Machine Learning Should Combat Climate Change
by Veda Sunkara, Software Engineer and Data Scientist, Cloud to Street
As the climate crisis worsens, so does the magnitude of destruction caused by natural disasters. All over the globe, floods, droughts, fires, and storms are worsening; it is widely accepted by scientists that the ongoing devastating wildfires in the western U.S. and the hurricanes in the eastern U.S. have been made worse by increased climate instability. As these crises deepen, so does the need for aid. Yet convoluted or delayed reporting causes potentially devastating lags in its delivery. It is imperative that disaster response systems and relief efforts are refined and perfected to minimize loss of life. Beyond policy changes, we must use science and technology to bridge these major information gaps and make disaster planning more robust. Against this backdrop, Cloud to Street is leveraging cutting-edge remote sensing and machine learning technologies to help people on the ground to respond to flooding.
At Cloud to Street, our mission is to ensure that all vulnerable communities have the information they need to prepare for and respond to disasters. This means reducing scientific barriers and presenting accessible — and therefore actionable — information about flooding. The frequency and magnitude of flooding are increasing at an alarming rate, affecting growing populations of climate-vulnerable people. Flooding affects more people than any other environmental hazard and hinders sustainable development (for further reading, see: The Human Cost of Weather Related Disasters, Resilience of the Poor, Review Disaster Events). Our recently published research shows that relative property loss for floods are highest in the U.S. in places of high social vulnerability — specifically, locations with larger proportions of Hispanic, Black, and Native American communities that live below the poverty line.
We build flood mapping systems from remote-sensing data collected by commercial and public satellites and turn them into design tools. We have worked in fifteen countries, mostly in sub-Saharan Africa and Southern and Southeast Asia, building near real-time flood monitoring systems that show the extent of historical flooding events, current flood conditions, and who is affected — even when ground data is not available.
Our work provides novel insights and near real-time information to government leaders and disaster response coordinators to help communities on the ground. However, there are challenges to flood mapping, monitoring, and analyzing based on satellite remote sensing, as we do here at Cloud to Street. Satellites — the very technology that makes it possible to remotely create scalable, low-cost, and high-quality flood maps and impact assessments — have structured revisit times that do not capture every part of the earth every day, and clouds and their shadows can block them from imaging the ground. Beyond navigating adverse weather conditions and limitations in availability, traditional remote sensing approaches also require hand-tuned parameters and thresholding to achieve accurate results. These processes can often be painstaking and labor intensive; to circumvent this problem, we are exploring how to use machine learning not only to improve on our existing results, but also to provide new insights and improve access to accurate flood information on a larger scale.
Most machine learning research is focused on gains for the wealthiest populations and people. Focus on security, defense, and projects like self-driving cars and facial recognition consume immense resources but rarely benefit — and in fact, often harm — vulnerable communities. In order to address the immediate existential threat of climate change, we must redirect our energy and resources. There are immense gains to be made by applying even the most well-studied machine learning techniques to disaster response, as discussed below, and the possibilities for subsequent research and solutions are necessary and pressing areas of work that must be invested in. This is beginning: NeurIPS and other large AI conferences are promoting research into climate change, organizations such as Radiant Earth are taking steps to leverage machine learning to respond to disasters, and researchers are taking steps to identify distinct consequences of climate change and leverage cutting edge technologies to combat them. Dr. Catherine Nakalembe, for example, recently won the 2020 Africa Flood Prize for her machine learning work in assessing food security and smallholder agriculture. As researchers, environmentalists, and activists, we must come together to redirect our resources, creativity, and expertise to address this existential threat.
At Cloud to Street, we have already made significant gains simply by implementing traditional machine learning algorithms in a remote-sensing context. Using a Random Forest classifier, we are able to detect the presence of water through thin clouds that traditional remote sensing algorithms consistently miss.
We have also been able to use Random Forest classifiers to pinpoint where croplands actually are, creating a database for the Republic of Congo with an F1 accuracy rate of .94, up from the publicly available Global Food Security-support Analysis Data (GFSAD)’s .20. With this newly improved data, the World Food Project Congo was able to secure $12.5 million in multilateral aid for flood response and recovery in just eight months.
Using a Fully Convolutional Neural Net (FCNN), we are able to successfully perform semantic segmentation on images of flood events to produce novel gains in accuracy. We published a paper and a hand-labeled training dataset of flood images, in which we describe our methodology as well as the potential future applications of this technique. Insights and the associated datasets that spur algorithm innovation like these are groundbreaking; they significantly increase both the accuracy and reach of our work. This is just the tip of the iceberg; if we can achieve gains like this using existing machine learning algorithms, what gains in disaster preparedness and relief can be made once we invest time into developing techniques for these cases specifically?
These initial successes have enabled us to explore possibilities that, without the capabilities of machine learning, would have been prohibitively difficult. Can we use CNNs and crowdsourcing to develop an iterative semantic segmentation technique that will give us flood maps at the level of roadmaps, thus directly involving both remote sensing and communities on the ground? What gains in efficiency can be made using semantic segmentation with weak supervision in data-sparse areas? Can we use Generative Adversarial Networks to augment coarse satellite images and help fill key data gaps in urban areas? There is an abundance of satellite data sources available; is there a way to construct super-resolution images on which we can perform novel segmentation tasks irrespective of weather conditions? Can microwave, radar, and optical imagery be combined to construct novel early warning systems? This is an enormous area of research, the surface of which has barely been scratched.
It is imperative that we invest in technologies and research to develop climate resilience. At Cloud to Street, I am working on a project to incorporate crowd-sourced data into semantic segmentation models to reduce the immense manual overhead of generating hand-labeled training data. Our hope is that this project will allow us to make better flood maps with semantic segmentation models that are more universally applicable. These maps will provide novel insights in urban areas where the majority of populations and assets exist (and flood mapping using only remote sensing struggles the most), and involve communities directly with disaster reporting. We are also working on productionizing and stress testing the algorithms discussed earlier, as well as vetting a number of new projects for which we see immense potential. Fighting climate change will take a historic effort–and a historic amount of collaboration between communities, researchers, and policymakers. Machine learning has the power to play a (literal) life-saving role in that process; we simply need to redirect our energy and resources to address this universally existential threat.
Veda Sunkara is a Software and Data Engineer at Cloud to Street interested in developing fairly applied and thoughtfully designed models to real-world applications that address the critical resource disparities which characterize the growing disproportionate effects of climate change.
