On Mon, Oct 31, 2016 at 12:07 PM, Sajid Pareeth <[email protected]> wrote: > Hi Markus >> >> >> [ currently trying to get a grip on MODIS version 6 time series ] >> >> In theory, extracting seasons such as cropping cycles is quite easy to >> implement: whenever a parameter in a time series is above/below a >> given threshold, start/stop the season. The question is how to store >> the results for multiple cropping cycles: a separate raster for each >> cycle and each start and stop date? >> > > Yes, Output could be number of rasters equal to the maximum crop cycles > found in the scene. For those pixels with only one dominant cycle can be > represented with null() in the start and stop DOY maps of the next cycles. > > I couldnt find any tool or study which captures multiple SOS and EOS in case > of crops. This would be really great to have. For example the phenology > parameters computed using R package greenbrown > (http://greenbrown.r-forge.r-project.org/phenology.php) considers only one > cycle (Not really for the crop phenology change). The method they use is > explained in page 5 of the associated paper:
Multiple SOS and EOS would be easy to implement with an option setting the maximum number of cycles/seasons. As Vero said, "The problem is to define what you mean by start and end of growing season first, i.e.: a certain increasing or decreasing growth rate, a certain level of accumulated biomass, only a threshold in VI values, etc." > > Quoting the paper: > > "In the third step, we used the smoothed and daily interpolated time series > to estimate start of growing season (SOS) and end of growing season (EOS) by > either using 50% thresholds on the seasonal greenness curve (approach Trs) > (White et al., 1997) or the derivative of the seasonal curve (approach > Deriv) (Tateishi & Ebata, 2004) (Data S1). Both approaches are based on the > definition of SOS and EOS as the mid-points of spring > greenup and autumn senescence, respectively" The derivate can be a problem in case of noisy data with values jumping around the threshold in a short period of time, particularly with daily data. IMHO, the preparation of a time series (outlier removal and gap filling) is quite important because it influences any subsequent results. This thread mentions different approaches. They can be grouped into some kind of global fitting of the whole time series or local fitting (moving window over time). I tend to prefer local fitting methods (i.lmf, local weighted regression) for time series that do not show clear and regular cycles. Further on, approaches can be grouped by the kind of fitting: harmonics (r.hants, i.lmf), polynomial regression (local weighted regression), or PCA (DINEOF). The nice aspect of DINEOF is that it combines the spatial and temporal domains. I made an attempt to port DINEOF to GRASS 2 years ago, but did not finish because it is computationally too intensive (too much memory, too slow). DINEOF uses tricks to work with reduced matrices to overcome this problem, but the tricks are tricky (a lot of work to implement). Local weighted regression on the other hand is relatively easy, I am nearly done with the implementation. LWR seems to be more robust against outliers than HANTS, and there are no overshoots as in HANTS. Markus M > > Co-dominant water control on global inter-annual variability and trends in > land surface phenology and greenness. Available from: > https://www.researchgate.net/publication/275050767_Co-dominant_water_control_on_global_inter-annual_variability_and_trends_in_land_surface_phenology_and_greenness > [accessed Oct 31, 2016]." _______________________________________________ grass-user mailing list [email protected] http://lists.osgeo.org/mailman/listinfo/grass-user
