The classification results was accessed and evaluated. The assessment was based on site-specific accuracy assessment including error matrix, time consumption (processing time) used by each algorithm and the pixel count (i.e. evaluating the pixels which each algorithm allocated to each class during the classification process).
The comparison of the performance of the algorithms indicates the following facts:
First, the support vector machine algorithm of the pixel and object-based methods had better predictions with an overall accuracy of 63.1% and 68% respectively, although few misclassifications were observed. The random trees algorithm of both methods had more misclassifications in all cases and had an overall accuracy of 58.4% and 52.8% for the pixel and object-based methods.
The pixel-based and object-based maximum likelihood algorithms could not handle the overprediction and underprediction effects hence had overall accuracies of 49.8% and 38.7% respectively. The maximum likelihood algorithms overpredicted the slums in some of the classification results thereby categorized non-slum buildings as slum buildings.
The deep learning classification had an overall accuracy of 60%. The result obtained for the drone imagery showed misclassifications hence predicted the slum area as non-slums areas but the algorithm performed optimally in the orthophoto. Noteworthily, the deep learning algorithm did not achieve any result with the Sentinel-2 imagery.
In terms of the processing time, the support vector machine algorithm consumed more processing time than the random trees and maximum likelihood algorithms in both methods but the deep learning model used lots of processing time to achieve the classification.
The result of the pixel count calculation indicated that the support vector machine algorithm performed better in allocating the features to their respective classes than the other algorithms.
