The circular economy of urban settlements is essential for sustainability and limiting climate change to under 2°C. While the circular economy in building stocks can be achieved through design optimized for resource recovery, challenges persist due to insufficient data on existing building stocks, material weights, and geo-locations. Urban building stock estimation and mining address these issues by mapping and analysing building resources. However, traditional urban mining methods are time-consuming, costly, and rely on expert assumptions. This study applied a novel computer vision approach and tested ResNet and DenseNet deep learning models to classify typical building facades in Australia. It automatically estimated the façade material stock and its embodied carbon. The ResNet-50 model achieved the highest performance (F1-score: 0.77, Accuracy: 0.803) and was deployed in Southport (SA 3) of the city of Gold Coast. Of the available image data, 78% was successfully inferred in Southport, while 22% remained unidentified due to the coverage limitations of open-street view images. Using model predictions, semantic data, and expert consensus, façade material stock weight was estimated, and embodied carbon coefficients were applied to account for the total embodied carbon. Results show that larger-scale autonomous stock estimation is feasible, offering significant advantages over traditional methods. This approach supports data-driven circular economy applications for sustainable urban planning and climate change mitigation.