Regression in R

Very high-resolution reference data are usually difficult to obtain or only available for small areas of the study area. However, low-resolution data, such as Landsat 8 (30 m), are available in a high spatio-temporal resolution. Using a regression method, we can create sub-pixel information by relating the high-resolution information to very low-resolution Landsat 8 pixels.

We want to perform a Support Vector Regression in order to regress proportions of imperviousness for each Landsat 8 pixel in Berlin. For this we will use two data sets in this section:

  1. a shapefile containing very high-resolution land cover information (including imperviousness), based on a digitized digital orthophoto of 2016 (Berlin Environmental Atlas)
  2. a Landsat 8 acquisition (ID: LC08_L1TP_193023_20170602_20170615_01_T1), which you may already have acquired during the L8 Download Exercise

Landsat 8 scene overlaid by the shapefile (colored by classes) in QGIS. The different detail levels become very clear in this visualization

Section in a Box

Prepare the samples for training
– learn how to preprocess your shapefile
– extract raster features and percentages of your target class (e.g., imperviousness)
– create your training data set for regression analysis
SVM Regression
– learn how to perform a Support Vector Regression (SVR) in R using the e1071 package
– predict the whole image data based on your regression model