Parametric model for intra-annual reflectance time series

摘要

Imaging and sensing technologies are constantly evolving so that, now, the latest generations of satellites commonly provide with earth snapshots at very short sampling periods (daily images). It is unquestionable that this tendency towards continuous time observation will broaden up the scope of remote sensing activities: not only will it enable real-time detection of abrupt changes (e.g. forest fires, natural catastrophes) but also it will allow for accurate estimation of slowly varying trends (urban growths, desertification). Inevitably also, such increasing amount of information will prompt methodological approaches that combine spatial image processing with time series analysis. In the present study, we propose a consistent, while very simple, mathematical model to fit intra-annual reflectance time series. It is a parsimonious parametric model corresponding to the nonlinear harmonic solution of a chaotic attractor. The linear component reflects the specific seasonal periodicity associated to any given land cover, while the nonlinear part allows reproducing waveforms exhibiting possible non-symmetries, phase shifts or amplitude modulations. To support adequacy of our model, we report on a land cover classification task based on weekly composite multi-spectral reflectance images acquired by MODIS sensor at a 500m nominal resolution. For different combinations of the available spectral bands and dates, we show that using raw measurements or model-based synthesized data, yields similar classification performances, demonstrating thus good agreement between our nonlinear model and reflectance time series.