Stone to Sea

Let me show you the tide
that comes for us faster
than we’d like to admit.

Let me show you
airports underwater
bulldozed reefs, blasted sands
and plans to build new atolls
forcing land
from an ancient, rising sea,
forcing us to imagine
turning ourselves to stone.

Rise By Kathy Jetñil-Kijiner and Aka Niviâna

The land masses inhabited by 0.001% of the population are being swallowed by the seas because of carbon emissions the other 99.99% of the population has been releasing into the atmosphere for the past 50 years.

The Marshallese make up a part of that 0.001% of the population. They are no longer talking about mitigating effects of climate change. They have to survive the catastrophic effects of climate change. They do not have any more time. They have no choice but to build upwards on existing land, since relocation is not an option.

Poet and activist Kathy Jetñil-Kijiner uses her voice and words to create verbal snapshots of a changing landmass, her home, being consumed by rising waters. Satellite data gives us a perspective of this violent change from above and portraits of indigenous peoples allow us to see who is directly affected.

Technical Implementation

Using training data from NASA’s LandSat satellite imagery and a model set of FFHQ portraits (StyleGAN pre-trained model), we explored facial features amidst a changing landscape.

Through multiple phases of training, a creative blending of visual motifs across the two datasets are created. In the first phase of the training, the GAN is presented with a set of human portraits symbolizing the residents of Micronesian islander. In the second phase of training, satellite imagery of coastlines are introduced to the model representing the progress of sea over land.

The choice of when to switch from the first to the second phase of training, as well as the selection of training image examples and hyperparameter values were made based on their ability to produce creative results.

Our code runs in Python and uses GPU acceleration. Additional Python code was developed for selecting training images, filtering out noise, and transforming images to a standard format. Minor post-processing was performed on the output image to adjust the color and tone.

Stone to Sea (2019):

Maryam Ashoori
IBM Research AI

Oceane Boulais
MIT Media Lab

Brandon Leshchinskiy
MIT Media Lab