New AI Research Introduces Proxy Consistency Loss for Earth Observation Data Fusion

Researchers have developed a novel machine learning method called proxy consistency loss (PCL) to address data scarcity in Earth observation applications. The approach integrates correlated geographic variables as proxies through a trainable location encoder, enabling learning from abundantly available proxy data independently of limited labeled training samples. A key innovation involves using the location encoder as a flexible mechanism to incorporate these proxy variables, which are often correlated with but distinct from target variables of interest. The method also requires appropriate regularization of the location encoder to achieve robust performance with sparse labeled data. Experimental validation has been conducted on air quality prediction and poverty mapping tasks, demonstrating the technique's practical applications. The research addresses fundamental limitations in supervised learning with Earth observation inputs, where high-quality labeled or in-situ measured data remains scarce despite geographic data abundance. This work represents a technical advancement in geographic AI methodologies.