Characterizing analyst bias in unsupervised classification of Landsat images

Unsupervised classification is a classification technique used to process remotely sensed images. Products generated from this technique are used for monitoring changes in earth's vegetation, urban settlements, and water bodies. One of the drawbacks of this classification technique is operator bias, which can influence the area of map classes. This study examined the operator bias in calculating the surface area of Keyhole Reservoir through unsupervised classification of Landsat images. Using a set of Landsat data, two analysts generated maps with water and non-water classes. Each map pair was compared to quantify the operator bias in terms of percent agreement and disagreement. Between analysts we found that the differences in distinguishing water were minimal. However, most of the bias was found along the shorelines in classifying the boundary of where the water ended and land started. Results from this study will provide insights for minimizing operator bias in future projects.