1. Introduction
Computational Sustainability (CS) seeks to balance environmental, economic, and social needs by developing optimization and machine learning methods that support sustainable decision-making.
This project focuses on leveraging remote sensing data (satellite imagery and related datasets) to produce high-resolution, low-cost socioeconomic and environmental maps, with a particular emphasis on scalable algorithms rooted in submodularity and combinatorial optimization.
The broader goal is to transform the growing availability of satellite imagery (e.g., from Planet Labs, SkyBox, SpaceX Starlink) into actionable insights for sustainability domains such as poverty mapping, pollution monitoring, and transportation planning, where ground-truth data is scarce or expensive to collect.
2. Research Gap and Novelty
Prior work has shown that neural networks on satellite imagery can predict poverty indicators (e.g., Jean, Science, 2016; Yeh., Nature, 2020), and that remote sensing supports pollution mapping (e.g., Duncan et al., Environmental Science & Technology, 2014), the key challenges remain:
Scalability: Current methods rely heavily on deep learning with extensive ground-truth data. These models are computationally expensive and struggle to adapt across regions with limited labeled data.
Interpretability: Deep learning approaches often provide little transparency in feature selection, which is critical for policy adoption in sustainability contexts.
Efficient Use of Limited Ground Data: Ground-truth surveys (e.g., World Bank LSMS) are sparse, costly, and unevenly distributed. Methods to optimally select where and what data to collect remain underdeveloped.
Novelty of this Project: Introduce submodularity-based optimization into the remote sensing for sustainability pipeline. Submodular functions, with their property of diminishing returns, are well-suited for:
Data subset selection (choosing the most informative satellite images or features to annotate/ground-truth).
Sensor placement & survey design (optimizing where to deploy costly ground measurements).
Interpretability (since submodular optimization often selects representative and diverse features, offering insights into spatial drivers of poverty or pollution).
Integrate nonnegative matrix factorization (NMF) with submodular priors to generate interpretable latent factors from high-dimensional satellite data.
Develop a low-cost pipeline that requires minimal ground data but achieves comparable or better predictive performance than purely deep learning-based approaches.
This combination of remote sensing + submodular optimization + interpretable factorization has not yet been systematically applied to computational sustainability, filling a research gap between “black-box” predictive models and actionable, resource-aware methods.
3. Project Objectives
Design scalable submodularity-based algorithms for:
Selecting informative subsets of satellite images for poverty/pollution estimation.
Guiding sparse survey collection to complement imagery.
Integrate remote sensing with interpretable optimization methods (NMF + graph-theoretic models) for socioeconomic mapping.
Benchmark against state-of-the-art methods on poverty prediction (DHS, LSMS datasets) and air quality mapping.
Demonstrate applications in Smart Cities:
Poverty and inequality mapping at neighborhood-level resolution.
Traffic congestion monitoring using road network submodularity.
Urban air pollution estimation.
4. Methodology
Input: Multi-modal satellite imagery (optical, SAR, nightlights) + sparse socioeconomic surveys (LSMS, DHS) or pollution sensors.
Output: Scalable, high-resolution poverty/pollution/transportation maps that are interpretable, resource-efficient, and cost-effective.
Methods and Tools
Submodular Optimization: For data subset selection, sensor placement, and interpretable feature discovery (Bilmes, 2022).
Graph Theory: To model spatial and network interactions (e.g., traffic networks, spatial poverty clusters).
Nonnegative Matrix Factorization (NMF): To extract interpretable latent factors from high-dimensional imagery.
Remote Sensing Analysis: Handling multi-spectral and temporal satellite data streams.
References:
Bilmes, (2022). Submodularity in Machine Learning and Artificial Intelligence. arXiv:2202.00132.
Jean,, et al. (2016). Combining satellite imagery and machine learning to predict poverty. Science, 353(6301), 790–794.
Yeh, et al. (2020). Using publicly available satellite imagery and deep learning to understand economic well-being in Africa. Nature Communications, 11(1).
Duncan, et al. (2014). Satellite data of atmospheric pollution for U.S. air quality applications. Environmental Science & Technology.
World Bank LSMS Initiative. https://www.worldbank.org/en/programs/lsms