Monitoring Illicit Rare Earth Mining in Myanmar via Self-Supervised Learning

Heavy Rare Earth Elements (HREEs) are critical for the production of most electronic devices. Rapidly increasing demand for these minerals has led to a proliferation of highly polluting makeshift HREE extraction in Myanmar. Monitoring the spread of these mines is important for the preservation of human health and the environment. This paper utilizes a geospatial foundation model pre-trained using self-supervised learning to detect hundreds of rare earth mines using a single template example. This is achieved through the development of a novel method for embedding similarity search-- Cosine Contrast-- which leverages both positive and negative templates to yield more relevant results.