Mapping individual dwellings is important for emergency management, population mapping, planning and maintenance of off-grid services particularly for unplanned settlements but the process is manual & time consuming. The aim of this study was to automatically map building footprints from high resolution satellite and aerial images using artificial intelligence in unplanned urban settlements. A deep learning model called Mask Region-based Convolutional Neural Network (Mask R-CNN) was implemented to automatically extract dwellings in unplanned settlements of Nairobi (Kenya), Cap-Haïtien (Haiti) & Lima (Peru). The model was first trained using SpaceNet global dataset and further retrained using customised dataset of the study sites. Results show that the retrained model performed with an overall accuracy of 0.88 and F1 score of 0.63 representing a 35% and 37% improvement respectively compared with Mask R-CNN trained on SpaceNet data. When the retrained model was fine-tuned, the F1 score further improved from 0.63 to 0.66 whilst overall accuracy remained the same. Despite achieving a high precision (0.69) for detecting dwellings, the model struggled to differentiate classes achieving a recall of 0.64 when fine-tuned. This is because buildings in study sites had different sizes, patterns, shapes, and roof colours that made it difficult to differentiate them from background objects. The results also demonstrate that it is possible to automatically map unplanned dwellings using Mask R-CNN provided parameters are fine-tuned and model is trained on customised data. This project contributes to a broader understanding of applications of artificial intelligence in remote sensing for objects detection which in turn can improve off-grid service provision in unplanned settlements.