Abstract: This study presents a pipeline for estimating small-area population in Melusi informal settlement using high-resolution UAV imagery and machine learning. A 3 cm UAV orthomosaic of Melusi was tiled and paired with hand-digitised binary building masks. A DeepLabV3 semantic segmentation model was trained to extract building footprints across the settlement. Outputs were evaluated alongside a Gaussian Process Regression (GPR) model that estimated total population from household survey data (GRT-I, 2021-2023, n=21,470 individuals), tested under RBF and Matern kernels. DeepLabV3 reproduced the spatial structure of the settlement despite class imbalance. GPR predicted a total population of 21,470 against the observed 21,470, with MAE of approximately 1.3 and RMSE less than 2 persons per spatial unit. The pipeline is computationally viable and accurate at settlement scale, contributing a replicable framework for other South African informal settlements.