← All cases

AI Empowers · Environment

Predicting air quality from satellite images

Is deep learning on satellite images worth it to predict CO₂, O₃ and NOₓ? We tested it.

  • Environment
  • Computer vision
  • Predictive analytics
ClientEuropean Environment Agency (with Microsoft)
SectorEnvironment
CategoryAI Empowers
StatusStudy completed
≈ sameaccuracy as deep learning, far lower cost

The problem

The agency wanted to predict CO₂, O₃ and NOₓ levels and asked whether deep learning on satellite imagery was worth the cost.

The solution

We compared a CNN on satellite images with classic machine learning on metadata plus heuristics, tracked in MLflow on Databricks. Axel also gave a five-day Databricks training.

The outcome

Traditional ML matched the CNN's accuracy at a fraction of the cost. The right model is not always the biggest one.

The full story

  1. Two approaches

    A CNN built with Keras learned from satellite images. A classic model used structured metadata, such as sensor readings, weather and geography, plus domain heuristics.

  2. What we found

    The metadata model performed about as well. The CNN picked up spatial patterns and gained a little, but needed far more data, compute and training time, and is harder to explain.

  3. Into production

    Experiments, tuning and model comparison ran in MLflow on Databricks. The best model was deployed on Azure Machine Learning.

  4. The lesson

    Before you choose deep learning, weigh data readiness, cost against benefit and the need to explain results to regulators. Classic machine learning with strong domain knowledge is often enough.

Tech

Keras CNNMLflowDatabricksAzure ML
Facing something similar?We would shape your version around your own data.

Ask Pai

Ask Pai about this case

Pai is an AI. It only knows what is written on this page.

Baby Pai · AIHi, I'm Baby Pai, an AI assistant. Ask me about the European Environment Agency (with Microsoft) case.

Pai is an AI. Answers come from this case page. Don't share sensitive data.

Chats are kept for 30 days unless you switch that off in Privacy and choices.