Overview
This project performs a supervised Maximum Likelihood Classification (MLC) of a Landsat 9 image to map seven land cover classes across a coastal study area in British Columbia. Training and validation polygons were manually delineated, spectral signatures were explored per class, and the classification was evaluated using a confusion matrix with overall, producer, and user accuracy metrics.
Objectives
- Delineate training and validation polygons for seven land cover classes
- Explore and visualize spectral signatures of each class across Landsat bands
- Train a Maximum Likelihood Classifier using a 70/30 training/validation split
- Evaluate classification accuracy using a confusion matrix
- Interpret per-class producer and user accuracies and identify sources of misclassification
Data Sources & Tools
| File | Description |
|---|---|
LC09_L2SP_047026_20240716_..._SR_BSTACK.tif | Landsat 9 surface reflectance image — 6 bands (blue, green, red, NIR, SWIR1, SWIR2), 30 m resolution |
classification_polygons_AC.shp | Manually delineated training and validation polygons for 7 land cover classes |
bands_wavelength.csv | Lookup table of wavelength values for each Landsat band |
| Package | Purpose |
|---|---|
terra | Reading and plotting raster imagery |
sf | Reading and handling vector polygon data |
RStoolbox | Maximum Likelihood Classification via superClass() |
tidyverse | Data wrangling, reshaping, and ggplot2 visualization |
caret | Model training support used by RStoolbox |
Methods
A Landsat 9 image was loaded and 271 polygons were manually delineated across seven land cover classes: Broadleaf Forest, Coniferous Forest, Exposed Soil and Rocks, High Density Developed, Low Density Developed, Non-forest Vegetation, and Water. Polygons were split 70/30 into training and validation sets using stratified random sampling. Pixel values were extracted and spectral signatures (mean reflectance with 5th–95th percentile ribbons) were visualized per class across all six bands. The Maximum Likelihood Classifier was trained using RStoolbox::superClass() with 500 pixels sampled per class. Classification accuracy was assessed using a confusion matrix, with overall, producer, and user accuracies calculated from the matrix diagonal, row sums, and column sums.
Outputs
Spectral signatures showing mean reflectance and 5th–95th percentile uncertainty ribbons for all 7 land cover classes across Landsat 9 bands
Classified land cover map showing 7 classes across the coastal BC study area
Confusion matrix showing prediction vs reference classification results
Key Findings
- Water and Coniferous Forest achieved the highest accuracies due to their spectrally distinct signatures
- Broadleaf Forest had the lowest user accuracy (36.5%), frequently confused with Non-forest Vegetation and Low Density Developed areas
- Low Density Developed had the lowest producer accuracy (61.3%), likely due to its mixed spectral composition of vegetation, bare soil, and built surfaces
- Balancing pixel samples per class is critical — unbalanced sampling can inflate overall accuracy and mask poor performance in minority classes
Skills Learned
- Loading and visualizing multi-band Landsat imagery using
terra - Reading and validating vector polygon data with
sf - Stratified random sampling for training/validation splits using
dplyr - Extracting pixel values from rasters at polygon locations with
terra::extract - Summarizing and visualizing spectral signatures with
ggplot2andgeom_ribbon - Training a Maximum Likelihood Classifier with
RStoolbox::superClass() - Computing overall, producer, and user accuracy from a confusion matrix using
diag(),rowSums(), andcolSums() - Interpreting classification accuracy and identifying sources of misclassification