Generated flooded house image from the paper Visualizing the Consequences of Climate Change Using Cycle‑Consistent Adversarial Networks
We are pleased to share that two of our research papers were presented at the AI for Social Good Workshops held at the International Conference on Learning Representations (ICLR) 2019 in May and the International Conference on Machine Learning (ICML) 2019 in June. Both works explored how artificial intelligence could help address challenges associated with climate change, from making its potential consequences more tangible to improving how we monitor environmental conditions and support decision-making in agriculture and disaster management.
The first paper, Visualizing the Consequences of Climate Change Using Cycle‑Consistent Adversarial Networks, explored how AI‑generated images could be used to help people better understand the potential effects of climate change on places familiar to them. The research used Cycle‑Consistent Adversarial Networks (CycleGANs), a type of generative model, to transform images of houses and streets into visual representations of how they might look after flooding. By learning from images of flooded and non‑flooded locations, the model could generate these transformations without requiring before‑and‑after images of the same location. The aim was to make the potential consequences of climate change more concrete and personal.
The work combined these generated images with flood projections derived from climate models, with an initial focus on visualizing potential flooding around 2050. The resulting prototype illustrated the feasibility of this approach: about 70% of the 80 test images were successfully mapped to realistically flooded houses. Performance was strongest for single‑family suburban houses surrounded by grass, and further work was needed to improve image quality and account for a wider range of environments and climate‑related events. The authors also noted that climate models operate at a regional resolution, much coarser than individual households, and that their outputs are difficult to translate into photo‑realistic images, which motivates exploring physical constraints in GAN training. The broader vision was to develop an interactive tool that could help people explore possible climate futures and understand how individual choices and policy decisions might influence them.
We were particularly fortunate to contribute to this work alongside researchers from the Montreal Institute for Learning Algorithms (Mila), including Yoshua Bengio. Collaborating with researchers working at the forefront of machine learning on a problem with such significant societal implications was a valuable experience for our team, and we are proud to have contributed to this effort to explore how emerging AI techniques could support public understanding of climate change. The paper was authored by Victor Schmidt, Alexandra Luccioni, S. Karthik Mukkavilli, Kris Sankaran, and Yoshua Bengio from Mila, Narmada Balasooriya from ConscientAI, and Jennifer Chayes from Microsoft Research.
Our second paper, AI for Mitigating Effects of Climate and Weather Changes in Agriculture, presented a proposal for an AI‑powered platform to support disaster management and agriculture in Sri Lanka, along with preliminary work on its core modules. The proposed system would use remote sensing data and aerial imagery, with machine learning and computer vision, to identify features such as land cover, land use, water levels, and crop stage. These features would then feed predictive models. The long‑term vision was to help improve disaster preparedness, crop monitoring, and agricultural planning through more timely, data‑driven decisions. The paper was authored by Narmada Balasooriya, C.D. Athuraliya, and Janak Gunatilleke from ConscientAI.
These two projects reflect an early part of our journey exploring how AI research can contribute to real‑world challenges. They also illustrate two complementary approaches: using generative models to communicate the possible consequences of climate change, and using data‑driven systems to help communities and decision‑makers respond to environmental risks.
Read the first paper below and access the second paper here. You can also find out more about ConscientAI’s research on our Research page.