PythonGeoPandasShapelyDEM AnalysisFoliumCDC SVISARRemote Sensing
Days 3 and 4: the shaded bands are where surge could reach, the blue is where the land is actually low enough to flood. Hurricane Claude was a fictional simulation.
Role
Planning: flood mappingOperations support: imagery analysis and SAR ship tracking
Team
MIT BWSI Remote Sensing for Disaster ResponseMulti-team emergency-response simulation
Timeline
July–August 2026Four simulated forecast days
Tools
Python, GeoPandas, Shapely, Folium30 m DEM, UTM 19N, CDC SVI, SAR backscatter
Outcome
4,390 km² exposed, 370 km² modeled floodingDay 4, displayed study area

For BWSI's final project, our class ran Hurricane Claude, a fictional hurricane-response simulation across New England. Different teams handled planning, logistics, operations, and public information. I was on Planning, where my main responsibility was flood mapping. I also helped with imagery analysis on the Operations side, including deep-learning work and tracking a ship using synthetic aperture radar (SAR).

The question I kept coming back to was fairly simple: if we know where the storm might go, how do we figure out what actually needs attention on the ground?

A forecast track can tell you where the center of a hurricane might travel. It does not tell you which streets flood, whether high ground stays dry, or which neighborhoods will have the hardest time recovering. Even a storm-surge boundary can be misleading if you treat every bit of land inside it as underwater.

The ensemble forecast on each of the four simulation days. By Day 4 the storm had moved largely offshore, and the planning maps had to move with it.

Our simulation supplied changing forecasts over four days. My job was to make sense of that uncertainty and translate it into a geographic assessment that could inform the rest of the response.

  • Each daily forecast contained 20 ensemble members, not one certain storm path.
  • Surge was approximated with distance bands around the central forecast path: 6 m within 80 km, 4 m between 80–160 km, and 3 m between 160–300 km.
  • The Massachusetts coast has harbors, islands, river inlets, and narrow peninsulas. A simple inland buffer can look much more expansive than you expect.
  • The model needed to distinguish potential exposure from elevation-based flooding and then consider the communities living inside those areas.
  • This was a time-limited simulation. The outputs were screening estimates, not a hydrodynamic forecast or a real evacuation order.

Start with the ensemble, then build the surge zones

I averaged the positions of 20 forecast members at each forecast hour and connected them into a central path. Using Shapely and GeoPandas, I built the three surge-distance bands with buffer and difference operations. This gave us a consistent way to compare the storm's changing reach from day to day, while keeping in mind that the average line was only a summary of many possible tracks.

Day 2 surge bands around the central forecast path: 6 m within 80 km, 4 m out to 160 km, 3 m out to 300 km.

Debug the coastline before trusting the flood map

A five-kilometer buffer around the Massachusetts coastline looked wildly too large, especially near Boston Harbor. I checked my projection, reprojected the geometry into UTM Zone 19N so the distances were in meters, and verified the buffer independently. The math was right. The geography was what I had misunderstood: buffers from opposite sides of narrow inlets and coastlines were merging on the map.

Day 1: the 5 km coastal zones where surge could reach land, before elevation was applied. Everything shaded here is exposed, not necessarily flooded.

That issue made it pretty obvious why a surge zone could not be my final output. I brought in a 30 m digital elevation model and used a simple bathtub-style comparison between surge height and terrain elevation to estimate where flooding could occur. It was a rough model with clear limits, but substantially more useful than coloring every exposed coastal area blue.

Put people on top of the hazard map

I overlaid the CDC Social Vulnerability Index (SVI) with the hazard and flood layers. Chelsea and East Boston repeatedly stood out: the maps indicated significant coastal exposure overlapping with communities that could face additional barriers during a disaster. The point was not to rank people by a single index; it was to make sure our physical hazard map wasn't the only thing guiding attention.

Social vulnerability inside the 6 m and 4 m surge zones around Boston on Day 2. Darker red means higher vulnerability.

Simulation daySurge-exposed areaModeled floodingFlooded share
Day 35,845 km²629 km²11%
Day 44,390 km²370 km²8%

These are modeled values from the Planning Team's displayed map extent, not observed hurricane impacts. In the resulting maps, flood risk concentrated in low-lying parts of the Boston Harbor area, including East Boston, Chelsea, Quincy, and the Charles and Mystic river basins. The team also used hazard, population, vulnerability, and shelter information in its broader evacuation-planning work.

Student-team simulation outputs, not measurements from a real hurricane or instructions for a real evacuation.

Planning was my main contribution, but I also helped on Operations. That team worked on disaster imagery, including ResNet-50/ResNet-152 damage classification, U-Net segmentation, and Grad-CAM visual explanations with the RescueNet and xView2 datasets. I contributed to this side of the work, although the models and reported performance figures were team results, not models I am claiming to have built alone.

Damage classification on RescueNet aerial imagery. The top row is correct; the bottom row mistakes medium damage for superficial. Operations Team results.
Grad-CAM heatmaps: red marks the part of each image that most influenced the damage label, which let the team check the model was looking at buildings and not the background.

I also helped investigate a ship caught in Hurricane Claude using radar imagery. The team sorted 83 SAR scenes by timestamp and worked with VH/VV polarization data and land masking to reduce bright coastline false positives. Early detections made an implausible zigzagging track. After filtering and geolocating candidate detections, the eight mapped positions formed a more coherent offshore path. The resulting assessment supported leaving evacuation resources focused on higher-priority areas rather than treating every bright radar pixel as a ship.

The coastline bug is the part I still remember best. I spent time questioning the code, only to find that the five-kilometer buffer was correct and my picture of the coastline was wrong. A map can be mathematically valid and still be interpreted badly.

The same problem showed up in the SAR work. A detection pipeline can produce clean-looking coordinates that are completely implausible once plotted. I came out of the project much more skeptical of outputs that look convincing without being checked against the geography and the actual problem.

A shorter walkthrough of the same project: the flood-mapping sequence first, then the Operations work.

Loading slides…