Question : what features does Iota2 compute from S2 images and how to analyze random forest model
I'm using Iota2 to compute ground classification over Guadeloupe. I recently added DEM derived features :
- elevation (from SRTM 30m)
- slope (computed with gdaldem)
- cosinus & sinus of the aspect angle (to avoid a discontinuity at 0/360).
This had really improved the visual result, over the area that was covered by clouds.
I wanted to see if these new values were taken into account in the RF model. To do that, I simply counted the occurrences of each "var:xxx" in the "model.txt" where xxx is the index of the feature in the datacube. I don't know if it's a good idea, but the results are pretty interesting :
- the other features (S2 bands and derived indices) are used between 5000 and 10000 times in the model
- the elevation and the slope are used respectively 36000 and 20000 times.
- the cos and sin of aspect are used around 6000 times
I had plotted a graph (see below) that shows a kind of repetitive pattern (36 peaks - the temporal resolution is 10 days). Furthermore, the peaks in the middle of the graph seem a bit higher. If it's time ordered, it should be summer.
What are the features computed (by default) by Iota2 from Sentinel2 bands ? I suppose there are NDVI, NDWI, brightness, but cannot find the list in the doc.
Is that right that these features are stacked "after" the time serie of Sentinel 2 radiometric bands ?
And last not least : did you already make this kind of analysis over the Random Forest model file ? I suppose there are better / other methods to interpret the model...