Tackling climate change with machine learning: Day 1

“Machine learning is not a silver bullet but it can facilitate many climate change strategies from policy and engineering,” according to Lynn Kaack, chair of the Climate Change AI group of academics and industry representatives which is hosting a workshop at the International Conference on Learning Representations (ICLR) event which started yesterday.

The online workshop has taken inclusion and outreach as its themes and speakers and panelists have emphasized the need for researchers and practitioners to collaborate with stakeholders, customers and governments to ensure locally-accurate predictions result in appropriate climate change actions.

The Climate Change AI group working is chaired by Priya Donti, Kaack and David Rolnick, Ph.D. students and post-doctoral fellows at Carnegie Mellon University, in Pittsburgh; ETH Zürich; and the University of Pennsylvania, respectively. The workshop runs until Thursday and will focus on the application of machine learning to energy; land use, including agriculture and forestry; climate science; and climate change adaptation.

The event, originally planned for Addis Ababa, in Ethiopia, has been moved online because of the Covid-19 pandemic. It is free to view at the Climate Change AI ICLR 2020 website.

“We felt this was an opportunity to try to ‘do digital right’,” Donti told pv magazine, “especially in an era when there has been so much conversation around reducing the carbon impact associated with conferences. The digital format is perhaps more accessible to people from non-machine learning backgrounds than an in-person workshop. We actually expanded our workshop from one day to five days in the digital format in order to add more content that would engage individuals from backgrounds outside of machine learning, and thus facilitate cross-disciplinary conversations.”

Addis Ababa had been selected as a venue to broaden AI participation in the nation and throughout Africa, and that was a theme echoed by Donti and her colleagues at the online event. “We aimed to highlight climate change-plus-machine-learning applications from outside the North American and European contexts from which most of our current community hails,” said the workshop chair.

More with less

Crafting most machine learning models consists of training them on datasets of inputs and labels and during the workshop, Ciira wa Maina, senior lecturer at the Dedan Kimathi University of Technology in Kenya, discussed a model he and his team built to identify bird species based upon audio recordings. The recordings were the input data and the bird species names the data labels. Maina is using the model as a more efficient way of monitoring ecosystem health in Kenya’s national parks and explained “with limited budgets, conservation efforts must become better targeted and traditional attempts are not able to scale because they involve large teams taking on-the-ground surveys.”

Maina is monitoring birds because they are an “indicator species” that serve as a proxy for broader ecosystem health. The lecturer and his colleagues are deploying remote audio recording hardware and using machine learning models to automate the detection of bird species to gain more widespread and granular insights into ecosystem health at a lower cost than traditional methods.

Data scarcity

Many emerging economies do not collect regular information on the incomes and wellbeing of citizens and the consequent lack of data presents problems for those interested in understanding the impact of development interventions and how wellbeing changes over time.

Stefano Ermon, assistant professor of computer science at Stanford University, took a novel approach to the problem of data scarcity by leveraging two widely available data sources: global satellite imagery and Wikipedia articles. Ermon was able to make accurate predictions of economic development and crop types, and to access infrastructure in data-scarce areas using just satellite imagery as input data.

“We used the geolocation information – latitude and longitude – of Wikipedia articles to match the article’s location with a specific satellite image,” said Ermon. “The article can then be thought of as a detailed caption for that satellite image.” The assistant professor created a model to process the satellite image and Wikipedia article together which resulted in a newly processed form of the satellite image which would contain the information of the Wikipedia article. The newly-processed satellite images enabled Ermon’s team to make more accurate predictions of economic development in areas with very little data. The World Bank has been able to verify the effectiveness of the model and, in partnership with fossil fuels philanthropic association the Rockefeller Foundation, Ermon and his colleagues have started a company called Atlas AI to provide economic and agricultural insights.

Insights into action

A theme discussed by many speakers and panelists at the workshop was the need to build trust and a strong connection with stakeholders including local partners and governments. Georgina Campbell Flatter leads emerging market activities at weather insights company ClimaCell. Flatter discussed the importance of localized weather prediction to the livelihoods of small farmers. “Nearly 500 million households still depend on rule-of-thumb for making critical farming decisions,” said Flatter.

In some parts of East Africa, Flatter explained, farmers rely on the received wisdom that three days of heavy rains indicates the onset of the rainy season, however, changing weather patterns due to climate change have blurred the picture.

In January, said Flatter, “it rained for three days in a region of Uganda and then went dry for two weeks. Farmers who planted during this time lost all of their crops and for many this meant famine.” ClimaCell is trying to get small farmers to use weather forecasts.

In emerging markets, Flatter sees her work as trying to unlock value. “To tackle poverty, we must focus on prosperity,” she said. “And we must ask the right questions. Human-centered design approaches must be used to understand the needs of our customers and end-users.” Many farmers have some access to weather predictions but those forecasts often go unheeded because they are not communicated in actionable ways or the farmers do not trust them.

Call to action

Given the far-reaching consequences of climate change and machine learning, there have already been many opportunities to see how the topics intersect, at the ICLR 2020 workshop.

In his closing remarks, Ermon emphasized the importance of actionable insights to address global issues.

“Big problems like ending poverty and climate change often have limited data available and, as a result, programs are poorly targeted, social impacts are poorly understood, products are not developed, money is wasted and opportunities are lost,” he said. “The key challenge to addressing big problems like ending poverty and climate change is how to use all the datasets that are becoming available globally to extract actionable insights that can be used to come up with better approaches. That is where we need more help and that is why I am so excited to be part of this workshop because we need more people like you to work on these problems and to build a bridge from data to insights to guide better policies.”

By Dustin Zubke

This article was amended on 28/04/20 in line with changes requested by the workshop organizers. The changes included removal of a reference to Climate Change AI as a lobby group and removing the statement that people wishing to ask questions of speakers and panellists at the workshop would have to register by paying a $50-100 fee.