Teaching a four-band camera to see more
Ongoing research into reconstructing hyperspectral detail from multispectral drone imagery over mangroves.
Hyperspectral cameras record hundreds of narrow bands and can separate species and stress conditions that look identical to an ordinary sensor. They are also costly, heavy and slow to fly. Multispectral drones are the opposite: cheap to deploy, but limited to a handful of bands.
The question
Can a model trained on paired data learn to estimate the full spectrum from the few bands a multispectral drone records, well enough to be useful over mangrove forest?
What we are doing
We collect matched multispectral and hyperspectral imagery over the same canopy and train deep-learning models to reconstruct one from the other. Much of the work is in the data rather than the network: radiometric calibration, removing bands affected by atmospheric absorption, and making a model trained at one site hold up at another.
Why it matters
If it works, large areas of mangrove could be surveyed with affordable aircraft and still yield spectral information that today needs specialised equipment.
This is research in progress. We will publish results when they have been reviewed.