Open metasim opened 3 years ago
this first one (unsupervised) seems similar to #511 .. not real sure but maybe
Can confirm the unsupervised is only happening in the branch not on develop.
About raster join: i suspect this is an artefact of using cubic convolution near areas of NODATA. FWIW this is not specific to the branch. THe current develop branch nighly build is showing the same. Link. Screenshot:
Compare to nearest neighbor (in prod docs). Note the regions of NODATA ...
Unsupervised learning: can collect the incorrectly displaying dataframe / row/ Tile to the driver, and it shows up nicely with the PNG repr... hopefully that is an important clue
another intersting thing about that is the celltype of the rf_agg_overview_raster
is int32
, for the rf_agg_overview_raster
but the retiled.prediction
column is int8.
A further note here is that just a table view of retiled
also shows a completely purple / 0 tile in the table preview
In unsupervised, retiled.select(rf_convert_cell_type('prediction', 'float32'))
shows us what we expect....
possibly something to do with the way histograms are computed here https://github.com/locationtech/geotrellis/blob/bb5f8642bd0ae04913e1742d65d86e095500486d/raster/src/main/scala/geotrellis/raster/render/PngRenderMethods.scala#L52-L61
Refer to this build: https://3415-174014757-gh.circle-artifacts.com/0/rf-site/index.html
In![image](https://user-images.githubusercontent.com/131013/97301633-5fd09700-182e-11eb-9e56-341bec3709e4.png)
unsupervised-learning.pymd
the final prediction raster is a single color:In![image](https://user-images.githubusercontent.com/131013/97315901-8fd46600-183f-11eb-9b15-3a960c4edac7.png)
masking.pymd
the masks fromSCL
aren't rendered properly as tiles. However, application of mask looks correct:In![image](https://user-images.githubusercontent.com/131013/97316228-e3df4a80-183f-11eb-8d20-2caa84c0eb14.png)
raster-join.pymd
need to confirm the MODIS columns is correct.... looks like multiple resolutions are being combined?