search

gee-contrib / gee-wiki

Wiki to develop users' contributed material for GEE documentation
Other
20 stars 4 forks source link

Using LandTrendr - identifying forest disturbance, mapping, and exporting data #1

Open jdbcode opened 5 years ago

jdbcode commented 5 years ago

I can provide an example application of running LandTrendr to identify forest disturbance. The results of which will be displayed in the map and optionally exported to Google Drive

gee-contrib commented 5 years ago

This sounds like a good application. Does it build on in-situ data to validate or calibrate LT?

jdbcode commented 5 years ago

RE: build on in-situ data to validate or calibrate... no, but we do have interactive tools to help parameterize the LT algorithm and validate LT-identified disturbances. I'll include these in the tutorial as part of the current best practice.

gee-contrib commented 5 years ago

How about using some block areas from USGS Land Cover Trends project - https://pubs.usgs.gov/ds/844/ combined with https://landcovertrends.usgs.gov/download/LandCoverTrendsData_updated.html These give land cover for at least (within Landsat 5-8) 1986, 1991, 2000, 2006 & 2010 with 11 classes.

Do you have a good metric to validate a segmented time series?

jdbcode commented 5 years ago

These sound like great resources - thanks! I'll check them out soon - one concern for using them as validation of LT-identified disturbance is that many disturbances don't constitute an LULC change. For instance, it is likely that many forest disturbances will not register as an LULC class change - imagine the scenario where there is a partial harvest - the LULC before and after remains forest/industrial forest, the same for the insect outbreak, windthrow, fire, avalanche, etc. In this case, our LandTrendr data would appear to have high commission error.

As for a metric to validate a segmented time series, we recommend using the TimeSync application. This application has not kept up to the scaling of LandTrendr to Earth Engine, but there are some similar solutions being worked on and a legacy version can still e used.

gee-contrib commented 5 years ago

I understand this. Also if your validation data is only every few years but the time series you enter LT is annual there may be missed transitions. That said, even if you have a substantial set of 'full' (e.g. annual) validation data e.g. trajectories such as AAAABBCABBCDBB where A,B,C,D are land cover/disturbances. What numerical metric do you use to compare to LT segmentation which say e.g. AAAABBAABBCDDB (notice 2 differences...) ? It is not enough to do a confusion matrix ignoring this is a time series...