosystems in two sequential stages:

1. Stage 1 — Zero-Shot Feature Extraction: Gemma 4-E4B (Google DeepMind, April 2026) is used as a multimodal perception layer. Given a raw underwater survey image, the model is prompted as an expert marine biologist and returns 25 quantitative ecological features in JSON format — no fine-tuning required.

2. Stage 2 — Fuzzy Cognitive Map Inference: The feature vector is fed as clamped input activations into a Fuzzy Cognitive Map (FCM) with an expert-initialised causal weight matrix. Iterative sigmoid dynamics propagate causal influence through 31 concept nodes until convergence, producing a six-dimensional ecological risk profile.

The entire pipeline runs in a zero-shot regime on a single NVIDIA T4 GPU (Kaggle platform) using exclusively open-weight models and public datasets.

---

Pipeline

Coralscapes Image (1024×2048 px)
        │
        ▼
┌─────────────────────────────┐
│  Stage 1 · Gemma 4-E4B      │  ← zero-shot, JSON output, no fine-tuning
│  25 ecological features     │
│  normalised to [0, 1]       │
└────────────┬────────────────┘
             │  feature vector x ∈ [0,1]²⁵
             ▼
┌─────────────────────────────┐
│  Stage 2 · Fuzzy Cognitive  │  ← expert weight matrix, sigmoid dynamics
│  Map (31 nodes, 94 edges)   │
│  clamped input activations  │
└────────────┬────────────────┘
             │
             ▼
     Risk Profile r ∈ [0,1]⁶
  ┌─────────────────────────┐
  │ ecosystem_health        │
  │ pollution_risk          │
  │ thermal_stress_risk     │
  │ habitat_degradation_risk│
  │ recovery_potential      │
  │ heritage_integrity_risk │
  └─────────────────────────┘

---

Ecological Features

The 25 features extracted by Gemma 4-E4B are grouped into three thematic categories:

Group	Features
Coral-specific (12)	coral_cover_alive, coral_bleaching, coral_morphology_diversity, dead_coral_algae, juvenile_coral_recruitment, coralline_algae, substrate_rugosity, rubble_cover, sediment_stress, water_turbidity, algae_coverage, sand_cover
General marine (7)	fish_abundance, seagrass_coverage, macro_invertebrate_diversity, sponge_coverage, coral_disease_signs, crown_of_thorns_starfish, soft_coral_abundance
Anthropogenic (6)	anchoring_mechanical_damage, ghost_fishing_gear, blast_fishing_damage, image_assessability, sea_urchin_density, marine_debris

---

FCM Dynamics

Risk concept activations are governed by:

$$c_j^{(t+1)} = \sigma!\left(\sum_{i \in C_{\mathrm{in}}} w_{ij}, x_i + \sum_{k \in C_{\mathrm{risk}}} w_{kj}, c_k^{(t)}\right), \quad j \in C_{\mathrm{risk}}$$

where $\sigma(z) = (1 + e^{-z})^{-1}$ is the sigmoid activation function. Input concepts are clamped throughout inference. Convergence is defined as $|c^{(t+1)} - c^{(t)}|_\infty < 10^{-5}$, with a maximum of 50 iterations. In practice, all risk concepts converge within 4–5 iterations.

---

Dataset

The pipeline is evaluated on the Coralscapes dataset (Sauder et al., ICCVW 2025):

• 2,075 high-resolution images (1024×2048 px)
• 35 dive sites across 5 Red Sea countries
• 174k expert-labelled polygons, 39 benthic classes
• Experiments reported here use the test split (N = 50 images)

---

Results

Scenario Analysis

FCM risk profiles for four synthetic ecological scenarios, validating the ecological consistency of the expert weight matrix:

Scenario	EH	PR	TSR	HDR	RP	HIR
Healthy reef	0.986	0.451	0.511	0.068	0.959	0.476
Post-bleaching	0.339	0.616	0.556	0.819	0.144	0.697
COTS + blast fishing	0.081	0.692	0.500	0.962	0.034	0.855
Urchin barren	0.674	0.500	0.500	0.540	0.459	0.568

EH = ecosystem health · PR = pollution risk · TSR = thermal stress risk · HDR = habitat degradation risk · RP = recovery potential · HIR = heritage integrity risk. Higher EH and RP indicate healthier conditions; higher PR, TSR, HDR, and HIR indicate greater risk.

Sample Result — test_0

ecosystem_health       = 0.69
pollution_risk         = 0.59
thermal_stress_risk    = 0.52
habitat_degradation_risk = 0.48
recovery_potential     = 0.51
heritage_integrity_risk = 0.50

Consistent with a moderately degraded reef exhibiting residual structural complexity (substrate_rugosity = 0.70, coral_morphology_diversity = 0.60) alongside early-stage stressors.

---

Repository Structure

coral-reef-risk-fcm/
│
├── gemma4-fcm.ipynb       # Main notebook (Kaggle T4, Python 3.12)
├── README.md
└── outputs/
    └── results.jsonl      # Per-image feature vectors and risk scores

---

Requirements

The notebook is designed to run on Kaggle with a T4 GPU. The required Kaggle datasets and model sources are declared in the notebook metadata:

• google/gemma/keras/gemma4-e4b (Gemma 4-E4B model weights)
• Coralscapes dataset (dataset version ID 15929973)

Main Python dependencies:

torch
transformers
Pillow
numpy
pandas
networkx
matplotlib
tqdm

For larger Gemma 4 variants (26B-A4B, 31B), 4-bit NF4 quantisation via bitsandbytes is supported.

---

Data

The Coralscapes dataset is not included in this repository. Download it via one of the following:

• HuggingFace (recommended): EPFL-ECEO/coralscapes
• Zenodo (DOI): https://doi.org/10.5281/zenodo.15061505
• Kaggle: pre-configured in the notebook metadata (no action needed)

Quickstart

# Clone the repo and open the notebook on Kaggle
# Make sure the T4 GPU accelerator is enabled

# Stage 1: extract features from a single image
features = extract_features_gemma4(image_path)  # returns dict with 25 features in [0,1]

# Stage 2: run FCM inference
risk_profile, history = simulate_fcm(features, n_iter=50)
print(risk_profile)  # {'ecosystem_health': ..., 'pollution_risk': ..., ...}

Results are written to /kaggle/working/results.jsonl (one JSON record per image) and a summary CSV is exported via summary.csv.

---

Citation

If you use this code or the framework in your research, please cite:

@inproceedings{porcelli2026coral,
  title     = {A Hybrid Framework for Automated Coral Reef Risk Assessment
               by Zero-Shot Visual Perception and Fuzzy Cognitive Map Inference},
  author    = {Porcelli, Fabio and Cioffi, Simone and
               Di Nardo, Emanuel and Ciaramella, Angelo},
  booktitle = {Proceedings of the ...},
  year      = {2026},
  institution = {University of Naples Parthenope}
}

---

License

The code in this repository is released under the Apache 2.0 License, consistent with the Gemma 4 model license.

The Coralscapes dataset is subject to its own terms — see the original dataset page for details.

---

Acknowledgements

This work uses Gemma 4 by Google DeepMind (Apache 2.0) and the Coralscapes dataset by Sauder et al. Experiments were conducted on the Kaggle platform with free T4 GPU access.