Anonymization of Private and Confidential Ground Data for Earth Observation Data Analytics such as Socioeconomic Data
David A. Petit, Elisabeth Petersen, David D. Smith, Andrew Harvey, Anna Burzykowska, Rogério Manuel Lemos Pereira Bonifácio · 2024
This article examines recent research utilizing Earth Observation to forecast socioeconomic data. It proposes a novel and pragmatic definition of data anonymization, which uses differential privacy and synthetic data generated with a CTGAN. By extracting features from satellite imagery, machine learning models such as neural networks can be trained on anonymized datasets to predict socioeconomic data using Earth Observation.The research examines the effects of anonymization in a hypothetical scenario to assess the capacity of a model trained on anonymized datasets to deduce the connection between private data and imagery and, thus, the risk of information leakage. While the demonstration of the concept is based on socioeconomic data, the underlying principles remain applicable to various forms of private and confidential data that are gathered locally (e.g., from IoT sensors and mobile devices) and utilized as labels to train EO models.