Title
Towards Foundation Models for Agricultural Remote Sensing: From Spatial-Temporal-Spectral Pre-training to Multimodal Agricultural Intelligence
Abstract
Foundation models are reshaping computer vision and multimodal intelligence, yet agricultural remote sensing remains challenging due to complex spatial-temporal dynamics, heterogeneous data sources, and limited cross-region generalization. Traditional task-specific pipelines are often isolated, labor-intensive, and difficult to scale.
In this talk, I will present our recent efforts toward foundation models for agricultural remote sensing. First, I will introduce A2-MAE, a spatial-temporal-spectral self-supervised pre-training framework for large-scale multimodal Earth observation data, enabling unified representation learning across heterogeneous satellite imagery and downstream geospatial tasks. Next, I will present our multimodal agricultural intelligence benchmarks, including AgroMind for agricultural scene understanding and AgroCoT for chain-of-thought reasoning evaluation.
Our studies show that existing large multimodal models still exhibit substantial limitations in agricultural perception and reasoning tasks. I will further introduce AgroNVILA, a domain-adapted agricultural vision-language model enhanced through fine-tuning and reinforcement learning strategies, demonstrating strong performance in crop monitoring, scene understanding, and spatial reasoning. Finally, I will discuss how next-generation supercomputing infrastructure supports large-scale agricultural foundation models and conclude with perspectives on open and trustworthy AI for sustainable agriculture and food security.
Bio
Juepeng Zheng is an Associate Professor at the School of Artificial Intelligence, Sun Yat-sen University, and a jointly appointed Research Scientist at the National Supercomputing Center in Shenzhen. His research focuses on large-scale geoscientific intelligent computing, with applications in disaster prevention and mitigation, and natural resource management.
His work aims to deeply integrate artificial intelligence methods with domestic supercomputing platforms to enhance Earth system observation and forecasting capabilities, including weather and climate prediction, and large-scale remote sensing interpretation. Dr. Zheng has published more than 40 papers as first or corresponding author in leading journals and conferences across both geoscience and artificial intelligence, including RSE, ISPRS P&RS, IEEE TGRS, ICML, ICLR, NeurIPS, and CVPR. His publications have received over 2,000 citations on Google Scholar.
He has received the Wu Wenjun Artificial Intelligence Science and Technology Progress Award (Second Prize) and the nomination award for the Best Application in Chinese Supercomputing. His student teams have also won multiple national competition awards in AI and high-performance computing.