>_ Jorge Uc / projects← All projects
~/projects/satellite-disasters // DEEP LEARNING
◆ FloodScope AISentinel-1PyTorchResearch case study

Seeing water.
Measuring uncertainty.

FloodScope AI is an end-to-end research project for satellite flood segmentation. I trained and compared U-Net and DeepLabV3+ models, evaluated their performance across geographic regions, and built an interactive Streamlit dashboard. A separate Nepal experiment exposed how badly a model can fail when its input domain changes.

FloodScope AI Streamlit dashboard showing segmentation research metrics
FLOODSCOPE AI
Dashboard screenshot: dashboard-overview.png
RESEARCH PROJECT / FS-001PYTORCH + STREAMLIT
LABELED TILES446Sen1Floods11 training source
BEST VALIDATION DICE0.7301U-Net V2 · 512 px
V4 TEST DICE0.7185Exploratory comparison
EXPERIMENTS04Architectures and losses
01 / PROJECT CONTEXT

Can a model identify flooded terrain beyond the places it has seen?

Satellite imagery can support disaster monitoring, but pixel-level segmentation must be evaluated carefully. A visually convincing mask does not automatically represent a dependable flood map.

01SEGMENTATION

Identify water-related flood labels

Learn a pixel-wise binary mask from labeled satellite images and inspect spatial errors instead of relying on a single accuracy figure.

COMPUTER VISION
02GENERALIZATION

Evaluate unseen regions

Hold out geographic regions during training so the evaluation tests more than memorization of nearby imagery.

GROUP-AWARE SPLITS
03TRANSPARENCY

Show when a model breaks

Compare predictions with ground truth and investigate out-of-domain inputs rather than presenting every output as correct.

RESPONSIBLE MODELING
02 / DATA ENGINEERING

From satellite tiles to reproducible experiments.

The project uses the prepared RGB PNG imagery and binary masks supplied in a Sen1Floods11-based dataset. The PyTorch loader reads aligned pairs directly from ZIP archives.

> preprocessing.pipelineIMPLEMENTED WORKFLOW
01 / INPUT

Load paired tiles

RGB images and matching binary masks.

02 / SPLIT

Group by region

Train, validation and test use distinct regions.

03 / RESIZE

Prepare pixels

256 or 512 px; mask nearest-neighbor resize.

04 / TRAIN

Optimize models

Binary loss combinations and horizontal/vertical flips.

05 / REVIEW

Inspect outputs

Dice, IoU, threshold tuning and region-level errors.

Input format: 3-channel RGB scaled to [0, 1] · Masks: 0 / 255 converted to binary labels · Region-level separation

TRAIN2927 regions
VALIDATION98India · Spain
TEST56Mekong · Somalia

The region-aware split limits geographic overlap, but it does not by itself establish global generalization. Split sizes are based on the recorded project dataset.

03 / MODEL DEVELOPMENT

Four experiments. Different trade-offs.

I compared image resolution, loss functions and network architecture rather than selecting a model from a single run.

V1 / BASELINE

U-Net · 256 px

ResNet18 encoder with BCE + Dice loss. Established the first segmentation baseline.

V2 / RESOLUTION

U-Net · 512 px

Higher input resolution, BCE + Dice and mixed precision training. Highest validation Dice.

V3 / LOSS STUDY

U-Net · Tversky

Compared a different loss formulation to study the precision/recall trade-off.

V4 / ARCHITECTURE

DeepLabV3+ · 512 px

ResNet18 encoder and atrous spatial processing. Highest observed exploratory test Dice.

> inference.flowCONCEPTUAL OVERVIEW
INPUTRGB tile3 × H × W
→
FEATURE EXTRACTIONResNet18Convolutional features
→
SEGMENTATIONU-Net / DeepLabV3+Dense pixel-wise logits
→
OUTPUTProbability mapSigmoid + chosen threshold
04 / QUANTITATIVE RESULTS

Performance depends on how you measure it.

Validation performance guided model selection; test results are reported transparently as exploratory comparisons because the test set was consulted during development.

> model.metricsDICE SCORE ↑
MODELVALIDATIONTEST
V1 · U-Net 2560.71980.6528
V2 · U-Net 5120.7301 BEST VAL0.6799
V3 · U-Net Tversky0.7270Not evaluated
V4 · DeepLabV3+0.71890.7185 OBSERVED

Scores use tuned thresholds where recorded. The held-out test was not preserved as a strictly untouched final benchmark after multiple comparisons.

MODEL SELECTIONValidation and test tell different stories.

V2 achieved the highest recorded validation Dice. V4 produced the highest observed test Dice, but that observation alone is not proof of superior real-world performance.

V1 / VAL
.720
V2 / VAL
.730
V3 / VAL
.727
V4 / VAL
.719
Evaluation note. Validation-selected thresholds and exploratory test comparisons must be distinguished. Dice and IoU depend on class balance, annotation conventions and geographic domain.
05 / INTERACTIVE APPLICATION

From experiments to an explorable dashboard.

I built a Streamlit research interface to bring model results, segmentation examples, regional performance and the Nepal failure analysis into one place.

REAL APPLICATION · STREAMLIT

Explore the model.
Not just a metric.

Choose images from the labeled dataset, inspect ground truth, run V2 or V4 with local checkpoints, and examine prediction errors.

The interface uses a warm cream-and-blue visual identity, while this portfolio case study retains the site's orange terminal aesthetic.

Model ExplorerRegional EvaluationComparisonNepal Study
Actual FloodScope AI Streamlit dashboard overview
Add dashboard-overview.png to your satellite-disasters assets folder.
06 / GEOGRAPHIC GENERALIZATION

The same model. Very different regions.

Performance was not uniform across geographic areas. Region-specific evaluation helped expose failure cases that an aggregate metric would obscure.

MEKONG / TEST0.8648

DeepLabV3+ Dice

The model performed substantially better on the recorded Mekong samples.

SOMALIA / TEST0.2934

DeepLabV3+ Dice

Performance fell sharply, highlighting a limitation in cross-region generalization.

INTERPRETATION

One global score can hide local weaknesses.

Averaging across regions can conceal geographic disparities. These observations motivate per-region diagnostics and more representative evaluation rather than claims of universal flood detection.

07 / OUT-OF-DOMAIN EXPERIMENT

The Nepal experiment: when transfer fails.

Using a separate UNOSAT/Nepal VV/VH tile set, I reconstructed exploratory three-channel composites and applied V4. Input channels differed substantially from training imagery.

PAIRED VV/VH TILES2,073Available in the case-study ZIP
DIAGNOSTIC SAMPLE80Investigated channel distributions
EXPLORATORY INFERENCE4Example tiles assessed with V4
A CRITICAL FAILURE MODE

High confidence is not the same as correctness.

With a visualization threshold of 0.45, V4 marked approximately 97.7% to 99.7% of pixels positive across the four filtered Nepal examples. These numbers are model outputs—not verified flooded-area percentages.

The original RGB training transformation was not fully reconstructed from the available material. Matching tensor shape and numerical range does not establish SAR equivalence.

TRAINING DOMAINSen1Floods11 RGB PNGsKnown labels · evaluated
↓ ≠ ↓
TRANSFER DOMAINNepal VV/VH compositesUnverified preprocessing · no ground truth
Responsible disclosure. FloodScope AI does not claim validated flood maps, affected population totals or flooded area in Nepal. This experiment documents domain shift, not disaster-impact estimates.
08 / PERSONAL RETROSPECTIVE

What I learned building FloodScope AI.

The most useful outcome was not a single score. It was learning how dataset design, model selection and uncertainty affect the validity of a computer vision system.

BEYOND THE TRAINING LOOP

A model's apparent success depends on what the evaluation actually tests. This project pushed me to investigate geographic bias, investigate failures, and communicate results without overclaiming.

01 / DATA DESIGN

Splits define the question

A random tile split can exaggerate generalization if nearby imagery appears across sets. Grouping by region made the experiment more meaningful and exposed geographic weaknesses.

Data leakageGroup-wise splitsDataset design
02 / MODEL DEVELOPMENT

Better architecture is not automatically better

Changing resolution, loss function and network architecture produced different trade-offs. Validation and test rankings did not agree, reinforcing disciplined model selection.

U-NetDeepLabV3+PyTorch
03 / ERROR ANALYSIS

Aggregate metrics need context

The gap between Mekong and Somalia demonstrated why I need to inspect regional results, false positives and false negatives—not just present a headline Dice score.

Dice / IoUDiagnosticsRegional bias
04 / DOMAIN SHIFT

Inputs can look compatible but mean different things

Exploratory Nepal inference produced overwhelmingly positive predictions. I learned to check SAR preprocessing and input distributions before trusting cross-dataset transfer.

Sentinel-1OOD detectionResponsible AI
05 / PRODUCT ENGINEERING

Research is more useful when others can inspect it

Turning training scripts and JSON metrics into a Streamlit interface made it easier to compare experiments, inspect masks and understand model limitations.

StreamlitVisualizationReproducibility
06 / SCIENTIFIC INTEGRITY

Knowing when not to claim success matters

Without Nepal labels and a verified input transformation, it would be misleading to report flooded area or population estimates. Communicating that boundary became part of the engineering work.

ValidationUncertaintyScientific reporting
KEY TAKEAWAY

Reliable geospatial AI starts with careful evaluation—not a convincing visualization.

FloodScope AI connected deep learning, satellite imagery, data engineering and deployment with a broader understanding of how and why models fail outside their training conditions.

09 / SOURCE CODE & DOCUMENTATION

Explore the project.

The public repository contains the application and experiment code, documented results and instructions for running the local dashboard. Datasets and model weights are handled separately.

PUBLIC REPOSITORY

FloodScope AI on GitHub

Browse the PyTorch models, training workflows, evaluation scripts, documentation and Streamlit source code.

Open repository ↗
REPRODUCIBLE RESEARCH

Setup & methodology

Review dataset preparation, model comparison, evaluation assumptions and the documented transfer-domain limitations.

Read documentation ↗