When I was a kid, I spent an unreasonable amount of time looking at maps.

Google Maps was probably the main one. Google Earth was cooler, obviously, but my computer had an Intel i3 and 8GB of RAM, so it would turn into a slideshow pretty quickly. I’d zoom into random places, follow highways for no reason, then zoom back out and wonder why everything was where it was.
During COVID, that turned into the game GeoGuessr. My parents probably thought I was just scrolling through maps like normal. Technically they weren’t wrong.
Whenever my family traveled, I’d do the same thing. I’d wonder why a highway took one route instead of another, why a city developed where it did, or why a road followed a river instead of cutting through a mountain. Even when reading history, I’d end up staring at maps. The maps in my Bible are probably more worn out than some of the actual pages.
I didn’t have a name for what I was interested in yet. I just knew I liked seeing things from above.


A road becomes a transportation network. A river becomes a watershed. A neighborhood becomes part of a larger pattern. A disaster becomes a relationship between geography, infrastructure, people, and time.
One of my first real introductions to academic geography came through environmental studies, where I was introduced to ArcGIS. For the first time, I wasn’t just looking at a map someone else had made. I could build one myself, layering roads, elevation, population, land use, hazards, and infrastructure to ask questions that a normal map couldn’t answer.
I eventually competed in the national ArcGIS StoryMaps competition, and AP Environmental Science pushed me to think about environmental problems spatially. Pollution wasn’t just a chemical process. Where was it happening? What was around it? How did water move? How did land use change the risks people faced?
Geography was becoming more than something I enjoyed looking at. It was becoming a tool.
The Word “Geospatial”
When it came time to choose my Technology Student Association events, I wasn’t particularly interested in another computer science event. Then I saw Geospatial Technology and had no idea what it meant, so naturally I Googled it. Five minutes later I was down a rabbit hole.
It was basically everything I’d already been interested in: maps, geography, environmental science, programming, and data analysis. For TSA, I built Geospat, a project focused on earthquake risk across Seattle, using historical earthquake data, the ETAS model for aftershock probabilities, simulated M7.0 shaking, and vulnerable unreinforced masonry buildings.

What hooked me was seeing what happened when all those layers came together. A fault map tells you where faults are, a population map tells you where people live, and a building inventory tells you what’s there. Together, they start telling you where a disaster could actually hit hardest.
Geospatial Technology 2025 Portfolio.pdf
I ended up placing first in Geospatial Technology at Washington TSA and qualifying for nationals, and I knew I wanted to keep going deeper into geospatial work.
Going Deeper
After TSA, I kept getting deeper into geospatial work, messing around with GeoPandas and Python and learning more about GIS and remote sensing. I started getting curious about what happened before the map: how satellites collect imagery, how it gets processed, and how it becomes something useful on the ground.
Then, the summer before my junior year, I found MIT Beaver Works Summer Institute. I hadn’t planned my summer that well and wasn’t really interested in doing research just to have it on my resume, but Remote Sensing for Disaster Response caught my attention. It combined satellite imagery, GIS, Python, image processing, machine learning, and actual disaster scenarios, which felt like a pretty natural next step.
Beaver Works Summer Institute @ MIT - Where we teach tech before the textbooks
Looking back, it’s kinda funny how everything connected. Google Maps got me obsessed with looking at places from above, ArcGIS showed me I could actually build and analyze those maps, environmental science got me thinking more about how geography affects real problems, and TSA gave me a way to use geospatial data to model earthquake risk. BWSI felt like the next step in all of that.
Then I Got into BWSI
A month after my surprising acceptance letter, I was sitting in BWSI’s Remote Sensing for Disaster Response course, and pretty quickly I realized how much bigger the field was than I had understood from making maps.
The first few weeks
were spent building up the tools we would eventually need for the final project. We started with basic geometry using Shapely, creating points, lines, and polygons and learning how operations like buffers, intersections, and unions actually work. From there we moved into GeoPandas and coordinate reference systems, then raster data and satellite imagery, spatial joins, Google Earth Engine, and image processing. At first these felt like separate topics, but as the program went on, they started fitting together into one way of thinking about geographic data.

One of the biggest shifts for me was realizing that a map isn’t really just a picture. A vector dataset can represent something like a coastline as geometry, while a raster can store a value at every pixel, whether that’s elevation, reflectance, or radar backscatter. That distinction became important later when we started working with actual satellite imagery and Digital Elevation Models. We also did a spatial analysis of the 1938 New England hurricane where we combined wind hazard with the CDC’s Social Vulnerability Index. That showed me something I would keep coming back to during the final project: knowing where a hazard is isn’t enough. You also need to know what and who that hazard overlaps with. A neighborhood experiencing the same wind or flooding as another can face a completely different outcome depending on things like infrastructure and social vulnerability.
import geopandas as gpd
import numpy as np
from matplotlib import pyplot as plt
!unzip -o WindField.zip -d windfield
# load windfield into gpd geodataframe
windfield_df = gpd.read_file('windfield/WindField.shp')
rows, columns = windfield_df.shape
# load SVI2018 from course shared folder
!wget https://files.bwsi-remote-sensing.net/data/SVI2020_US_tract.gdb.zip
svi_df = gpd.read_file('SVI2020_US_tract.gdb.zip')
svi_df
# overall vulnerability score is in RPL_THEMES; -999 = missing data
svi_df['RPL_THEMES']
# replace missing data with None values
svi_df['RPL_THEMES'] = svi_df['RPL_THEMES'].replace({-999: np.nan})
# take a look at overall SVI (takes a minute to run because it's the whole country)
fig = plt.figure(figsize=(20,10))
ax = fig.add_subplot(1,1,1)
svi_df.dropna(subset=['RPL_THEMES']).plot('RPL_THEMES', legend=True, ax=ax)
ax.set_xlim([-180,-60])
# spatial join — merge dataframe rows based on how their geometries interact
joined_df = gpd.sjoin(svi_df, windfield_df, how='inner', predicate='intersects')
joined_df
# joined dataframe only includes the intersection of the two datasets
# at a census-tract level
joined_df.plot()
# compute wind-vulnerability score: Risk = Hazard × Vulnerability
joined_df['wind_risk'] = joined_df['RPL_THEMES'] * joined_df['GRIDCODE']
joined_df.plot('wind_risk', legend=True)
We eventually started working with Google Earth Engine, and for that module I pulled data from the 2018 Kerala floods — a dataset that felt personal given my connection to India. Using Sentinel-1 SAR imagery, I compared a radar scan of the region before the flood against one taken after, looking at changes in backscatter — the strength of the radar signal bouncing back off the surface. Calm water reflects almost nothing, while land reflects much more, so a sharp drop in backscatter between the two images signals that dry land had become standing water. From there I cleaned up the raw flood mask, filtering out isolated noise pixels and masking out steep terrain where flooding wasn’t physically plausible, to get an accurate map of exactly where the flooding occurred. It was probably the first time in the program I really felt the connection between something I’d cared about for years and the technical mechanics of remote sensing.
All of that built toward our final project, a month-long disaster response simulation centered around a fictional hurricane that our class named
Hurricane Claude.

We split into different teams covering planning, logistics, operations, and public information, basically mirroring the different kinds of work that would happen during an actual emergency response. I was on the Planning Team, where I took ownership of flood mapping. My job was to take the hurricane forecast and turn it into a map that could answer a much more useful question than simply “where is the storm?” We needed to figure out where storm surge could reach, which areas would actually flood once elevation was considered, which communities were most vulnerable, and even which shelters might themselves be affected.
Every simulation day, we received a new forecast containing 20 ensemble members, each representing a possible path for the hurricane. I averaged the positions of those ensemble members at each forecast hour and connected them into a central forecast path. We were then given three surge thresholds based on distance from the hurricane: up to six meters within 80 kilometers of the path, four meters from 80 to 160 kilometers, and three meters from 160 to 300 kilometers. I used buffers and difference operations to turn those distances into clean surge zones, which gave us a first approximation of where the storm could potentially push water.
The problem was that a surge zone isn’t the same thing as a flood map. Storm surge is ocean water, so it obviously can’t just appear indefinitely inland. I needed to figure out where that water could physically reach the Massachusetts coastline, which meant working with much more detailed coastline data than I had initially used. When I buffered the Massachusetts coastline by five kilometers, though, the result looked completely wrong. The flooded area seemed to extend much farther than five kilometers, especially around Boston Harbor and the islands. I initially assumed my code was broken, so I went back through the projection and geometry, confirmed that everything was correctly reprojected into UTM Zone 19N where the coordinates were measured in meters, and independently checked the five kilometer distance using an outside mapping tool.
The buffer was correct. The problem was my mental model of what a five kilometer buffer should look like. Massachusetts has an incredibly complicated coastline, with harbors, river inlets, islands, and narrow pieces of land everywhere. A uniform buffer follows all of those shapes, meaning that the buffer coming in from opposite sides of a narrow inlet can overlap and visually merge into what looks like one enormous flooded area. It wasn’t a bug in the code. The geometry was doing exactly what it was supposed to do, and I simply hadn’t accounted for how much the actual geography would affect what five kilometers looked like on a map. That was probably one of my favorite debugging moments of the program because I had to stop looking at the code and zoom out to understand the geography that the code was actually representing.


Once the coastline made sense, I brought in a Digital Elevation Model and built a simple bathtub flood model, comparing surge height against elevation pixel by pixel to separate areas that could actually flood from areas that were only theoretically exposed. The result concentrated flooding in low-lying areas around Boston Harbor, East Boston, Chelsea, Quincy, and the Charles and Mystic river basins. I then spatially joined the CDC’s Social Vulnerability Index to the flood zones, and Chelsea and East Boston consistently stood out where high flood exposure and community vulnerability overlapped.
That was where the whole idea of zooming out came back for me. I had started the project asking where the hurricane was going, then had to zoom out and think about what its distance from the coastline actually meant, then zoom out again and account for elevation, and then zoom out even further and ask who was actually living in those areas. The useful answer wasn’t simply “this place will flood.” It was something closer to “this place is likely to flood, and these are the people and communities that may be most affected.”

That felt like the point where remote sensing stopped being just about mapping a physical event and started becoming a way to make better decisions about it.
The flood mapping was my main responsibility, but I also helped with a different problem:
tracking a ship caught inside Hurricane Claude.
Optical imagery is useless in a storm — total cloud cover — so we used Synthetic Aperture Radar instead, which measures radio wave backscatter and works through clouds and at night. We had 83 SAR scenes, each with land-mask and dual-polarization backscatter layers, and were hunting for the ship as a bright, isolated radar return against darker open water. Our first attempts produced heavy false positives, mostly from complex coastlines reflecting radar just as strongly as metal does. We fixed this by indexing all 83 scenes chronologically by timestamp, turning an unordered pile of imagery into an actual searchable track.

Even after that, the detections we plotted — eight timestamped positions across the storm — produced a route that zigzagged in ways no real ship would travel. That was the real lesson: an algorithm returning a result doesn’t mean it returned the correct result, and we hadn’t actually proven we’d found the ship. That idea — that a fast wrong answer is worse than a slow correct one — came up constantly in the disaster response guest talks, and seeing it play out in our own detection pipeline made it land differently than just hearing it secondhand.


The fix was combining the two SAR polarization layers, VH and VV, and masking out land after resizing the land mask to match the imagery. That cut down the huge number of coastline false positives we were getting. We then tracked bright pixels, grouped them into objects, filtered out noise by size, and converted the surviving coordinates into real latitude and longitude. That finally gave us a ship track that actually made sense. The tanker moved from Hour 1 to Hour 4, then stayed mostly stationary through Hour 6, well offshore and away from populated areas. Based on that, we recommended no evacuation and focused resources elsewhere.

We closed the presentation on that point: the interesting part wasn’t that we found the ship, it’s that we could finally trust that we had — which took throwing out our first two “working” detection pipelines to get there.
Guest Speakers
The guest speakers and seminar series were easily one of my favorite parts of BWSI. We heard from Dr. Jerry Wohletz at Draper, Dr. Melissa Choi at MIT Lincoln Laboratory, Dr. Devon Maloney at FEMA, Maj. Scott Kaplan from Civil Air Patrol and NGA, Dr. Raymond Richards from RTX BBN Technologies, Neale Batra from Applied Epi, April Mills from Engine for Change, and Dr. Monique Kuglitsch from Fraunhofer HHI.

Dr. Jerry Wohletz’s talk at Draper probably stuck with me the most. He talked about leadership and how the people you lead can roughly fall into three groups: a third who will support you and listen to basically anything you say, a third who are skeptical and need to be engaged, and a third who just won’t like you no matter what. There’s no point wasting all your energy trying to convince that last group. The important part is finding the middle, listening to them, and keeping the team moving. That was a pretty different lesson from everything else we were learning technically, and honestly one I’ll probably remember for a lot longer. Dr. Melissa Choi completely changed how I saw MIT Lincoln Laboratory. I knew it was an MIT lab, but hearing how deeply its work connects to U.S. defense, national security, sensing, satellites, and autonomous systems was honestly crazy. Dr. Devon Maloney from FEMA showed us how satellite, aircraft, and drone imagery actually come together during real disasters, which made our Hurricane Claude simulation feel much more real.
By the end of the month, I had picked up a ton of technical skills, but honestly the bigger thing I took away was how much changing your perspective can change the problem. When my flood map looked wrong, I had to stop staring at my code and look at the geography. When I modeled flooding, I had to look past the water and think about who was actually there. When we had 20 hurricane forecasts, I had to think about uncertainty instead of treating one path like it was guaranteed. Even with the ship, we had to step back from individual detections and ask if the whole story actually made sense.
It was pretty much the same thing I’d been doing with maps since I was a kid, just with a lot more data, code, and actual problems to solve. I’ve been zooming in and out for years. BWSI finally gave me a reason to do something with it.