Tackling Climate Change with Machine Learning

Source: Deep Learning on Medium

Tackling Climate Change with Machine Learning

If the title of this blog post seems somewhat familiar, it’s likely because you’ve heard of or read this thoroughly-sourced paper released back in June of this year. The paper lays out an overview of the myriad areas machine learning can provide impactful solutions to mitigate the effects of climate change. While the entire paper is worth summarizing (and reading!), for this blog post I will focus on two specific areas that I found interesting: carbon emission capture/reduction and climate prediction.

Carbon Emission Capture/Reduction

In 2018, the Intergovernmental Panel on Climate Change (IPCC) estimated that, within 30 years, the world will be facing catastrophic consequences if we do not limit and severely reduce global greenhouse gas emissions. Despite international accords, global protests, and the overwhelming scientific consensus that we need to reduce our emissions if we are to avoid catastrophe, our global emissions continue to increase. If governments are unwilling to act as quickly as necessary to decrease emissions, then investment in carbon-capturing technologies is a necessity. While the technology itself exists, it is in its infancy. But machine learning can aid this new technology in a variety of ways.

The paper outlines three options for reducing carbon emissions. It also admits that, while these technologies do currently exist, the applications they outline, specifically as it pertains to machine learning, are speculative.

1) Natural or Semi-Natural Methods

While global emissions have been increasing year over year, deforestation has also added fuel to the fire. About half of the world’s tropical forests have already been cleared. Even worse, estimates put total deforestation at 18.7 million acres of forest per year, the equivalent of 27 soccer fields worth of forest being cleared every minute. In total, deforestation accounts for roughly 15% of all greenhouse gas emissions.

Given all the deforestation occurring, providing tools to help track deforestation can provide valuable data for policy-makers and law enforcement. According to the paper, machine learning can help “differentiate selective cutting from clearcutting using remote sensing imagery.” It can also be used to “detect chainsaw sounds within a radius of a kilometer and report them to nearby cellphone antenna” to alert law enforcement officials of illegal deforestation.

Additionally, reforestation can help reduce the impacts of deforestation. It is estimated that there is a capacity for 1.2 trillion trees to be planted in existing forests and abandoned lands. ML can be used to help locate appropriate planting sites, monitor plant health, assess weeds, and further analyze trends.

2) Direct Air Capture (DAC)

Direct Air Capture is a technique for extracting CO2 from power plant exhaust, industrial processes, or ambient air. Facilities are built to extract the CO2 by having air blown onto sorbents (basically sponges), which then use heat-powered chemical processes to release the CO2 in a purified form for sequestration. The image below outlines this process.

Machine Learning can help increase the efficiency of this in a number of ways. According to the paper, it can be used to “accelerate the materials discovery process to maximize sorbent reusability and CO2 uptake while minimizing the heat required for CO2 release.” It could also help to develop “corrosion-resistant components capable of withstanding high temperatures, as well as optomize their geometry for air-sorbent contact.”

3) Sequestering CO2

Unless permanently stored, any captured CO2 will inevitably be released back into the atmosphere. Therefore, captured CO2 must be sequestered. The current best methods for doing so are by direct injection into geologic formations, such as saline aquifers (similar to oil and gas reservoirs), and sequestering in volcanic basalt formations. So where does ML factor into this? In the same way that oil and gas companies have utilized ML for subsurface imaging based on seismograph traces to find extraction points, ML can be used to help identify potential storage locations. The models used by oil and gas companies can be repurposed to help trap and inject rather than extract. Additionally, ML can be used to monitor and maintain sequestration sites, especially for CO2 leaks and overall emissions detection. This process is done by using sensor measurements that “must be translated into inferences about subsurface CO2 flow and remaining injection capacity.” There has also been success “using convolutional image-to-image regression techniques for uncertainty quantification in global CO2 storage simulations.”

Climate Prediction

Current climate prediction models are used to inform local and national government decisions, help individuals calculate their risks and footprint, and estimate potential impacts of our emissions. Machine Learning has helped advance the ability to more accurately model these predictions. Much of these advances stem from the availability of data. While more data does not always equate to more accurate models, having a greater availability of the same types of data from more areas around the globe can help increase accuracy. For example, according to the paper, “newer and cheaper satellites are creating petabytes of climate observation data.” The models based on this information also generate petabytes of simulated climate data, so these models tend to build off of each other. Finally, these forecasts are computationally expensive and time-consuming. As a result, climate scientists have begun using ML techniques to combat these issues.

1) Uniting Data, ML, and Climate Science

Why are climate scientists increasingly relying on machine learning models? According to the paper, these models “are likely to be more accurate or less expensive than other models where: 1) there is plentiful data, but it is hard to model systems with traditional statistics, or 2) there are good models, but they are too computationally expensive to use in production.”

ML models have already been used for a variety of purposes among climate scientists. They have been used to calibrate satellite sensors, classify crop cover, and identify pollutant sources. It has also been proposed that deep learning could be used for “pattern recognition, super-resolution, and short-term forecasting in climate models,” along with compiling the data of environmental imagery to further accelerate ML work in the field.

Additionally, deep neural networks, combined with existing thermodynamics knowledge and previously collected data, can be used to model some of the greatest sources of uncertainty in climate models: clouds, ice sheets, and sea level rise. Bright clouds reflect light and help cool the earth, while dark clouds absorb the light and keep the earth warm. Physical models of these processes and predicting their eventual effects as the climate changes is too computationally expensive to include in global climate models, but machine learning models are not. According to the paper, “Gentine et al. trained a deep neural network to emulate the behavior of a high-resolution cloud simulation, and found that the network gave similar results for a fraction of the cost, and was stable in a simplified global model.”

Ice sheets and sea level rise have been difficult to model for different reasons. The biggest issue has been collecting data. Because these regions are isolated, dark, and cold, they are difficult to observe and thus collect data on. However, new satellite campaigns have allowed us to collect terabytes of data on these areas in the last few years. Still, the biggest problem is accurately modelling mass loss from these ice sheets and their impact on sea level rise. The largest area for improvement in these models involves snow and sea-ice reflectivity, as much of the light from the sun is reflected off of these sheets and the heat is not being absorbed. As these sheets melt, the earth will absorb the heat and continue melting the ice sheets, thus leading to greater sea-level rises.

2) Forecasting Extreme Events

Weather models are far easier and more accurate to produce than climate models. Since weather models track rapid changes in the atmosphere, and there is abundant information available to predict the effects these changes will have, models are tested and updated every day to accurately predict short-term weather conditions. Accurate climate models, however, are far more difficult to produce. They can only be tested against long-term observations, and anything longer than a week is exceptionally difficult to predict. Additionally, all available data sets used for predictions are strongly skewed because extreme events are rare, and historical data is severely lacking. While current climate models can help predict changes in long-term trends, the accuracy of these models needs to be improved greatly.

Machine Learning has been used successfully to classify some extreme weather events. Deep convolutional neural networks have been used to count cyclones and weather fronts in past climate data sets, and some techniques have been used to track storms and tornadoes. Additionally, ML is widely used to make local weather model predictions. “Various authors have attempted this using support vector machines, autoencoders, Bayesian deep learning, and super-resolution convolutional nerual networks.” These models have already been used to predict flooding patterns in various areas. With more data becoming increasingly available, long-term climate predictions will hopefully increase in their accuracy.


While the whole paper is worth reading, this post specifically focused on two sections: carbon capture and climate forecasting. Machine learning can be used in the same manner as oil and gas companies for finding injection sites and monitoring sequestered CO2. As for climate forecasting, ML can be better utilized as more data is made available, and used to better model the uncertainties that melting ice sheets and rising sea levels will bring.