Thematic Mapping in Lund Municipality 

The aim of this project is to produce a thematic map (or land cover map) for the municipality of Lund using remote sensing, ancillary data, and data collected in the field. 

Quantitative analysis (pre-classification):

• • Digitized 8 land cover classes using FCC.

  • Created signature files for classification.

  • Signature separability (SEPSIG): assessed for 8 classes (Urban, Water, Bare Soil, Quarry, Cropland, Coniferous, Deciduous, Grassland) across bands B2, B3, B4, B8.
    
    

Field Data Collection 

• Random points: 40 generated in ArcGIS; 30 visited in the field.

• Sampling method: Simple random sampling for unbiased representation.

• Data cleaning: Removed outliers and misaligned entries.

• Data split: 85% training, 15% validation (proportional representation).

• Repeated for 6-class classification (merged Bare Soil & Grassland into Cropland).
  
  

Supervised Classification

• Methods applied:

  • MLC (Maximum Likelihood Classification) – done in Terrset.

  • RF (Random Forest)—done in RStudio.

  • SVM (Support Vector Machine) – done in RStudio.
    
    

• Maximum Likelihood Classification in IDRISI TerrSet:

  • Used 85% randomized training data points for 8-class and 6-class.

  • Training polygons digitized; signature files created to guide classification.
    
    

• RF & SVM in RStudio:

  • Combined 4 Sentinel-2 bands (TIFFs) into a single raster.

  • Defined 8 and 6 land cover classes for classification.

Post-Classification Analysis

• Accuracy assessment:

  • Error matrix.

  • Overall accuracy.

  • Kappa statistic.

  • User’s accuracy (class reliability).

  • Producer’s accuracy (class completeness).
    
    

Classification Challenges

• Urban and quarry classes are spectrally similar, making them difficult to differentiate.

Machine Learning Differences

• SVM vs RF (8 classes):

  • SVM classifies more grassland than RF

• Reduced classes (6-class):

  • Grassland pixels shift to cropland.

  • SVM ends up with more cropland than RF.

Accuracy Results

• MLC (Maximum Likelihood Classification):

  • Accuracy improved from 25% → 79% when classes reduced from 8 to 6.

  • Improvement due to:

    • Additional cropland training points.

    • Merging grassland, bare soil, and cropland → reduced spectral overlap.

  • MLC sensitive to overlapping signatures → unreliable with 8 classes.
    
    

• SVM & RF:

  • Performed better than MLC overall.

  • SVM stronger with 8 classes due to handling nonlinear class boundaries (Pal & Mather, 2005).

  • Both improved with fewer classes, though less dramatically, since they already handle complexity well (Abdi, 2020).

Key Findings

• 6-class setup outperformed 8-class due to reduced spectral confusion.

• Cropland is the dominant land cover class in the 6-class classification.

• Land cover shaped by:

  • Elevation → forests dominate higher areas.

  • Human land use → cropland & urban areas concentrated in flatter, lowland zones near water/infrastructure.