Abstract

While text classification can classify tweets, assessing whether a tweet is related to an ongoing flood event or not based on its text remains difficult. Inclusion of contextual hydrological information could improve the performance of such algorithms. In this study, we designed a multilingual multimodal neural network that can effectively use both textual and hydrological information. The classification data was obtained from the Twitter-streaming API using flood-related keywords in English, French, Spanish and Indonesian. Subsequently, hydrological information was extracted from a global precipitation dataset based on the tweet’s timestamp and locations mentioned in its text. We performed three experiments analyzing precision, recall and F1-scores while comparing a network that uses hydrological information against a network that does not. Results showed that F1-scores improved significantly across all experiments. Most notably, when optimizing for precision the network with hydrological information could achieve a precision of 0.91 while the network without hydrological information failed to effectively optimize. Moreover, this study shows that including hydrological information can assist in the translation of the classification algorithm to unseen languages. Tweets filtered using this network can be used to more effectively organize disaster response, validate and calibrate flood risk models, and task satellites among other applications.

Cite as

<< citation >>

How to run

1. Setup
   • Install Python 3.6 and all modules in requirements.txt.
   • Install PostgreSQL (tested with 10.1) and POSTGRESQL_HOST, POSTGRESQL_PORT, POSTGRESQL_USER and POSTGRESQL_PASSWORD in config.py.
2. Hydrological input data
   • Register at https://sharaku.eorc.jaxa.jp/GSMaP/registration.html and obtain username and password for ftp-server.
   • Fill out GSMaP_USERNAME and GSMaP_PASSWORD in config.py
   • Run 1. download_GSMaP_hourly.py. This file will download hourly GSMaP data to the folder classification/data/GSMaP/raw. This will take a while.
   • Run 2. process_rainfall.py. This will compile a NetCDF4-file out of the downloaded files.
   • Download the HydroBasins dataset for all contintents from https://www.hydrosheds.org/downloads. Select "Standard (Without lakes)" level 9.
   • Unpack the HydroBasins dataset to data/hybas (e.g., data/hybas/hybas_af_lev09_v1c.shp)
   • Download hydrosheds_connectivity.zip from https://doi.org/10.5281/zenodo.1015799
   • Unpack the data to data/hydrorivers (e.g., data/hydrorivers/afriv/rapid_connect_HSmsp_tst.csv)
   • Download riverPolylines.zip from https://doi.org/10.5281/zenodo.1015799
   • Unpack the data to the corresponding folders in data/hydrorivers (e.g., data/hydrorivers/afriv/afriv.shp)
   • Run 3. create_river_and_basin_maps.py
   • Run 4. find_basin_indices.py
3. Textual input data
   • Download the MUSE repository (https://github.com/facebookresearch/MUSE) to source folder (e.g., MUSE/supervised.py)
   • Run 5. get_word_embeddings.py. This will download word embeddings from Facebook and create multilingual word embeddings using MUSE. This process is a lot faster when Faiss is installed.
4. General
   • Obtain a TWITTER_CONSUMER_KEY, TWITTER_CONSUMER_SECRET, TWITTER_ACCESS_TOKEN, TWITTER_ACCESS_TOKEN_SECRET from Twitter by registering as a developer at Twitter (https://developer.twitter.com/) and set keys in config.py.
   • Run 6. hydrate.py. This obtain text, date and language for tweets in the labelled data and place the hydrated data in data/labeled_data_hydrated.csv.
   • Run 7. create_input_data.py This will read data/labeled_data_hydrated.csv and outputs a pickle with word embeddings and hydrological data per tweet. The data is placed in data/input.
   • Run 8. experiments.py. This will run all experiments.
   • Run 9. analyze_results.py. The results of the experiments are placed in the results folder.