OpenNEX: Science-guided applications of machine learning in the Earth sciences

Xiaowei Jia (July 26, 2018)

Growth in the world’s population and the acceleration of industrialization are straining already scarce natural resources and food supplies, which must scale up to keep pace with growing demand. Large-scale changes driven by these complex factors have led to tremendous stress on the availability and distribution of these resources at a global scale. Hence, it is essential to create systems and analytic capabilities that can provide the much needed spatially and temporally explicit information about changes at a global scale.

Advances in Earth observation technologies enables the acquisition of vast amounts of accurate, timely, and reliable Earth system data that can be used for monitoring changes at a global scale. Scientists from a number of disciplines have increasingly been using state-of-the-art machine learning (ML) techniques to monitor human activities and the impact that they are having on the environment. However, existing ML methods face multiple challenges in this task: 1) ML algorithms are commonly used as “black-box” method and are limited in their ability to interpret the results. 2) The existing machine learning techniques cannot be used to jointly analyze multi-scale remote sensing data with different spatial and temporal resolutions. 3) The data heterogeneity across space and time limits the ability of traditional ML algorithms to achieve high performance for years or regions not used for training.

In this talk, we will introduce a customized deep learning approach for analyzing spatio-temporal remote sensing data. We demonstrate the effectiveness of our approach in several important applications including deforestation in Southeast Asia and cropland monitoring in US. Moreover, we will explore the combination of the laws of physics and traditional ML models to better capture environmental changes over space and over time.