nedhorning / randomforestforremotesensing Goto Github PK

# LOAD LIBRARIES
require(sf)
library(raster)
library(rgdal)
library(tidyverse)
library(tools)
library(data.table)
library(furrr)
options(warn=-1)
set.seed(2012)

# COLLECTION FUNCTION
collectTrain <- function(classNums     = c(11, 12, 13),
                         classSampNums = c(0, 0, 0),
                         inImageName   = "___.tif",
                         shapefile     = "___.shp",
                         attName       = 'id',
                         nd            = 0) {
  
  # Read vector file
  vec <- st_read(shapefile)
  vec <- st_zm(vec)# To prevent possible errors
  
  # Load the image then flag all no-data values(nd) so they are not processed
  satImage          <- brick(inImageName)
  NAvalue(satImage) <- nd

  # Create vector of unique land cover attribute values
  polygons   <- vec[which(vec[[attName]] %in% classNums),]
  points     <- st_sf(st_sample(polygons, classSampNums))
  # Getting Point Attributes
  pointTable <- st_join(points, polygons, join = st_intersects)
  # Preparing Columns
  coords     <- st_coordinates(pointTable)
  predictors <- within(raster::extract(satImage, pointTable, df=TRUE), rm('ID'))
  attributes <- within(as.data.table(pointTable), rm('geometry'))
  # Combine Columns
  trainvals  <- cbind(attributes, coords, predictors)
  
  return(trainvals)
}

# SET VARIABLES
args <- list(classNums     = c(11, 12, 13),
             classSampNums = c(1000, 1000, 1000),
             inImageName   = "20190703-04_plamsc_sub.tif",
             shapefile     = "Karacabey_Dataset_FULL.shp",
             attName       = 'id',
             nd            = 0)

outputCSV   <- "TrainSet_200320.csv"

# COLLECTION
cat("Create training data to train model\n")
list_train <- furrr::future_pmap(args, collectTrain)
trainvals <- dplyr::bind_rows(list_train)