Questions: Image Classification in Remote Sensing

Question 1

A supervised classification of Landsat imagery produces a land cover map with 85% overall accuracy, but the 'wetland' class has only 40% producer's accuracy. What does this metric indicate?

• A. Only 40% of the map's wetland pixels are actually wetland
• B. Only 40% of the actual wetland areas were correctly classified as wetland (the rest were misclassified as other classes)
• C. The wetland class covers 40% of the study area
• D. 40% of training samples were wetlands

Answer: (B)

Producer's accuracy (recall) measures how well a class is detected -- it is the percentage of reference samples for that class that were correctly classified. A 40% producer's accuracy means 60% of actual wetland areas were misclassified as other classes (errors of omission). This differs from user's accuracy (precision), which measures how reliable the classified wetland pixels are.

Question 2

Unsupervised classification is always inferior to supervised classification because it does not use training data.

• T. True
• F. False

Answer: False

Unsupervised classification can be superior when the analyst lacks sufficient knowledge of the study area to define classes or collect training data, when the goal is exploratory data analysis to discover natural spectral groupings, or when the study area is too large or inaccessible for training data collection. It also avoids biases introduced by subjective training sample selection. The best approaches often combine both -- using unsupervised clustering to refine class definitions before supervised classification.

Question 3

Explain why object-based image analysis (OBIA) often outperforms pixel-based classification for high-resolution satellite imagery.

Model answer: At high resolution (sub-meter), individual pixels capture only fragments of real-world objects (part of a roof, a single tree branch, a shadow). Pixel-based classification of these fragments produces noisy, speckled results because spectrally similar pixels from different classes are intermixed. OBIA first segments the image into meaningful objects (groups of spectrally and spatially similar pixels), then classifies these objects using spectral, shape, texture, and contextual properties. This reduces noise, incorporates spatial information that pixel-based methods ignore, and produces results that match real-world objects and boundaries.

As resolution increases, the ratio of within-class to between-class spectral variance increases, degrading pixel-based classification. Segmentation reduces this variance by averaging pixels into meaningful objects.