REDNET-ML: A Multi-Sensor Machine Learning Pipeline for Harmful Algal Bloom Risk Detection Along the Omani Coast

The REDNET-ML project presents a reproducible, multi-sensor machine learning pipeline that fuses Sentinel-2 and MODIS satellite data with object detection evidence to generate calibrated, operational HAB risk probabilities for the Omani coast, validated through rigorous evaluation metrics and drift analysis.

The Gist

Imagine the coastline of Oman as a giant, busy kitchen. In this kitchen, there are huge desalination plants (which turn seawater into drinking water) and fishing fleets that rely on the ocean. Suddenly, a dangerous "soup" starts to form in the water: Harmful Algal Blooms (HABs). These are massive explosions of toxic algae that can kill fish, clog water intake pipes, and poison the water supply.

The problem is that these "soup explosions" are hard to predict. They happen fast, they are hard to see from the ground, and the data we have about them is often messy or late.

REDNET-ML is a new, smart "Kitchen Safety System" designed to predict these disasters before they happen. Here is how it works, explained simply:

1. The Three Sets of Eyes (Multi-Sensor Data)

Instead of relying on just one camera, this system uses three different types of "eyes" to watch the ocean, like a security team with different specialties:

• The High-Def Drone (Sentinel-2): This is a satellite that takes super-sharp, zoomed-in photos of the water near the shore. It looks for specific colors and textures that look like algae, kind of like a detective looking for a specific stain on a carpet.
• The Weather Reporter (MODIS): This is a broader satellite that doesn't zoom in as much, but it tracks the "big picture" health of the ocean, like water temperature and general greenness (chlorophyll). It's like checking the weather forecast to see if conditions are right for a storm.
• The Pattern Spotter (AI Detectors): This is a special AI trained to recognize the shape of an algal bloom. It doesn't just look at colors; it looks for the swirling, patchy patterns that algae make, acting like a security guard who recognizes a suspicious gait.

2. The "No-Cheating" Rule (Non-Leaky Evaluation)

In many computer projects, the AI gets "cheated" by studying the test questions before the exam. For example, if the AI learns the specific cloud patterns of one day, it might just memorize that day instead of learning what algae actually looks like.

The REDNET-ML team was very strict: They never let the AI see the future. They trained the system on old data and tested it on brand-new, future data. This ensures that when the system says "danger," it's actually smart, not just memorizing the past.

3. The "Safety Committee" (Decision Fusion)

Once the three "eyes" gather their clues, they don't just vote; they have a meeting.

• The CatBoost model acts as the Committee Chair.
• It listens to the High-Def Drone, the Weather Reporter, and the Pattern Spotter.
• It weighs all the evidence together to give a single Risk Score (from 0 to 100%).

4. The Two-Alarm System (Watch vs. Action)

The system doesn't just say "Yes" or "No." It uses a two-step alarm system to avoid panic but ensure safety:

• 🟡 WATCH (Yellow Light): The system sees something suspicious. It's not 100% sure yet, but it's worth a closer look. The human operators are told to "keep an eye on this area."
• 🔴 ACTION (Red Light): The evidence is strong. The risk is high. The operators are told to "take immediate action," like shutting down a water intake pipe or deploying cleanup crews.

5. Why This Matters

Think of this system as a smart smoke detector for the ocean.

• Old methods were like waiting until you smelled smoke (toxic algae) to realize there was a fire.
• REDNET-ML is like a detector that smells the smoke before the fire starts, giving the "kitchen staff" (desalination plants and fishermen) time to move the pots and pans to safety.

The Bottom Line

This project proves that by combining sharp photos, broad weather data, and smart pattern recognition, we can build a reliable early-warning system. It's designed to be transparent (you can see why it made a decision) and robust (it won't break when the ocean conditions change slightly). It turns a chaotic, dangerous ocean problem into a manageable, data-driven safety plan.

Technical Summary

1. Problem Statement

Harmful Algal Blooms (HABs) pose significant threats to coastal infrastructure, fisheries, and desalination plants (critical for water supply in Oman). While satellite remote sensing offers a scalable monitoring solution, translating raw imagery into actionable risk signals faces three major challenges:

• Sparse and Noisy Ground Truth: HAB labels are often delayed, heterogeneous, and limited in quantity.
• Data Leakage: Naïve train/test splits in remote sensing often lead to overestimated performance because models learn scene-specific artifacts (e.g., specific overpass conditions) rather than generalizable bloom patterns.
• Temporal Distribution Shift: Environmental conditions change over time, causing models trained on historical data to degrade in performance when deployed on future data.

The project aims to build a reproducible, non-leaky, multi-sensor machine learning pipeline to detect HAB risks specifically for coastal plants along the Omani coastline.

---

2. Methodology

The REDNET-ML pipeline is an end-to-end system that fuses data from multiple sources and employs strict evaluation protocols to ensure operational reliability.

A. Multi-Sensor Data Fusion

The system integrates three distinct data streams:

1. Sentinel-2 Optical Data: High-resolution (10–20m) "chips" extracted around coastal plants. Features include spectral indices (NDWI, Floating Algae Index/FAI), texture signals, and red/NIR contrast ratios.
2. MODIS Level-3 Data: Regional context predictors including ocean color (Chlorophyll-a, Kd490, NFLH) and Sea Surface Temperature (SST).
3. Learned Image Evidence: Instead of using deep learning detectors for final classification, they act as evidence generators. Models (Faster R-CNN, SSD) detect bloom-like spatial patterns and output compact score summaries (max/median confidence, detection counts) rather than binary labels.

B. Label Strategy & Dataset Construction

To address label scarcity, the project uses a hybrid labeling approach:

• Trusted Labels: A conservative subset from reliable manual checks.
• Mined Labels: Heuristic candidates derived from ocean color/thermal anomalies, filtered to reduce false positives.
• Non-Leaky Splits: The pipeline strictly avoids temporal and spatial leakage by using Group-Safe Cross-Validation (ensuring chips from the same satellite overpass are not split between train/test) and Temporal Cross-Validation (training on earlier years, testing on later years).

C. Decision Fusion Model

• Algorithm: CatBoost is selected for its robustness with tabular data and ability to handle non-linear interactions.
• Input Features: A fusion of Sentinel indices, MODIS predictors, detector score summaries, and seasonal encodings (sin/cos of month).
• Operational Logic: The model outputs a calibrated probability ($\hat{p}_t$). This is processed through a weighted blend involving:
  • HAB probability ($hab\_prob$)
  • Detector evidence ($det\_mean$)
  • Adjusted Ocean Color Index ($OCI_{adj}$)
  • Seasonal anomaly ($Season_{adj}$)
• Alert Policy: Two thresholds define the operational state:
  • WATCH: High sensitivity for candidate surfacing.
  • ACTION: High-confidence escalation for operational response.

D. Evaluation & Drift Analysis

The system is evaluated using AUPRC (primary metric due to class imbalance) and AUROC. Crucially, it includes Drift Analysis (PSI and KS distance) to quantify distribution shifts between training data (2017–2024) and recent data (2025), ensuring thresholds remain valid over time.

---

3. Key Contributions

• Non-Leaky Evaluation Framework: A rigorous methodology that prevents data leakage in time-series remote sensing, providing a realistic estimate of model performance.
• Evidence-Based Fusion: A novel approach where object detectors serve as feature extractors (evidence generators) rather than final classifiers, improving robustness under sparse labeling conditions.
• Operational Risk Viewer: An end-to-deployment workflow that includes a risk-field viewer for site-specific exploration, supporting decision-making for desalination plants and coastal infrastructure.
• Drift-Aware Thresholding: A mechanism to detect and account for temporal distribution shifts, recommending a "WATCH" policy for robust candidate surfacing while flagging "ACTION" thresholds for periodic recalibration.

---

4. Results

• Performance: Under group-safe cross-validation, the fused model achieved a mean AUROC of 0.842 and a mean AUPRC of 0.731.
• Operational Metrics: At an operating threshold optimized for minimum recall (≥0.60), the system achieved:
  • Precision: ~0.682
  • Recall: ~0.605
  • Confusion Matrix: 398 True Positives, 185 False Positives, 887 True Negatives, 260 False Negatives.
• Detector Benchmarking:
  • Faster R-CNN (ResNet50) and SSD MobileNet provided the best balance of regional recall (0.94) and area overlap.
  • YOLOv8 was evaluated but excluded as it did not significantly improve ensemble coverage.
• Drift Analysis: Significant distribution shifts were observed in 2025 data (PSI up to 5.2 for individual plants), confirming that static thresholds may degrade over time and necessitating the dual-threshold (WATCH/ACTION) strategy.

---

5. Significance

REDNET-ML represents a shift from purely academic HAB research to operational decision support. By addressing the "last mile" problems of data leakage, label sparsity, and temporal drift, the system provides a transparent, reproducible, and inspectable tool for managing coastal risks.

Its significance lies in:

1. Safety & Economics: Directly protecting critical infrastructure (desalination) and fisheries in the Gulf of Oman.
2. Methodological Rigor: Setting a new standard for remote sensing ML by prioritizing non-leaky evaluation and drift analysis over raw accuracy metrics.
3. Reproducibility: The entire pipeline is implemented as a modular repository with notebooks and scripts, exporting all evaluation artifacts (SHAP values, calibration curves, confusion matrices) for full transparency.