London NO2 Pollution Analysis: Earth Observation and Machine Learning Practice in the AI4EO Course

muimui9's London NO2 Pollution Analysis Project is a practical assignment for UCL's GEOL0069 course "Artificial Intelligence for Earth Observation (AI4EO)". It demonstrates how to use satellite remote sensing data (e.g., Sentinel-5P TROPOMI) and machine learning techniques to monitor urban air quality, reflecting the application value of the intersection between Earth observation and artificial intelligence.

Concepts and Challenges of AI4EO

Traditional Earth observation relies on expert visual interpretation and simple statistical models, which are limited in efficiency. AI4EO applies AI technologies such as deep learning and computer vision to remote sensing data analysis to achieve automated monitoring, but faces challenges like multispectral data processing, time-series analysis, multi-source fusion, and small-sample learning.

Scientific Background of NO2 Pollution Monitoring

NO2 is a major air pollutant derived from fossil fuel combustion and is associated with respiratory diseases. The TROPOMI instrument on the Sentinel-5P satellite can measure NO2 column concentrations with daily global coverage, filling the gap of insufficient spatial coverage of ground monitoring stations.

The core data of the project comes from the TROPOMI instrument on the European Space Agency's Sentinel-5P satellite, which provides daily measurements of tropospheric NO2 column concentrations with a spatial resolution of approximately 5.5×3.5 km.

Advantages of satellite data: wide coverage, high update frequency, and no dependence on ground infrastructure; Limitations: limited spatial resolution, cloud interference, only measuring column concentrations rather than ground-level concentrations, often requiring combination with ground data or use for trend analysis.

The project uses machine learning to process NO2 data, including:

• Data preprocessing: Cloud masking, quality control filtering, spatial resampling, time-series synthesis (e.g., monthly averages);
• Spatial analysis: Identifying pollution hotspots, analyzing distribution patterns, correlating with urban features (road networks, industrial areas);
• Time-series analysis: Detecting pollution trends, identifying seasonal patterns, anomaly event detection;
• Predictive modeling: Predicting NO2 concentrations based on meteorological/traffic data, spatial downscaling (mapping coarse resolution to high-resolution grids).

As a megacity, London faces air quality challenges, and policies like the Ultra Low Emission Zone (ULEZ) provide natural experimental scenarios. Analyzing London's NO2 data through AI4EO methods can quantify the spatial heterogeneity of pollution, evaluate the environmental effects of traffic policies, identify exposure differences among vulnerable groups, and support urban planning decisions.

Typical AI4EO project toolchain:

• Data acquisition: Copernicus Open Access Hub, Google Earth Engine;
• Data processing: Python ecosystem (xarray, rasterio for raster processing, pandas for time-series processing);
• Machine learning: scikit-learn (traditional ML), PyTorch/TensorFlow (deep learning);
• Visualization: matplotlib, seaborn, geopandas (map visualization).

Educational Significance

Course assignments cultivate interdisciplinary thinking (environmental science + data science), practical problem-solving skills, and scientific communication skills. They require completing the full process from data acquisition to visualization, simulating real research workflows.

Future Directions

Cutting-edge directions in the AI4EO field:

• Foundation models: Remote sensing pre-trained large models (e.g., Prithvi, SatMAE);
• Multimodal fusion: Combining optical, SAR, LiDAR, and ground/social media data;
• Causal inference: Evaluating the real effects of policy interventions;
• Real-time monitoring: Edge computing + satellite data streams for near-real-time early warning.