Inspiration
The inspiration was mostly the result of wanting to work on either tracks 3 or 4, because we thought that those best aligned to our interests and also probably had the most open sourced training data. After brainstorming through several ideas and weighing pros and cons, we decided to land on the idea of some sort of AQI map, as we'd all seen them before and we were curious as to how they actually worked. We wanted to work towards trustworthiness within the AQI map by leveraging AI.
What it does
Performs like a normal AQI map, but instead of glossing over gaps between sensors and interpolating values, it accurately provides a range of potential values and dynamically adjusts it's readings every hour based on how those ranges fall within EPA standards.
How we built it
We built everything in Python. Every hour, a collector pulls live PM2.5 data from three sources (PurpleAir, NYC Health, EPA AirNow); 95 sensors, from around the city. It drops faulty ones and saves every raw timestamp and location for tracing. A Gaussian process model then uses these sensors to estimate air quality across the city like a regular AQI map within a range of 95%, but our model goes further and makes sure it's right and knows when it's wrong. We control posted AQI ranges for each square half kilometer of land by hiding one of the sensors' and surrounding ones' data and having the model guess it blind, over and over again per hour. We ended up with 92.7% accuracy, so once we widened the range up 1.2x, we hit 95% A placement algorithm then decides where to recommend placing sensors next to shrink uncertainty, weighted by asthmatic population density. We refuse to label a block's AQ if ranges fall between values.
Challenges we ran into
Messy sensors. Cheap PurpleAir sensors often read high compared to the official ones, and we had to calibrate accordingly. Our first blind test didn't quite hit 95% accuracy, so we had to widen the ranges to reach it. Our data collection script also randomly crashed at 3:24am and Naman had a panic attack. Midtown West was a direct miss of our model due to it not knowing there's a bus station there, but we leveraged it into a learning experience and plans for the future.
Accomplishments that we're proud of
We think that the data collected and also what we discovered through our process of predicting, masking, and grading helped uncover blind spots in the NYC sensor data. We're proud of the fact that our project has real-world grounding and impact, and highlights our passion for responsible, safe AI.
What we learned
We learned a lot about predictive models, but also machine learning processes. We weren't aware of the vast amount of available data in the form of open-sourced .csv's or the ever-changing online datasets accessed through a free API key. Learning about the gaussian process was also very interesting, as we were able to construct a model with a relatively small training set and actually have it perform well.
What's next for SensorGap
AirNow actually has a lot of AQI/PM2.5 data across the country and even the globe. This gives us an exciting opportunity to collect and train a model with a lot more data to create bigger ranges of live AQI coverage, and especially in underdeveloped or heavily polluted areas, we can find the many areas without sensors, where these gaps in sensors could actually result in non-represented points that are incredibly varying in magnitude. Helping determine where exactly to add sensors for optimal coverage/estimation also gives a very cost-effective solution. We plan to add more context for the model to avoid more Midtown West events, such as traffic routes, bus stations, airports, zones with higher pedestrian frequency, and tracts with higher percentages of people with different types of respiratory ailments.
Built With
- folium
- geopandas
- matplotlib
- numpy
- pandas
- scikit-learn
- streamlit
- streamlit-folium
