Cellula Technologies • Earth Observation Division

Multispectral Water Segmentation Benchmark

Comprehensive comparative evaluation of 12-Channel Sentinel-2 U-Net, ImageNet Transfer Learning (ResNet-34 & EfficientNet-B0), Spectral Band Ablation, and Generative Optimal Transport Flow Matching (OT-CFM).

Pairs: 306

|

Split: 244 / 62

|

Pixels: 1.01M

PEAK VALIDATION IoU SOTA

82.10%

+8.76% vs 12ch Scratch Baseline

Model: Pretrained ResNet-34 (12ch)

EDGE EFFICIENCY FRONTIER 6.25M Params

76.56%

~4x smaller footprint than ResNet-34

Model: Pretrained EfficientNet-B0 (12ch)

BANDWIDTH CHAMPION 6 Golden Bands

91.83%

Peak Precision (50% less satellite data)

Model: Pretrained ResNet-34 (6ch)

OT-CFM SYNTHESIS YIELD 48.5% Pass

+3.18%

509 Scenes vetted via Physical Gatekeeper

Generative Heun ODE 25-step paths

Interactive Model & Loss Curve Inspector

Click any model card below to dynamically load and animate its exact epoch-by-epoch loss trajectories, validation IoU progression, and confusion matrix.

Multi-Model Overlay Mode

Inspecting: ResNet-34 12ch

Train Loss / Precision Val Loss / Recall ★ Best Checkpoint Warmup Boundary

Epochs 1-3: Encoder Frozen Warmup • Epochs 4+: Full Fine-Tuning

Accuracy Champion

Pretrained ResNet-34 U-Net

SMP U-Net • 12 Bands (Full)

Params 24.46M

Global Pixel IoU

81.66%

Peak Val: 82.10%

F1-Score (Dice)

89.91%

Harmonic Mean

Precision

91.48%

Low False Alarms

Recall

88.39%

Water Coverage

Training Duration: 209.0 s (~2.35 s/epoch)

Total Epochs Logged: 89 epochs

Best Checkpoint Saved: Epoch 29 (Val IoU: 82.10%)

Pixel Confusion Matrix

62 Val Scenes • 1,015,808 Pixels

True Positive (TP) 22.1%

225,012

Water Correctly Predicted

False Positive (FP) 2.1%

20,963

Land Mistaken for Water

False Negative (FN) 2.9%

29,568

Missed Real Water

True Negative (TN) 72.9%

740,265

Land Correctly Classified

Total Water Detection Accuracy: 81.66% IoU

Official Master Benchmark Matrix

Cross-architecture evaluation under identical split (244 train / 62 val) and deterministic seed (42).

Model / Experiment ID	Architecture	Bands	Params	Global IoU	Precision	Recall	F1-Score	Duration	Inspect

Pareto Efficiency & Metric Comparisons

Analyzing parameter footprints versus segmentation accuracy frontier, and grouped multi-metric rankings. Click any point or bar to inspect that model's exact curves.

Pareto Efficiency Frontier (IoU vs Params M)

Optimal models maximize IoU while minimizing parameter footprint. Click any point to inspect.

Interactive Points

Scientific Pareto Discovery:

Pretrained EfficientNet-B0 (6.25M) forms the lightweight Pareto vertex, surpassing the Scratch Baseline (7.77M) by +3.23% IoU while being 20% smaller. ResNet-34 (24.46M) forms the accuracy apex at 81.66% IoU, dominating all alternatives.

Metric Multi-Bar Rankings

Comparing Global IoU, Precision, Recall, and F1 across models. Click any bar to inspect.

Global IoU

F1-Score

Precision

Recall

Spectral Physics Studio & NDWI Calculator

Explore the 12 Sentinel-2 MSI spectral bands, water absorption mechanics, and test the real-time NDWI/MNDWI simulator. Click any band chip to highlight its role.

Sentinel-2 Multispectral Instrument (MSI) 12-Band Architecture

★ Golden 6-Band Subset Discarded (Aerosol / Red-Edge)

1. Core Absorption Law

Water strongly absorbs in the Near-Infrared (B8) and Short-Wave Infrared (B11, B12) with reflectance near zero ($R pprox 0$), while terrestrial vegetation strongly reflects ($R pprox 0.5$).

2. The 6-Band Golden Subset

Isolating B2, B3, B4, B8, B11, B12 preserves all essential indices while cutting training time by 37.2% and reducing satellite downlink volume by 50%.

3. Discarded Channels

B1 & B9 are coarse 60m atmospheric correction bands. B5, B6, B7 are narrow red-edge transitions that introduce redundant features without improving boundary delineation.

Live Normalized Difference Water Index (NDWI) Simulator

Adjust band surface reflectance sliders or choose natural surface presets to see real-time index calculations.

Presets:

Green Band Reflectance (B3 • 0.560 µm) 0.22

NIR Band Reflectance (B8 • 0.842 µm) 0.02

SWIR-1 Band Reflectance (B11 • 1.610 µm) 0.01

NDWI (McFeeters, 1996)

(Green - NIR) / (Green + NIR)

+0.833

Water Pixel (> 0.0)

MNDWI (Xu, 2006)

(Green - SWIR) / (Green + SWIR)

+0.913

Clear Water Body

Optimal Transport Flow Matching (OT-CFM) Lab

Generating multispectral satellite scenes along straight probability trajectories with automated 3-Stage Physical Gatekeeper vetting.

Straight-Path Probability Flow: $x_t = (1-t)x_0 + t x_1$

25 Heun Steps

t = 0.0

Gaussian Noise

N(0, I) [6, 128, 128]

→ dx/dt = v_θ(x_t, t, c) →

2nd-Order Heun Predictor-Corrector

t = 1.0

Sentinel-2 Scene

Synthetic Multispectral

Continuous Normalizing Flow vs Curved Diffusion:

Standard DDPM diffusion requires 1,000 curved steps. Optimal Transport CFM solves straight vector fields in 25 Heun steps without score approximation drift, generating 1,050 candidates in 312 seconds on Kaggle T4 GPU.

3-Stage Physical Quality Gatekeeper Funnel

48.5% Acceptance

Generated Candidate Scenes: 1,050 scenes (7 seeds)

Stage 1: Variance Filter (σ < 0.025)

Rejects degenerate flat/solid generations

-280 rejected

770 remaining

Stage 2: Physical NIR Absorption (Δ_NIR ≥ 0.015)

Enforces physical law: water must absorb NIR

-140 rejected

630 remaining

Stage 3: Spatial Alignment (IoU ≥ 0.15)

Enforces structural overlap with conditioning mask

-121 rejected

509 passed

Accepted High-Fidelity Dataset:

Scaled Training Pipeline: 244 real → 753 total scenes (+208.6%)

509 Scenes

+3.18% IoU Boost

Qualitative Water Regimes & Spatial Error Breakdown

Performance profiling across hydrological environments and physical spatial downsampling constraints.

REGIME 1 99.8% Prec

Open Ocean & Deep Lakes

Uniform water surfaces with pronounced NIR absorption. ResNet-34 achieves zero false alarms and 99.8% precision across expansive bodies.

Status: Fully Mastered

REGIME 2 91.8% Prec

Meandering Rivers & Coastlines

Curvilinear shorelines. The skip connections in U-Net successfully preserve high-frequency edge gradients along complex coastal sandbars.

Status: Sharp Delineation

REGIME 3 Challenging

Narrow Streams (≤ 2 Pixels)

The 32x encoder bottleneck ($128 ightarrow 4$ px) compresses thin streams to sub-pixel latent dimensions, causing occasional stream fragmentation.

Physical Bottleneck Identified

REGIME 4 MNDWI Key

Turbid Waters & Wetlands

Suspended sediments elevate green and red reflectance. Inclusion of SWIR-1 (B11) is mathematically required to avoid false vegetation classification.

Solved via SWIR-1 (B11)