search

WilhelmusLab / ice-floe-tracker-pipeline

Processing pipeline for IceFloeTracker.jl
1 stars 1 forks source link

Landmask appearing in the labeled floes image #117

Open danielmwatkins opened 6 months ago

danielmwatkins commented 6 months ago

I've been running the IFT-pipeline on a random sample of locations throughout the Arctic. I've run into lots of issues where the pipeline simply fails. Out of 189 cases, IFT-pipeline made it to the preprocessing step 148 times, and successfully made hdf5 files of the segmented image only 36 times (19% of the time).

There's some cases where it does an OK job, at least mainly finding sea ice floes. Case 007 below is about the best that the algorithm can do at the moment -- the segmented image mainly is showing identified floes aligned with real floes, albeit with a large negative area bias.

007_segmentation_results

Case 022 is stranger: there appears to be an issue where occasionally the dilated landmask gets labeled as a floe. This happens in 12 out of the 36 completed runs, so it's a pretty common issue. 022_segmentation_results

The land mask, for context: image

The common issues show up all of our test regions, so I've attached a CSV with the specifications for the 21 Baffin Bay cases. baffin_bay_100km_cases.csv

These cases were all run with the default settings of IFT.

tdivoll commented 2 months ago

See this issue in the IFT.jl source package for context. May need to add a param to the CLI or Cylc based on image size.