DeepAir-Dhaka: Spatiotemporal Air Quality Forecasting 🌍💨

DeepAir-Dhaka is a machine learning-driven system designed to provide granular, neighborhood-level air quality forecasts for Dhaka, Bangladesh. Moving beyond simple city-wide averages derived from sparse physical monitoring stations, this project utilizes spatiotemporal deep learning (ConvLSTM U-Net) to visualize and predict the diffusion of PM2.5 and other pollutants as dynamic, high-resolution heatmaps.

---

🌟 The Problem & Our Solution

Dhaka consistently ranks as one of the most polluted cities globally. However, existing monitoring systems are extremely sparse (only 9 major Continuous Ambient Air Quality Monitoring System (CAMS) stations), leaving many neighborhoods without accurate, hyper-local data.

DeepAir-Dhaka bridges this gap by:

1. Constructing "Virtual Sensors": Using mathematical Inverse Distance Weighting (IDW) interpolation and OpenStreetMap (OSM) spatial priors (roads, rivers, and industrial zones) to estimate pollution in unmonitored zones, turning 9 stations into 1,024 virtual sensors (32x32 grid).
2. Spatiotemporal Forecasting: Employing a ConvLSTM U-Net network equipped with Attention Gates to learn exactly how pollution "clouds" travel across the city over time.
3. Dynamic Visualization: Generating high-resolution heatmaps for real-time tracking and up to 7-day future air quality insights.

---

🧠 Methodology & Architecture

1. Data Collection & Preprocessing

• Sources: 9 CAMS Stations in Dhaka & Open-Meteo API.
• Timeframe: 2022 → 2026 (38,304 hours of continuous data).
• Interpolation: IDW translates sparse 9-point data into a dense 32x32 geographic grid.
• LUR Multiplier (Land Use Regression): Modifies grid values based on real-world priors:
  • Industrial zones (Gazipur, Narayanganj): +50% pollution weight
  • Major Highways/Roads: +20% pollution weight
  • Water Bodies (Buriganga, Turag rivers): -30% pollution weight

2. Feature Engineering (24-Channel Tensor)

Each cell in the 32x32 grid carries 24 distinct channels:

• Pollutants: PM2.5, PM10, CO, NO2, O3, Dust, AOD.
• Meteorology: Temperature, Humidity, Rain, Pressure, Cloud Cover, Radiation, Wind U/V vectors.
• Temporal Encodings: Cyclical sin/cos encodings for Hour, Day, and Month.
• Spatial Priors: Static maps for Industrial, Water, and Road densities.

3. Deep Learning Architecture (ConvLSTM U-Net)

• Input: A 24-hour historical window of the 32x32x24 tensor.
• Encoder: ConvLSTM2D (64 filters) extracts spatial patterns for each hour, followed by a ConvLSTM2D (128 filters) layer that compresses the 24 hours into a singular spatial representation.
• Attention Gate: Informs the decoder where to focus (heavily weighting high-pollution zones).
• Decoder: Refines the spatial prediction through Concatenation, Conv2D(128), and Conv2D(64) layers.
• Custom Weighted Huber Loss: The model is trained to heavily penalize errors during dangerous pollution spikes (PM2.5 > 107 µg/m³ gets a 3× weight penalty) ensuring the AI predicts extreme events rather than safe "averages".

---

📈 Results and Visualizations

The pipeline incorporates a robust ConvLSTM architecture, exceeding all initial project proposals:

• Correlation Coefficient (R): 0.9729
• Mean Absolute Error (MAE): 3.39 µg/m³ (±6% of the mean)
• Reliability: 96.9% of all predictions fall within ±10 µg/m³.

Visual Explanations

1. Project Pipeline
(From raw CAMS data → Preprocessing → Feature Engineering → ConvLSTM Training → Forecast)

2. Methodology Deep-Dive
(Detailed look at IDW Interpolation, Tensor formulation, Loss function, and Architecture)

---

🚀 Getting Started

Prerequisites

• Python 3.8+
• TensorFlow / Keras 2.0+
• pandas, numpy, requests, scikit-learn, matplotlib

1. Data Generation

To fetch live data, process it, and build the 32x32 spatiotemporal tensor maps, run:

python "Scipts/createDeepAir datasetinNPY.py"

This script will:

1. Fetch hourly data from the CAMS stations.
2. Clean and standardize the data with physical weather vectors.
3. Perform IDW Interpolation to generate the DeepAir_Dhaka_32x32_21_Channels.npy tensor.

2. Diagram Generation

To dynamically generate the custom visualization diagrams detailing the methodology:

python pipelineDarkBackground.py

This will output pipeline_diagram.png and methodology_diagram.png in the root directory.

---

📂 Project Structure

DeepAir/
│
├── Scipts/ 
│   └── createDeepAir datasetinNPY.py    # Core pipeline for data collection & preprocessing
│
├── results/                             # Contains evaluation metrics, GIFs, and plots (results.zip excluded from Git)
│   ├── epoch_visuals/                   # Heatmaps generated per epoch
│   ├── forecast_24h.gif                 # Animated 24h spatial PM2.5 forecast
│   └── forecast_7day.gif                # Animated 7-day spatial PM2.5 forecast
│
├── figures/                             # Automatically generated visualizations
├── Docs/                                # Project proposal and technical documentation
├── scratch/                             # Experimental scripts and temporary logs
│
├── DeepAir_Dhaka_Final.keras            # The fully trained ConvLSTM-U-Net model (Load via tf.keras.models.load_model)
├── pipelineDarkBackground.py            # Python script using Matplotlib to generate architectural diagrams
│
├── pipeline_diagram.png                 # Exported pipeline visualization
└── methodology_diagram.png              # Exported methodology/architecture visualization

---

📜 References & Acknowledgements

1. Islam, M. S., et al. "Air Quality Prediction in Dhaka City using Deep Learning Approaches." IEEE, 2023.
2. Shi, X., et al. "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting." NIPS, 2015.

This project contributes to the field of AI for Social Good, offering a scalable solution for cities with limited sensor infrastructure by leveraging Spatiotemporal AI.