Supervised & Pixel-Based Image Classification (UCLA Master's Coursework)
As one of my many coursework projects for the UCLA Master of Applied Geospatial Information Systems and Technologies (MAGIST) program, I practiced different methods of image classification.
The supervised object-based classification result is considerably more accurate than the supervised pixel-based classification result. This is apparent by comparing the confusion matrices: the object-based classification result has consistent higher user and producer accuracy (i.e. fewer false positives and false negatives).
This also makes intuitive sense based on a visual comparison of the two classification results. The pixel-based classification has several other inaccuracies clearly visible just from a visual survey of the map: for example, multiple patches of sand and rock are classified as “developed”, shown on the map as wispy gray patterns distant from any actual settlement. This seems to be a somewhat “organic” error, reflecting real-world similarities as opposed to errors in training data; scrutinizing the Landsat image reveals that there are many areas of gray sand and rock, particularly dry river beds, that are extremely similar in hue to roads and other paved impervious surfaces in developed areas. That error is nicely avoided by the object-based classification.
I considered and reconsidered my choice of classes for the Landsat image, opting to focus on the most clearly distinguishable aspects. The “barren sand and rock” category was so extensive, encompassing everything from gray sandy riverbeds to bright whitish sand patches to gray and ocher rock formations, that I wondered if I was making a mistake by lumping all of these features together. However, I couldn’t come up with a defensible subcategorization: the distinct land covers of snow, vegetation, human-created development, and water were clearly identifiable as separate entities, but the vegetation-free sand and rock that dominated the landscape seemed to occupy more of a spectrum, both in color and texture, in which they were more similar to each other than to any other land cover. Thus, even though it seems surprising that the object-based classification result overwhelmingly consists of the “barren sand and rock” class, I believe this is an accurate categorization of the landscape. As both the object-based and pixel-based classification use a Support Vector Matrix method for reclassification, as well as the same training data, differences between them are extremely likely to be due to object-based/pixel-based status alone.