The LSTM part refers to Long Short-Term Memory. This component handles sequences and keeps track of previous information. After the convolutional layers process each frame, the LSTM continues the work. It does not operate on individual frames independently. Instead, it analyses features from consecutive frames and models their temporal evolution (see Figure 3). Finally, the LSTM output is passed to several fully connected layers. The model was trained using a batch size of 8, the Adam optimiser, a learning rate of 0.001, early stopping, and a dropout rate of 0.3. These layers act as the decision maker. They take all the analysed information about space and time – what fire and smoke look like and how they change – and weigh the evidence.Their output is a probability score from 0 to 1. A score near 1 means the system believes there is a fire. A score near 0 means the scene is normal.

See full paper here: https://brain.edusoft.ro/index.php/brain/article/view/2030