Discussion

SEPAL vs. GEE

SEPAL GEE
pro - quick to learn con - takes more time to learn
pro - no programming experience needed con - requires understanding of programming basics
con - complicated workflow using multiple different tools pro - simple worflow that can be done in 2 scripts
con - limited choices for input data pro - can import any data sets from personal computer or GEE Data Catalogue (PALSAR, Harmonized Landsat Sentinel)
con - limited choices for pre- and post-processing pro - can access all pre- and post-processing GEE tools (smoothing, cloud masking, calculating indices)
con - limited options for altering model parameters pro - many options for altering model parameters (multiprobability mode, variable importance)
con - difficult to rerun to test out different versions pro - easy to rerun to test out different versions

Ways to Improve Accuracy

1. Preprocessing

  • increase the number of training and testing points
    • overall and/or using a stratified approach with more points in classes with low accuracies
  • change or combine cloud masking strategies
  • composite imagery for more than 1 year
  • merge or separate classes of interest (e.g. one combined forest class instead of separated forest classes, separate wetlands class)

2. Model Development

  • add and remove predictor variables based on variable importance and accuracy (e.g. different combinations of imagery sources or indices)
  • increase the number of trees in the random forest model
  • generate more training points in the areas of highest uncertainty and retrain the model with the addition of these new points

3. Postprocessing

  • change smoothing functions
  • overlay with hand-digitized data

2025 --- Spatial Informatics Group / NovaSphere --- Financial support from Environment and Climate Change Canada