Deep Neural Networks for Emergency Detection

摘要

The increasing deployment of sensor networks has resulted in an high availability of geophysical data that can be used for classification and predictions of environmental features and conditions. In particular, detecting emergency situations would be desirable to reduce damages to people and things. In this work we propose the use of a deep neural network architecture to detect pluvial-flood emergencies, building upon and extending our previous works [1, 2] in which we gathered a large set of rain measures coming from a sensor and surveillance network deployed in the last decade in the Italian region of Tuscany and built a database of verified emergency events using a manually annotated set of resources found on the World Wide Web. We used a stacked LSTM [3] network to classify 4-day-long sequences (the measures for a given day and the three days prior) of pluvial measurements gathered from the whole set of stations belonging to Servizio Idrogeologico Regionale in Tuscany. After empirical tests, we chose two 2 LSTM layers with 256 outputs each for the hidden part of the network. Using multiple layers we exploit the abstraction power for pattern recognition in time sequences that has been previously recognized for LSTMs: lower layers are able to detect the most significant variations, while the higher ones use these patterns to spot emergency events. As they are very infrequent, a balanced subset of quiet days has to be considered to build a binary classifier to avoid overfitting. To increase the number of relevant true examples we performed a linear interpolation of existing sequences, generating 10 new examples for each original one. After training the network using 560 examples, we tested its performance using 1276 sequences. We had 131 true positives, 1074 true negatives, 64 false positives and 7 false negatives. This leads to a precision of 0.67 and a recall of 0.95, so the F_1-score is 0.79. Accuracy is also high (0.94). We are planning to train this network on a bigger dataset, and then perform transfer learning to have an overall better classifier.