As I mentioned in a previous post, last September I participated to the Recent Advances in Quantitative Remotes Sensing Symposium. I presented several posters there. One of them was about the assessment of the classification accuracy of Venus and Sentinel-2 sensors for land cover map production.
While the results of the study are interesting, I think that the most important thing that this paper shows is how a time series with realistic reflectance values can be simulated.
The idea here is to find a good balance between image synthesis (low accuracy) and physically sound simulation (need for ancillary data and computational complexity). The choice made here is to use a real time series of Formosat-2 images (only 4 spectral bands) in order to simulated Venus and Sentinel-2 time series with the same temporal sampling but with more than 10 spectral bands.
The Formosat-2 time series is used in order to:
- Estimate the LAI (leaf area index) for each pixel
- Give a spatial distribution using a land-cover map
A database containing leaf pigment values for different types of vegetation is used together with the above-mentioned LAI estimation in order to drive a reflectance simulator. The simulated reflectances are then convolved with the relative spectral responses of the sensors in order to generate the simulated images.
The poster presented at RAQRS 2010 is here.