Things to incorporate at some point

[admahood]

• [ ] use ranger instead of randomForest
• [ ] make road mask pulling and mosaicing county data
• [ ] explore the idea of a smoother
• [ ] the thing where we start with an NA extent bigger than the scene to avoid extent-not-matching issues
• [ ] figure out an academically rigorous reason for the vegbank validation performing better despite lower accuracy
• [ ] get soilgrids.org data, get a later showing sand content
• [ ] fill in NAs on final predictions with years before and after (mean of the two?)