At NILU, machine learning is used on its own or combined with physics-based models, depending on what the problem requires.
Good models need good data, and environmental data is rarely perfect:
- reference instruments are few
- low-cost sensors are noisy
- satellite coverage is uneven
Calibration, correction and data fusion are therefore a core part of our modelling, including our work with satellite data and Copernicus land monitoring.
Our models do not stop at the prototype stage. They run in operational systems that feed services, forecasts and official reporting, and are continuously monitored and retrained. Training and operation take place in NILU’s own AI-lab and GPU cluster, which keeps sensitive and unpublished data in-house.
AI must be reliable, transparent and ethically sound. NILU is affiliated with the Z-Inspection® initiative, a holistic process for assessing the trustworthiness of AI-based technologies throughout their lifecycle. It focuses on identifying ethical issues and tensions through socio-technical scenarios. The process is published in IEEE Transactions on Technology and Society, distributed under a Creative Commons licence (CC BY-NC-SA), and listed in the OECD Catalogue of AI Tools & Metrics.