How Are Data Scientists Contributing to Social Good Right Now?

In disaster scenarios, information can be both scarce and overwhelming. With the central communications infrastructure compromised, ad-hoc communication technologies such as local Bluetooth networks can combine with social media to form an unconventional but useful lifeline for rescue services and volunteer organizations.

In certain cases, secondary surveillance and monitoring systems, from closed loop CCTV networks to satellite feeds, can provide additional data where normal channels have been disrupted.

The Artificial Intelligence for Disaster Response (AIDR) repository is a collaboration among data scientists and engineers across Asia and the west, with the aim of providing open-source monitoring and analysis tools to develop action maps from social network messages in the wake of a disaster.

The initiative was born in the wake of a 7.7 magnitude earthquake in Pakistan in 2017, when the United Nations asked the Qatar Computer Research Institute (QCRI) to help interpret the torrent of social media posts from the affected zone in order to help target relief efforts.

Within a few hours of the earthquake, over 35,000 geolocated tweets had been collected via QRCI's Micro Mappers NLP (Natural Language Processing) worker threads, which ran filters over the rapidly accruing dataset to identify disaster hot-spots to be passed on to the triage processes overseen by the UN and volunteer aid.

The Micro Mappers provided 14,000 tweets to the Digital Humanitarian Network's Standby Volunteer Task Force (SBTF), and within 30 hours 341 images from the tweets had been assimilated into AIDR in order to generate a human-tagged machine learning classifier algorithm capable of identifying pertinent messages with a near to 80% accuracy.

AIDR, which constitutes the interpretive framework for the data, leverages convolutional neural networks (CNNs) that are additionally capable of creating segmentation maps to visualize crisis spots, aided by UAV systems.

The system is collaborative: AIDR's tagger module initially relies on human-tagged posts as the ground truth for a classification algorithm which will then spontaneously classify tweets within its estimated catchment range of pertinence. AIDR has subsequently been used, together with diverse intelligence-gathering tools, in a number of disaster scenarios around the world.

In the interval, various other similar machine learning-driven disaster relief systems have been developed:

• The American Red Cross's Missing Maps project combines crowdsourced mapping with a Facebook machine learning layer to identify informal communities that are not registered in popular mapping systems.
• New research out of the University of Texas uses transfer learning via the VGG-16 CNN to collate and map social media posts in disaster scenarios.
• The US Department of Defense is in the third year of its XView challenge to AI engineers, with the most recent iteration concentrating on computer vision algorithms to more effectively analyze satellite and aerial image streams. The aim is to more effectively ingest and process data for post-disaster damage assessment.

Meteorological data presents one of the most confounding challenges for statisticians and data scientists seeking to create worthwhile prediction models. In the case of air quality, unnatural factors such as carbon monoxide, nitrous oxide, and methane gas add yet more variables to weather patterns and air flows, which are already difficult to model accurately.

Current approaches to model-building favor traditional long-term data evaluation, with little global consensus regarding investment in real-time air monitoring systems. While sensor networks are relatively affordable, machine learning equipment and expertise is usually more difficult to obtain, and connectivity for remote sensors presents an additional hurdle, as we will see.

A number of global initiatives and science competitions are leading the way in the development of potential affordable new workflows and methodologies for meteorological data science. Many are international initiatives, offering the hope for consensus regarding which machine learning and data science techniques will be most effective in tackling air pollution in real-time monitoring systems.

Unsurprisingly, investment is easiest to find where meteorological issues have a direct and uncontested effect on society. Therefore, the topic of air pollution in cities is currently the 'loss leader' for the sector.

Imperial College London's Data Science for Social Good Fellowship Programme (DSSG) includes an initiative to improve air traffic data via machine learning, with a view to developing new regulatory systems for improving air quality in cities.

Conventional traffic data is often low-resolution and regional, with the output data being too aggregated and infrequent to permit ad-hoc interventions or experiments with traffic regulation. So to achieve improved coverage and greater data resolution through frequency, the DSSG project uses a GPU to interpret the image streams of 600 Transport For London CCTV cameras in real time.

The project is aimed at informing better policy decisions, serving as background data for new green route developments in cities, and at providing open-source datasets for better air quality predictions for meteorological offices.

Machine learning is also powering the AirQo Ugandan Air Quality Forecast Challenge, which has spent three years building new sensor-driven monitoring services for urban areas in sub-Saharan Africa. The vast amounts of data generated require a great deal of computation, a logistical problem that has been alleviated by a 2019 grant from Google to provide a cloud-based platform for the project's data science needs.

Research out of China, one of the world's most affected areas in terms of air pollution, has proposed an air quality forecasting framework called DAQFF, comprising two deep neural networks. The architecture leverages one-dimensional convolutional neural networks (CNNs) and a bi-directional long short-term memory (LSTM) recurrent neural network (RNN).

Dozens of other research efforts around the world are making headway toward standardized approaches to air quality data analysis through AI.

In common with many proposed civic and environmental monitoring frameworks, live air quality analysis systems require effective sensor networks in outlying locations that may be poorly served in terms of bandwidth availability. Arguably the biggest bottleneck in a machine learning workflow is not the computing power, but the lack of even a meagre bandwidth from a sensor's location.

Thus many of the air quality monitoring systems envisioned are essentially awaiting the roll-out of 5G in the near future. The 5G spectrum includes 'low-band' or 'sub-6' 5G frequencies that will be suitable for lightweight data packets over considerable distances from the transmitter exchange.

Negative news about data science in politics risks to entirely frame the debate, as well as to define the regulatory landscape and poor public profile likely to ensue for the sector. Therefore, it's important to consider that nearly all potentially harmful new technologies have transpired to have beneficial uses, and that the good works of AI-driven data science are in evidence and growing all over the world.