Using Deep Learning to Detect Illegal Sand Mining in India
Joshua Blumenstock, Professor
Information, School of
Applications for Spring 2024 are closed for this project.
Sustainable sand mining is one of the most pressing ecological challenges currently facing the planet. After water, sand is the world’s most valuable natural resource – over 50 billion tons of construction-grade sand and gravel were mined globally from rivers and beaches in 2021, accounting for 85% of all mineral extraction. As the major ingredient of concrete and asphalt, sand is vital to economic growth, and will play a key role in aiding the transition to a low carbon society, especially in the Global South.
Over the past decade, the rapid proliferation of satellites and other digital sensors has created tremendous opportunities to measure human behavior. These data also provide the foundation for the growing interdisciplinary field of computational social science. Our team has pioneered the use of non-traditional "big" data for tackling some of the world's most important policy questions. See https://globalpolicy.science/
Furthering this research agenda, this project is focused on using satellite imagery (and other data sources) to measure and monitor sand mining and resource extraction in India. Excessive sand mining in India’s riverbeds has threatened local wildlife, accelerated the erosion of rivers, and amplified the effects of climate change. It is also associated with increased violence and disruption of local livelihoods. While monitoring and regulation is the need of the hour, efforts have been hindered by a severe lack of data.
Our current focus is on developing machine learning tools to detect mining activity in high-resolution satellite imagery. In collaboration with trusted partners, we have manually annotated thousands of images, spanning 39 river basins in India, to label the images that capture signatures of sand mining. Using these data, we aim to develop and train a custom machine learning model to detect sand mining activity using satellite imagery.
We are uniquely positioned to be able to answer important and pressing questions, such as: Can we leverage recent advances in semi-supervised learning (SSL) to effectively detect sand mines in medium resolution satellite imagery
Once we have a dataset of detected mining activity, can we then use this data to examine the socio-economic and environmental consequences of mining activity?
Role: The URAP apprentice will primarily be expected to:
- Assist in building and maintaining the machine learning pipeline and associated codebase
- Assist with the gathering and labelling of ground-truth data
conduct preliminary data analysis as and when required (statistical and descriptive)
- Depending on the student’s interest, there may be the opportunity to explore methods related to machine learning, causal inference, etc., as well as to develop new estimation and deep learning frameworks.
- The undergraduate would be expected to be able to complete well-defined research tasks, but will also have the freedom to propose and pursue promising directions of their own interest.
Qualifications: All URAP apprentices, irrespective of the specific role or assignment, are expected to be extremely self-motivated, attentive to detail, and meticulous in their approach to data analysis. Apprentices must be able to work independently and be excited to take responsibility and initiative for ensuring their work meets exacting standards of quality.
Specific qualifications for this position include:
- Proficiency in Python (required)
- Experience manipulating large data sets (required)
- Ability to derive insights from data and communicate these clearly (required)
- Knowledge of statistical/econometric methods (preferred)
- Proficiency with visualization tools (preferred)
- Knowledge/ prior experience working with deep learning models - ViTs/ CNNs (preferred)
- Experience dealing with geospatial datasets (preferred)
Day-to-day supervisor for this project: Ando Shah, Ph.D. candidate
Hours: to be negotiated
Related website: https://globalpolicy.science/
Related website: https://www.jblumenstock.com