Moisture stains signal progressive degradation processes and can indicate points or regions of the roof slab where infiltration is occurring, which compromises the performance and durability of the building system. During building inspections of roofing systems, the inspector's field of vision is different from the viewing field of the drone's cameras during overflight. For this reason, traditional inspections do not always identify the presence of stains and their severity, making maintenance on flat roofs a complex activity. In this context, the objective of this experimental study is to analyze the use of deep learning-based semantic segmentation with drone-acquired images to map and monitor damp patches during automated building inspection processes on flat roof systems. This experimental research tested two semantic segmentation convolutional neural networks (FCN with the ResNet50 backbone and DeepLabV3 with the ResNet101 backbone) and a transformer-type deep artificial neural network (SegFormer with the MiT-B1 backbone). In all cases, the Adam, Adagrad, and SGD optimizers were tested, and the learning rates 1e-2, 1e-3, and 1e-4 were evaluated. These models were compared using four performance metrics. The best results for the metrics obtained by FCN stand out when optimized by Adagrad with a learning rate of 1e-2. The averages obtained in this case for precision, recall, f-score, and IoU were 79.69%, 67.81%, 73.09%, and 57.70%, respectively.