For this assignment, image segmentation techniques were explored using different software programs namely eCognition and ArcGIS Pro. A contextual or bottom-up approach was used in the image pre-processing particularly the Multi-resolution Segmentation (MRS) in eCognition, and the Segment Mean Shift method in ArcGIS Pro.
Several parameters were tried using the Pleiades sample image with a 2-meter spatial resolution.
Parameters considered:
• Scale – determines the size of the resulting objects
• Shape/Color – defines the textural homogeneity or the digital value of the objects
• Compactness/Smoothness – optimizes the resulting image objects with regards to smoothness and compactness in terms of the shape criterion
Based on the results above, changing the scale level parameter produced the most pronounced change in terms of the size of the generated features. In Case 3, increasing the importance of color over shape, as well as compactness over smoothness produced a more detailed result as compared to Case 2 with equal importance given for the color and shape, as well as between compactness and smoothness. Since Case 3 considers the spectral properties more than the textural homogeneity, the output produced a more detailed segmentation of the vegetation features as compared to Case 2 wherein, bigger features were produced.
Parameters Considered:
• Spectral Detail - Set the level of importance given to the spectral differences of features
• Spatial Detail - Set the level of importance given to the proximity between features
• Segment Size - Merge segments smaller than this size with their best fitting neighbor segment.
For the first to attempts, only the spatial detail parameter was changed. It can be seen that by increasing the value of the spatial detail, the forested area was not delineated mainly because the algorithm was not able to distinguish the variation between the shape of forest and other vegetation area since they less variation in terms of the shape of the vegetation layers unlike when comparing other features i.e. roads vs buildings.
In the last attempt, the spectral detail was set to maximum resulting to a more detailed and accurate segmentation using the mean shift algorithm in ArcGIS Pro. The algorithm was able to distinguish different vegetation classes by looking at the spectral properties more than the spatial arrangement and configuration of features.
