GeoAI framework for real-time marine pollution monitoring
This implements the AI Remote Sensing Pollution Monitor (AI-RSPM). Upload a satellite image (SAR or optical) and the system runs a convolutional-style pixel classifier, applies DBSCAN density clustering to segment oil spill footprints, and visualises LSTM-based trajectory forecasts informed by wind and current drift in the Gulf of Suez.
CNN / YOLOv11 pixel classification
DBSCAN clustering
LSTM 72h forecast
Lagrangian drift
Sentinel-1 · Sentinel-2
OpenWeather forcing
1 · Upload & detect
Drop a satellite image (SAR dark-slick or optical iridescent film).
Oil pixels are identified by the convolutional detector and coloured overlays are drawn.
Click or drop an image here
JPG / PNG / TIFF preview · Sentinel-1 GRD recommended
tag your upload with the correct polarization below so the detector
uses the matching σ⁰ threshold and speckle filter.
Active polarization for next analysis:
VV (co-pol)
VV (co-pol): best sensitivity to capillary waves oil slicks appear as distinct
dark patches. Standard channel per ESA Sentinel-1 guidelines for marine pollution.
Original image
No image yet upload or pick a sample.
CNN detection + DBSCAN clusters
Detection results will appear here.
OIL COVERAGE
—
CLUSTERS (DBSCAN)
—
LARGEST SLICK
—
CNN CONFIDENCE
—
Technical Assessment
- Sensor
- —
- Polarization
- —
- Backscatter range (σ⁰)
- —
- Acquisition date
- —
- Oil type
- —
- Scene size
- —
- Sensitivity
- —
- Signal quality
- —
- CNN confidence
- —
- Run analysis to produce an operational recommendation.
Feature Inventory
| ID | Shape | Area (px) | % Scene | Severity |
|---|---|---|---|---|
| Run analysis to populate detected features. | ||||
Signature Discrepancies
Pending
Run analysis to cross-validate CNN pixel classifier against DBSCAN clustering and scene statistics.
Future enhancements — roadmap to lift analysis & performance
Practical upgrades that would materially improve the detector's accuracy,
runtime and robustness on real Sentinel-1 tiles. Each card notes an
expected gain relative to the current VV single-channel CNN baseline.
Dual-pol Bayesian fusion (VV ⊕ VH)
accuracy
Replace the single-threshold detector with a per-pixel log-likelihood
combining VV (dampening) and VH (cross-pol speckle floor). Cuts
biogenic-film false alarms — the dominant look-alike in the Gulf.
+8–12 % F1 · fewer look-alike alerts
Polarimetric decomposition (H/α, Freeman-Durden)
accuracy
For Sentinel-1 dual-pol acquisitions, use entropy/anisotropy/alpha
parameters to discriminate oil from wind-shadowed sea, rain cells and
low-wind patches — the top sources of SAR false positives.
−30 % false-positive rate in calm seas
Attention U-Net segmentation head
accuracy
Swap the thresholding stage for an attention-gated U-Net trained
on ERA5-aligned Sentinel-1 tiles. Learns slick morphology
(feather, blocky, ribbon) rather than relying on raw σ⁰.
+6 % mIoU at 20 m/pix
MC-Dropout uncertainty maps
accuracy
Run T = 20 stochastic forward passes and emit a per-pixel
uncertainty layer. The operator can gate auto-tasking on low
variance, reducing spurious drone deployments.
Calibrated confidence · explainable alerts
WebGL / WebGPU kernel for DBSCAN
speed
Move the per-pixel CNN score and the DBSCAN neighbour-query onto
the GPU via a compute shader. Opens the door to 10 k × 10 k
whole-tile processing in-browser.
≈ 25× runtime speed-up on 5 MP images
ONNX / TensorFlow-JS quantised CNN
speed
Ship a distilled INT-8 CNN (≈ 2 MB) as an ONNX blob so the real
thesis model runs client-side — no synthetic scoring, no server
round-trip. Uses WebAssembly SIMD on CPUs without GPUs.
True thesis-model inference · < 400 ms per scene
Copernicus DataSpace live tiling
data
Stream the most recent Sentinel-1 IW GRD tile covering the active
AOI directly from the CDSE STAC endpoint. Removes the upload step
for monitoring workflows.
Fresh imagery every ~6 days · no manual upload
ERA5 wind + HYCOM current fusion
data
Ingest 0.25° ERA5 winds and HYCOM surface currents at the scene
centroid to replace the 3% wind-only Lagrangian drift with a proper
leeway + current model improves 72 h trajectory skill.
−40 % mean position error at 48 h
Temporal change detection (Δt baseline)
data
Cache the last 4 passes over a AOI and flag dark patches that appear
against a rolling baseline. Rules out persistent dark sea-state
regions (upwelling cells, rain shadows).
Fewer seasonal false alarms
Class-weighted contrastive pre-training
accuracy
Use SimCLR/MoCo on unlabeled Gulf-of-Suez S1 tiles before
fine-tuning, addressing the severe class imbalance (≈ 1:5,000 oil
vs. sea pixels) without aggressive oversampling.
+4 % recall at same precision
Distributed inference on Kubernetes
scale
Package the CNN + drift engine as a horizontally-scalable service
(KFServing / Ray Serve). Supports basin-wide continuous monitoring
— the operational target of the thesis.
~10³ AOIs / hour throughput
Vessel-AIS cross-correlation
scale
Join detected slicks with nearby AIS tracks (MarineTraffic / Global
Fishing Watch) to flag probable-source vessels and estimate spill
age from ship trajectories.
Actionable attribution · supports enforcement
2 · Gulf of Suez situational map
Known production fields and recorded spill hotspots around Ras Gharib.
Click a marker for field details.
Active oil field
Historical spill site
Detection target (Ras Gharib)
Drift trajectory (72h)
Active AOI
Area of Interest (AOI) — spill origin
or
Lat
Lng
Radius (km)
No AOI set. Simulation uses default Ras Gharib origin (28.353°N, 33.080°E).
OpenWeather API — real-time wind & sea conditions for Ras Gharib
No key yet — simulation uses a default SE wind of 5.5 m/s. Get a free key at openweathermap.org.
3 · LSTM 72-hour trajectory forecast
Probabilistic drift prediction combining a physical Lagrangian model with an LSTM
recurrent network trained on historical Red Sea currents and wind fields.
Pick a scenario from the thesis case study:
TIME TO SHORELINE
—
PEAK DRIFT DIST
—
CROSS-TRACK σ (95%)
—
RISK LEVEL
—
4 · LSTM 72-hour trajectory curve
Parametric trajectory graph — displacement from the spill origin in kilometres
(Δeast vs Δnorth). The mean LSTM ensemble path is shown as a curve with time
waypoints at 0, 6, 12, 24, 48, and 72 hours; the 95% confidence envelope comes
from the N = 128 Monte-Carlo perturbations of wind and current forcing.
This is the "drift butterfly" used in the thesis §3.2.7 to visualise spatial
uncertainty over the 72-hour forecast horizon.
BEARING
—
MEAN VELOCITY
—
TORTUOSITY (L/D)
—
95% ERROR ELLIPSE @ 72H
—
Validation: in the thesis, drifter-buoy tests yielded a mean spatial error of
312 m and directional accuracy of 83% at the 72 h lead time (Ch. 3.2.8).
The curve updates when you switch scenarios, regenerate the ensemble, or when
live OpenWeather wind is merged.