The ubiquity of social media makes it a rich source for physical event detection, such as disasters, and as a potential resource for crisis management resource allocation. There have been some recent works on leveraging social media sources for retrospective, after-the-fact event detection of large events such as earthquakes or hurricanes. Similarly, there is a long history of using traditional physical sensors such as climate satellites to perform regional event detection. However, combining social media with corroborative physical sensors for real-time, accurate, and global physical detection has remained unexplored. This paper presents EventMapper, a framework to support event recognition of small yet equally costly events (landslides, flooding, wildfires). EventMapper integrates high-latency, high-accuracy corroborative sources such as physical sensors with low-latency, noisy probabilistic sources such as social media streams to deliver real-time, global event recognition. Furthermore, EventMapper is resilient to the concept drift phenomenon, where machine learning models require continuous fine-tuning to maintain high performance. By exploiting the common features of probabilistic and corroborative sources, EventMapper automates machine learning model updates, maintenance, and fine-tuning. We describe three applications built on EventMapper for landslide, wildfire, and flooding detection.