A research team led by Professors Zhang Yongjun and Li Yansheng from Wuhan University (WHU) has made significant strides in distributed collaborative learning for remote sensing foundation models (RSFMs).  

Their findings were published in The Innovation (IF=33.2) under the title Unleashing the potential of remote sensing foundation models via bridging data and computility islands. Professor Li Yansheng is the first author, with Professors Zhang Yongjun and Ant Group researcher Yang Ming as corresponding authors. The study involved collaborations with institutions including the Russian Academy of Sciences and Ant Group. 

The framework for cross-cloud collaborative training and collaborative inference of RSFMs for generalist EO intelligence.

To address the challenges of decentralized remote sensing data and computing power, the team introduced FedSense, the world's first privacy-preserving distributed collaborative pretraining framework for RSFMs. This framework combines contrastive learning and masked reconstruction self-supervised learning while incorporating an innovative federated mutual guidance learning method. By enabling bidirectional optimization between servers and clients, FedSense reduces communication overhead and improves model performance.

Schematic of the FedSense distributed collaborative pretraining framework for RSFMs and multi-institution remote sensing pretraining dataset.

Using this framework, the team successfully conducted large-scale collaborative pretraining on millions of remote sensing images from 10 institutions, significantly boosting performance in eight key tasks, including scene classification and object detection.  

The research team plans to further explore secure, reliable multi-party collaborative pretraining technologies to create an interconnected remote sensing data space and advance general intelligence for Earth observation.