Multiverse Computing

Analyzing deforestation in the Amazon Basin with machine learning

This project uses machine learning techniques to analyze data regarding deforestation in the Amazon basin. In a recent blog post, Ajitesh Kumar wrote:

Forest degradation is a major problem that contributes to climate change. It occurs when forests are damaged or destroyed, typically as a result of human activity. Deforestation, forest fires, pests, and poor forest management are all major causes of forest degradation. The loss of trees and other vegetation can have a significant impact on the local climate. In addition to absorbing carbon dioxide, trees, and plants help to regulate temperature and precipitation levels. When forests are degraded, these important functions are compromised, contributing to climate change…. In addition, forest degradation can also cause soil erosion, which can lead to desertification and loss of biodiversity. Therefore, it is essential to address the problem of forest degradation in order to protect our environment.

In principle, deforestation can be monitored automatically using high-resolution satellite imagery. At a Kaggle competition in 2016, the project “Planet: Understanding the Amazon from Space” evaluated Machine Learning algorithms for tracking human footprint in the Amazon rainforest (Luis Di Martino posted solutions to Part 1 and Part 2 of this problem; also see Nick Condo’s work on this project summarized here and repo here). This year’s M2PI project aims to update this solution to make a machine-learning algorithm using tensor network techniques inspired by quantum many-body physics.

Proposal

The solutions explained in the above links employ standard off-the-shell convolutional neural networks (CNNs) such as ResNet. We propose to (hopefully) improve the solution by replacing vanilla CNNs with “Tensorized” CNNs (TCNNs), or at least develop a detailed benchmarking of the performance and accuracy of TCNNs vs. CNNs for this problem. The basic ideas underlying TCNNs are explained here.

Skills

Required
  • Deep learning techniques, especially CNNs
Preferrable
  • Familiarity with low rank tensor factorization and tensor networks

The best skill match would be familiarity with this paper’s general ideas, particularly some of the references cited under the subcategory CNNs in Table 1.

Resources

In addition to the links above, the data set from the original Kaggle competition is available here.

Mehdi Bozzo-Rey
Mehdi Bozzo-Rey
Chief Revenue Officer
Steven Rayan
Steven Rayan
Professor of Mathematics, Director of Centre for Quantum Topology and its applications (quanTA)
Kristine Bauer
Kristine Bauer
Committee Chair
Li Xing
Li Xing
Assistant Professor
C Shijia Yu
C Shijia Yu
Graduate Student
Youssef Mousaaid
Youssef Mousaaid
Postdoctoral Researcher
Kimathra Reddy
Kimathra Reddy
Undergraduate Student
Santanil Jana
Santanil Jana
Ph.D Candidate