Field: Project Title
Paryavaran — Delhi Greenspace Suitability Platform
Field: Elevator Pitch / Tagline (1-2 sentences)
An AI-driven geospatial decision platform that synthesizes satellite thermal imagery, real-time pollution sensors, and demographic density into an actionable micro-forest suitability index for Delhi (NCT).
Field: Inspiration
Over 56% of humanity lives in urban areas today, projected to reach 70% by 2050. Few places on Earth illustrate the acute tensions of hyper-urbanization more starkly than Delhi (NCT). Every winter, hazardous smog episodes drive PM2.5 and PM10 concentrations 15 to 25 times above WHO safe limits. In summer, severe Urban Heat Island (UHI) effects push localized surface temperatures past 45°C.
Scientific research demonstrates that strategically planted urban greenspaces, tree corridors, and Miyawaki micro-forests lower local ambient temperatures by 3°C to 5°C, filter particulate matter, and restore community health. However, municipal authorities lack unified, high-resolution spatial decision-support tools. Planners face a critical question: Where should new green infrastructure be prioritized to maximize human health without conflicting with critical civic infrastructure?
That question inspired Paryavaran (the Sanskrit/Hindi word for Environment)—a spatial intelligence engine that turns satellite, demographic, sensor, and infrastructural data into an actionable greenspace suitability score.
Field: What it does
Paryavaran evaluates 42,300 spatial micro-zones across the National Capital Territory of Delhi on a 250-meter equidistant grid. The platform continuously processes:
1. Demographic Exposure: High-resolution spatial population density from WorldPop 2024.
2. Air Pollution Deficit: Spatial interpolation of continuous ground-level telemetry from CPCB monitoring stations (PM2.5, PM10, NO₂).
3. Surface Heat Stress: Summer daytime Land Surface Temperature (LST) derived from NASA MODIS thermal bands.
4. Greenspace Deficit: Spatial gaps relative to existing parks and nature reserves via OpenStreetMap and ESA WorldCover (10m).
5. Civic Infrastructure & Feasibility: Physical suitability multipliers that prevent unrealistic planting recommendations on airports, water bodies, hospitals, and railways, while rewarding bare vacant land.
The platform visualizes these insights through a high-performance, interactive dashboard featuring dynamic continuous overlays, 3D terrain pitch, and customizable urban planning layers.
Field: How we built it
1. Geodesic Grid Discretization:
We bounded Delhi NCT (28.40°N to 28.88°N, 76.84°E to 77.35°E) and discretized it into a uniform 250m grid (42,300 cells). Spherical geodesic distances were computed with the Haversine formula:
$$d = 2R \arcsin \left( \sqrt{\sin^2\left(\frac{\Delta \phi}{2}\right) + \cos(\phi_1)\cos(\phi_2)\sin^2\left(\frac{\Delta \lambda}{2}\right)} \right)$$
2. Spatial Air Quality Interpolation:
To map air pollution across every unmonitored grid centroid, we implemented an Inverse Distance-Weighted K-Nearest Neighbors (KNN) model using ground-truth CPCB stations:
$$\hat{Z}(p) = \frac{\sum_{i=1}^{k} w_i Z(s_i)}{\sum_{i=1}^{k} w_i}, \quad w_i = \frac{1}{\text{dist}(p, s_i)^2}$$
3. Multi-Criteria Priority Formulation:
The baseline environmental and demographic demand is modeled as:
$$\text{Raw Priority} = 0.25 \cdot P_{\text{need}} + 0.20 \cdot G_{\text{deficit}} + 0.20 \cdot H_{\text{stress}} + 0.15 \cdot AQ_{\text{score}} + 0.10 \cdot A_{\text{deficit}} + 0.10 \cdot B_{\text{density}}$$
Where P_need is population demand, G_deficit is greenspace deficit, H_stress is MODIS summer heat stress, AQ_score is particulate exposure, A_deficit is distance to the nearest park, and B_density is building footprint density from Google Open Buildings V3.
4. Land Feasibility Multipliers:
To ensure practical urban planning realism, we applied ground-truth feasibility penalties and rewards:
$$\text{Final Priority Score} = \text{Raw Priority} \times \prod_{j} f_j$$
- Airport Runway / Water Body: 0.00 (Hard Exclusion)
- Hospitals / Schools: 0.30 - 0.40 (Structural Conflict Penalty)
- Railway / Metro Transit Corridors: 0.50
- Existing Green Space: 0.50 (Need already satisfied)
- Wetlands: 0.60 (Ecological Buffer)
- Cropland: 0.80 (Agricultural Zoning)
- Neutral Urban Land: 1.00
- Bare / Vacant Land: 1.40 (High-Yield Opportunity Bonus)
5. Cloud Infrastructure on AWS:
- AWS CloudFormation: Declaratively provisioned the cloud infrastructure as code (IaC).
- AWS Systems Manager (SSM) Parameter Store: Dynamically tracks and resolves the latest Amazon Linux 2023 AMI without hardcoded image IDs.
- Amazon EC2 (t3.medium): Hosts the Python/Dash application using an auto-restarting Systemd daemon and Gunicorn production WSGI server bound to port 80.
Field: Challenges we ran into
1. The O(N²) Geospatial Compute Bottleneck:
Intersecting tens of thousands of OpenStreetMap polygons and satellite rasters against 42,300 grid points using standard iterative loops took over 4 hours per run. We refactored the pipeline to use vectorized spatial joins (geopandas.sjoin) and indexed coordinates using Scikit-Learn BallTrees in Haversine metric space, cutting execution time to under 90 seconds.
2. Cloud-Init GDAL Compilation Overhead:
Deploying to Amazon Linux 2023 initially caused cloud-init to freeze and fail with Out-Of-Memory (OOM) errors because heavy spatial packages (rasterio, fiona) were compiling C++ GDAL bindings on the instance. We solved this by decoupling offline spatial data processing from online web serving: spatial joins are pre-computed into final_df.csv, and a dedicated requirements-web.txt installs only lightweight, pre-built Python wheels. Instance boot time dropped from 20+ minutes to under 60 seconds.
3. Silent Breaking Changes in Plotly 7.0:
On our live EC2 instance, the server returned HTTP 200, but the browser rendered an empty Cartesian graph. Deep log inspection revealed Plotly 7.0 had recently removed all scattermapbox traces in favor of MapLibre traces. We pinned plotly==5.24.1 and dash==2.18.2, pre-rendered the initial figure directly in the layout to eliminate AJAX cold starts, and implemented structured exception logging.
Field: Accomplishments that we're proud of
- Sub-Second Geospatial Interactivity: Built a high-performance web dashboard that renders and filters 42,300+ spatial data points with smooth 3D pitch and pan capabilities in real time.
- Realistic Urban Planning Constraints: Unlike simple heatmaps, our model intelligently excludes airports and water bodies while prioritizing vacant, unbuilt urban parcels.
- Production-Grade Cloud Deployment: Achieved a zero-touch, self-healing AWS CloudFormation pipeline that boots, configures systemd, and goes live in under a minute.
Field: What we learned
- Geospatial Precision: Managing coordinate reference system (CRS) projections (EPSG:4326 vs. UTM 43N) and handling spherical distortions taught us the nuances of real-world GIS engineering.
- Infrastructure as Code is Indispensable: Managing our environment declaratively via AWS CloudFormation saved dozens of hours, allowing us to tear down and deploy clean stacks with a single CLI command.
- Domain Constraints Make AI Actionable: A purely statistical clustering model would have prioritized Indira Gandhi International Airport due to surrounding heat and pollution. Applying rule-based feasibility multipliers turned an academic exercise into a practical planning tool.
Field: What's next for Paryavaran
- Multi-City Expansion: Scaling the data ingestion pipeline to other dense South Asian megacities facing similar air pollution and heat crises, including Mumbai, Dhaka, and Lahore.
- Automated Sentinel-2 Satellite Ingestion: Integrating automated quarterly pipelines via European Space Agency (ESA) Sentinel-2 NDVI satellite imagery to monitor vegetation gains over time.
- Public Scenario Simulator: Enabling urban planners and community organizers to draw hypothetical micro-forest boundaries on the map and simulate projected cooling and air purification impacts before breaking ground.
Field: Built With (Tags / Technologies)
python, dash, plotly, pandas, geopandas, scikit-learn, scipy, rasterio, gunicorn, systemd, aws-ec2, aws-cloudformation, aws-ssm, amazon-linux-2023, openstreetmap, worldpop, cpcb-data, nasa-modis, esa-worldcover
Field: Try It Out (Links)
- Live Demo Link:
http://ec2-3-90-145-247.compute-1.amazonaws.com
- GitHub Repository:
https://github.com/uselessdevloper/paryavaran
Log in or sign up for Devpost to join the conversation.