Komorebi: Predicting Deforestation

David Dao, Ming-Da Liu, Catherine Cang, Johannes Rausch


Designing effective REDD+ policies, assessing their GHG impact, and linking them with the corresponding payments, is a resource intensive and complex task. Komorebi leverages video prediction with remote sensing to monitor and forecast forest change at high resolution.

Paper Draft (Climate Change AI ICML'19)


1. We train on Hansen dataset and segment forest cover using a U-Net architecture LANDSAT satellite imagery.

2. We receive the forest loss by subtracting adjacent covers.

3. We train a video prediction model to predict n frames given an initial number of frames.

4. Last, we segment forest cover from the predicted frames to receive a forest cover forecast.

Preliminary Results

Semantic Segmentation

Left: Forest Loss (Hansen), Right: Segmented Forest Loss (U-Net)

Video Prediction

Left: LANDSAT 2013-2017, Right: Video Prediction of LANDSAT

Video Prediction + Semantic Segmentation

Left: Forest Loss (Hansen), Right: Forecasted Forest Loss