From pixels to planning: Earth AI for nature restoration (research.google)

Google Research has unveiled a groundbreaking high-resolution deep learning framework that identifies fine-scale ecological features like hedgerows and copses, which traditional satellite detection typically overlooks. This new, vectorized dataset serves as a vital tool in addressing the pressing climate and biodiversity crises while ensuring food security. By revealing the often-invisible woody features interspersed within agricultural landscapes, this innovation opens a pathway for enhancing carbon storage and supporting biodiversity without competing for valuable arable land. The technology behind this advancement involves overcoming significant technical challenges, particularly those associated with spatial topology, semantics, and computational scale. Leveraging the Remote Sensing Foundations’ Vision-Transformer pre-trained on over 300 million satellite images, the model was fine-tuned with a limited annotated dataset to recognize specific features of the British countryside. Innovative solutions, such as a dual-layer labeling system and the Polsby–Popper compactness score, were implemented to accurately categorize and map complex spatial elements. Additionally, by utilizing Google Earth Engine to parallel process data across extensive geographical areas, the team successfully created a comprehensive ecological inventory. This open-access dataset aims to empower farmers, scientists, and policymakers in their efforts to restore nature, tackling environmental challenges effectively while preserving agricultural capabilities.