A cross-region transfer learning method is proposed to identify community (e.g. car parks, public spaces and shopping centers) fire hazards based on text input provided by community members. The key component of the method, which also accounts for data imbalance, is an improved transfer component analysis (TCA) that is embedded with a local discriminant analysis (LDA) to transfer non-local rich knowledge to the fire hazard identification of local communities with an insufficient number of samples. In addition, a fire hazard knowledge map is established and applied to supplement the missing key features for fire hazard identification, and ontology modeling is applied to standardize the text features and reduce the effect of semantic ambiguity brought by cross-region knowledge transfer. The proposed method is verified based on the text data of nine fire hazard classes from Lanzhou and Beidaihe in China. Machine learning experiments show that fire hazard identification performance of all nine classes were improved with the overall accuracy, precision, recall, F1 score and AUC increased by 12%, 15%, 16%, 15% and 15%, respectively. Under data imbalance scenarios, the proposed method outperforms the state of the art methods, such as sampling-based methods, FastText and ULMFiT. The results also show that the proposed method can achieve desired performance with only half of the training samples. These findings illustrate that the proposed method can assist regions by improving fire identification results significantly through knowledge transfer. The proposed approach can be followed to build smart systems for community fire risk management with reasonable performance and high efficiency.