Spatial Holdouts and Feature Selection

Findings Both model types (gbm and random forest) are very sensitive when using all variables and no spatial hold-outs; however, when increasing the hold out size (iid becomes more appropriate) we see that random forest takes some time before it stabilizes unlike gbm which is able to control for that effect early on. FFS variable selection at cluster holdout 10 was 9 variables while the new variable selection (unnamed) used only 5 variables at a cluster holdout of 10 while keeping the same accuracy; making the model more parsimonious.

By Josh Erickson

March 11, 2021

gsutil workflow

Working with gsutil Introduction This is a workflow on how to download multiple objects from a bucket on google cloud storage. Basically, so I stop forgetting how to do it. Moving objects Technically, you can create your own folders via some (above my head) method but I’ll just default to using the UI on GCS. From there, you can then start to move files that you’ve uploaded via desktop, R, earth engine, etc.

By Josh Erickson

March 4, 2021

Global Forest Change

Introduction Below is a workflow used to create a forest change raster (2011-2019) for the Watershed Condition Class (WCC) analysis in USDA-USFS Region 1. The workflow leverages Google Earth Engine (GEE) (Gorelick et al. 2017) to get large remote sensing products fast and easy. This allows for reproducibility in the future as well as quick return times for clients, i.

By Josh Erickson

March 4, 2021