Workflow

In the following, the most important steps of the image classification workflow of {mapme.classification} are described; the figure guides the reader through these descriptions. The goal of this tutorial is to classify Sentinel-2 images into different LULC categories in an agricultural landscape in Ethiopia, by using the {mapme.classification} package. Creating such LULC maps in different years would allow – for example – to back trace changes in the cropped area, which might be an intended outcome of a project being evaluated. Several algorithms, such as the Random Forest algorithm, can be calibrated based on user provided input data and reference data. The calibrated model is applied to classify the input data into user defined categories of land use / land cover (LULC). The assesses the accuracy of the LULC maps and estimates the area of each LULC category. Optimally, this LULC area estimation can be bias corrected by considering the class prevalence in the training data. The single processing steps are as follows:

pre-process Sentinel-2 images via the {[mapme.vegetation](https://mapme-initiative.github.io/mapme.vegetation/} package
extract reference data for known locations
split reference data into training and validation set
training a model possibly based on spatio-temporal folds and a Forward-Feature-Selection
run the final model to yield spatially comprehensive predictions
optionally run post-processing routine to remove isolated pixels
return evaluation metrics of the final model based on test data
calculate LULC area and optionally correct for biases by considering the class prevalence in the training data

Workflow

Last modified: 2021-10-21

Workflow