DryPath - Address-Level Evacuation Mapping from Multi-Frequency, Multi-Polarization SAR

View Our Presentation

Explore our comprehensive presentation about DryPath

View a PDF Version of Our Presentation

Problem

Floods evolve faster than conventional channels can deliver safe, actionable instructions on the ground. During storms, optical imagery is often unusable - clouds and heavy rain obscure the surface; gauges are sparse and uneven; and road closures, local inundation, and infrastructure failures can appear or shift within minutes.

Residents, first responders, emergency operations centers, schools, camps, and critical-infrastructure operators are left uncertain: evacuate or stay, which route is still passable, which facility remains reachable. People rarely die for lack of maps; they die for lack of timely, trusted routes. After-the-fact maps show where water was but don't answer the urgent, address-level question every household faces: "Can I get out now, and how?"

Critical Time Factor: Minutes lost between recognizing danger and choosing the correct path increase exposure to fast water, send emergency vehicles down dead ends, and strand vulnerable people without accessible shelter.

Traditional warnings are issued at county or town scale; they rarely provide per-address guidance that accounts for local topography, live road-network constraints, and the current inundation footprint. Meeting this operational need requires more than delineating flood extent - it requires converting remote-sensing and hydrometeorological inputs into human-scale recommendations: an explicit, safe path from a specific address to high ground or a designated shelter, with clear urgency and confidence. That conversion is the core challenge DryPath addresses.

How we solved it

We built DryPath, a web application that converts multi-frequency, multi-polarization radar observations and near-real-time rainfall into address-level flood alerts and safe evacuation routes, updating as new data arrive. Synthetic-aperture radar (SAR) operates through clouds, rain, and at night, providing crucial visibility when optical systems fail.

Our concept exploits differences in radar physics across polarizations and wavelengths to detect open water, urban double-bounce interactions, and under-canopy inundation; couples these signals with rainfall forcing and hydrologic context; and then translates the resulting geospatial products into street-by-street route decisions.

Operational Implementation:

Data Fusion: DryPath fuses NASA GPM IMERG rainfall with Sentinel-1 C-band (VV and VH) SAR for broad, frequent coverage, and ingests L-band SAR (where available, e.g., UAVSAR HH and HV) to separate saturated soils from true standing water beneath vegetation.

Radar Physics: C-band is sensitive to surface roughness and urban geometry - vertical walls produce characteristic double-bounce; smooth open water returns very low backscatter in both VV and VH. L-band's longer wavelength partially penetrates canopy and helps reveal water beneath trees - a common blind spot for optical sensors and shorter-wavelength SAR.

Processing Pipeline:

The processing pipeline begins with radiometric calibration to σ⁰, terrain correction, and precise co-registration of pre-flood, peak, and post-flood scenes to a common DEM. We then build informative composites and masks:

  • False-color RGB (e.g., VV, VH, VV and VH) to separate built areas, vegetation, and water
  • Temporal contrast composite R = VV(after), G = VH(after), B = VV(before) to make flood expansion visually "pop"
  • Adaptive thresholding on VH and VV with morphological cleanup yields robust open-water masks
  • HAND and slope filters suppress false positives from steep terrain or layover

In riparian forests where C-band is ambiguous, L-band HH and HV confirms under-canopy inundation (water remains dark while canopy volume-scattering patterns shift).

Change detection tracks flood expansion via strong decreases in VV and VH relative to pre-flood baselines, cross-checked against rainfall accumulations and susceptibility maps. The flood map connects directly to evacuation planning: road segments intersecting the water mask are closed; edges near the flood boundary or on steep slopes are risk-weighted; the system computes the safest available paths to shelters or high ground; and each address receives a clear alert tier - no risk, evacuate on foot, evacuate by car, or too late to drive (move to the highest floor and roof).

Why DryPath is useful

DryPath delivers value through three linked capabilities: timely coverage under adverse weather and darkness, physical discrimination of flood drivers across sensors, and direct translation of geospatial outputs into address-specific guidance.

All-Weather Operation: SAR's all-weather, day-night imaging provides actionable observations when decisions matter most.

Multi-Sensor Fusion: Multi-polarization, multi-frequency fusion improves discrimination: C-band captures open water and urban double-bounce affecting road conditions and access; VV and VH contrasts separate smooth water from roughened surfaces; L-band penetrates canopy to flag understory inundation that C-band and optical miss.

In practice, that means fewer unseen hazards - for example, a road that appears dry in a C-band-only product may be encircled by newly ponded water under trees, which L-band reveals.

Operational Beneficiaries:

  • Residents: Receive door-level guidance on whether and how to leave
  • First Responders and EOCs: Enhanced situational awareness and routing
  • Schools and Camps: Safe evacuation planning for vulnerable populations
  • Utilities and Public Works: Infrastructure protection and response coordination

Measurable Impacts: Shorter time-to-action and fewer wrong-turn dispatches.

View a Short Run-down of our presentation

Explore our comprehensive presentation about DryPath

Where it applies and what we did to make DryPath

We demonstrate DryPath in Kerr County, Texas, focusing on Kerrville neighborhoods along the Guadalupe River.

Input Data Sources:

  • Sentinel-1 GRD: C-band VV and VH
  • UAVSAR: L-band HH and HV scene for under-canopy confirmation
  • GPM IMERG: Rainfall accumulations for hydrologic forcing
  • DEM with derived HAND: Topographic susceptibility analysis
  • Road network and shelter inventory: Routing and evacuation planning

Methodological Approach:

Methodologically, we performed calibration and terrain correction, precise co-registration to the DEM, construction of false-color composites (e.g., RGB(VV, VH, VV and VH)) to visually separate urban areas, vegetation, and water, and adaptive thresholding to derive open-water masks.

Key Findings:

Findings (qualitative) highlight floodplain expansion along the Guadalupe, with L-band revealing water spreading beneath riparian canopies near key access points. In the urban core, bright corridors such as downtown blocks and parking lots often remain passable longer than low-lying approaches, while certain routes and bridge adjacencies show early flooding and require rerouting. Cross-polarized backscatter changes align with high-susceptibility zones and with recent heavy rainfall prior to peak flooding.

Data sources

  • Sentinel-1 (C-band, VV/VH) from ASF DAAC: scene IDs [S1_ID_t0] [S1_ID_t1] [S1_ID_t2]
  • UAVSAR (L-band, HH/HV) from JPL: [UAVSAR_ID_t1]
  • NASA GPM IMERG rainfall accumulations for the selected window
  • DEM & HAND (source and resolution specified)
  • Road network & shelters (source attribution; assumptions on capacity/availability)

Evaluation

Metrics:

  • IoU
    Overlap between predicted and reference water extent
  • Precision and Recall
    Water-class accuracy vs. vetted labels
  • Lead time (min)
    IMERG-triggered forecast time to SAR-confirmed extent
  • Differential exposure (km)
    Flooded road-km removed from suggested routes vs. a baseline router
  • Flooded road (km)
    Absolute length of roadway excluded by the water mask

Limitations and next steps

Known limitations include layover and shadow in steep terrain; mixed shoreline pixels that yield partial-water signatures; and revisit gaps that limit temporal resolution and alert latency. Assumptions about shelter radii or availability can propagate into routing; bridges and causeways may be overtopped without distinctive radar signatures at sensor resolution. Route guidance should complement, not replace, on-the-ground reports and official closures.

Next steps:

  • Automated scene-pair selection (pre, peak, post) and locally adaptive thresholds
  • Probabilistic water masks with pixel-wise uncertainty
  • Expanded L-band (and assessed X-band) contributions
  • Integration of agency and crowd reports to update edge weights in real time
  • Human-in-the-loop QA tools for rapid incident review

Closing synthesis

DryPath bridges radar physics and life-saving choices by converting multi-polarization, multi-frequency SAR and rainfall inputs into clear, address-level guidance. It is timely (works through weather and night), physically interpretable (wavelengths and polarizations map to distinct surface processes), and operational (pruned road networks and per-address alert tiers that residents and agencies can act on immediately). Where conventional maps leave people guessing, DryPath aims to deliver a simple, defensible answer at the scale that saves lives: "This is your level of evacuation; this route is safe; go now."