The significant advancements in the fields of image processing and machine learning to implement systems for early accurately smoke and fire detection still present research challenges for certain biomes. The cerrado is a fire-prone ecosystem that covers a large area in South America, where videobased fire smoke detection using cameras could be more efficient than sensor based systems. This work presents the SEMFOGO-DF fire monitoring system, a solution composed of a distributed processing architecture and a deep-learning computer vision algorithm for smoke detection and emergency alert generation. The solution performs smoke detection in image sequences using a two-phase algorithm that includes automatic zoom operations to confirm smoke predictions. The first phase analyzes image sequences and classifies regions of the image with a high probability of fire and in the second phase, a smoke classification is conducted on a zoomed image. The proposed solution underwent an experimental evaluation in the Brazilian Federal District, which is situated in the Midwest region of Brazil. Experimental results show that the two-phase algorithm can consistently reduce the number of false alerts generated by the first phase alone, with a relatively low reduction in the detection rate. The development of the solution also allowed the creation of two novel datasets. The first dataset consists of image sequences of the cerrado biome, annotated with smoke contours. The second dataset includes zoomed images of the cerrado landscape, annotated with labels indicating smoke and nonsmoke occurrences. These datasets provide a robust representation of the cerrado biome and have the potential to aid other research groups working on similar developments.