Classification of informal settlements using novel object-based machine learning algorithms and drone imagery, Slovo Park, Johannesburg, South Africa

The global proliferation of informal settlements is closely linked to urbanization, often leading to environmental degradation and socio-economic challenges for their inhabitants. This study introduces a novel method for classifying informal settlements, focusing on Slovo Park in Johannesburg, South Africa. High-resolution imagery was captured using an Unmanned Aerial Vehicle (UAV) and down-sampled from 0.0165 mm to 0.16 m to reduce heterogeneity for improved segmentation using the mean shift algorithm. Two versions of the down-sampled images were created: one for direct classification and another with an integrated Normalized Digital Surface Model (NDSM). Both images were classified using Random Forest (RF) and Support Vector Machine (SVM) algorithms. Accuracy was assessed using a Confusion Matrix, with overall accuracies ranging from 60% to 74%. SVM outperformed RF, particularly when NDSM was incorporated. This study introduces an innovative approach by integrating machine learning classification algorithms with various spatial metrics at the patch, class, and landscape levels to characterize informal settlement dwellings. Principal Component Analysis (PCA) was applied to reduce dimensionality at the patch level. The NDSM significantly improved classification accuracy, while class and landscape matrices provided insights into settlement characteristics, such as fragmentation and vegetation cover.