The surveillance and reports on accidents for the current traffic management are carried out by viewing and inspecting CCTV footage. This method is slow to detect incidents and requires a lot of manpower. Since nowadays communication technology and social media play a role in reporting such incidents this research has been conducted to investigate accidents of interest from text messages reported on social media in Thai. There is currently no research or platform that supports this type of work clearly. This research presented the development of a neuron network memory-learning model to solve the problem of incidence pattern classification and incidence severity identification from social media messages in Thai. Using deep learning models such as MLP CNN Bi-LSTM and LSTM+CNN the study was designed and divided into three experimental patterns. This included examining patterns to identify traffic incidents that may be reported as general news or traffic reporting; examining patterns to indicate the type of incidents such as traffic accidents disasters damaged roads or others other than those mentioned; and examining patterns to specify the severity levels of the incidents such as normal intermediate traffic lane blocking or immovable. The experiments demonstrated the capability of CNN+LSTM learning models with the best incidence detection result at 93.44 %. The CNN model gave the best result in identifying a pattern of incidence at 85.29 % and the LSTM model was best able to determine the severity levels of the incidence at 88.53 %.