A workflow for preprocessing, building training samples, classifying and assessing accuracy for Sentinel, Landsat or drone imagery.
Choose Sentinel-2, Landsat or other imagery based on study period, cloud cover, spatial resolution and sensor.
Before classifying, confirm that image dates match the dates of your ground samples.
Perform atmospheric correction, cloud masking, reprojection, clipping to the study area and band stacking.
Keep processing consistent across scenes to avoid introducing systematic differences.
Build classification samples from field surveys, visual interpretation or existing land use maps.
Samples should cover within-class variation, and training and validation samples must be kept separate.
Classify the imagery using random forest, SVM or other machine learning methods.
Random forest is usually a robust choice for a first classification model.
Compute the confusion matrix, overall accuracy, Kappa or F1-score, and check easily confused classes.
Don't report only overall accuracy in a paper; report per-class accuracy as well.
Apply patch filtering, set colours and tidy the legend for the classification result, then export the final thematic map.
Use intuitive class colours, e.g. blue for water, green for vegetation and red/grey for built-up land.
