Classification of satellite imagery

During the course on Advanced GIS Analysis, I also learned to analyze raster images. I needed to create thematic raster layers from RGB satellite orthophotos by image classification in ArcGIS Pro. This technique is often used for land cover mapping to understand the makeup of an area, change detection to identify changes in the landscape, resource management, and more.

The objective of this assignment was to learn about the use of pixel-based and object-based image classification techniques and how they differ. In very simple terms:

Pixel-based image classification is a per-pixel classification technique. This means that the properties of neighboring pixels are not taken into account, instead, each pixel in an image is assigned to a particular class based on its spectral properties. Once the classification is completed, the resulting image will have different classes assigned to different colors.

Object-based image classification is when pixels are grouped together based on their similarity with neighboring pixels. They are then divided into segments, or "objects", that are characterized by their color and shape. This segmentation approach is achieved by averaging the values of pixels and taking into account other geographic information. In doing so, the objects more accurately depict the features present in the imagery, leading to more precise classification results than the traditional pixel-based approach.  

The maps above show my first try with image classification, where the pixel-based approach was used. To do this, I needed to train the image classifier in ArcGIS Pro using training samples that were based on my manual input. This was a time-consuming and computationally-heavy process that required multiple iterations and generalization procedures to look good enough. The result clearly doesn't represent what is on the ground accurately, but rather shows a more general rendition.

The maps below show my second attempt at using image classification, but this time comparing both techniques. The object-based clearly yielded a more satisfying result than the pixel-based approach. Since the pixel-based approach operates on a per-pixel basis, it seems that the classification will be more inaccurate the more complex the scene is, producing misclassified pixels that either need to be reclassified or generalized heavily. In that sense, it is more logical to use object-based classification since it can deal better with the contextuality between pixels.