26 JAN 2026 (MON) 14:35 - 15:05

Towards Seamless High Resolution Global Land Surface Parameter Products Generation

Mr GUAN Shikang  

( Supervisor: Prof Shunlin Liang )

Abstract:

Land surface parameters derived from satellite observations have found broad applications in diverse sectors, including meteorology, agriculture, forestry, and transportation, where they play a critical role in informing development policies. Driven by advances in high-performance computing and the rapid evolution of artificial intelligence, there is a growing urgency for medium- and high-resolution remote sensing data to support emerging applications. The development and production of the High-resolution Global Land Surface Satellite (Hi-GLASS) product enable scientific communities and government agencies to proactively trace historical states of the Earth and to predict future transitions. However, due to the assumptions inherent in the underlying algorithms, certain Hi-GLASS parameters fail to fully recover signals obscured by clouds and cloud shadows, resulting in limitations for practical applications. To address this issue, further efforts are necessary in spatial gap-filling and time-series reconstruction of Hi-GLASS products. The core objective is to overcome the algorithmic assumptions and production inefficiencies associated with traditional inversion methods by leveraging remote sensing foundation models (RSFMs). This study enhances the practical value of Hi-GLASS and lays the groundwork for advancing very high-resolution remote sensing inversion technologies.

Spatiotemporal fusion (STF) is an effective approach to achieving the above objectives. However, existing STF methods have limited application scenarios due to their stringent input data requirements and restrictive algorithmic assumptions. Therefore, this research aims to combine STF with RSFMs to overcome theoretical limitations and enable the reconstruction and super-resolution of land surface parameters at medium- and high-resolution. Specifically, the research will focus on the following three objectives: (1) to develop an advanced model capable of accommodating cloudy observations and extended time series inputs. This model relaxes the strict alignment and clear-sky constraints typically required of input data pairs, thereby enabling STF methods to be applied at larger spatial scales. At this stage, surface reflectance serves as the supervised target to demonstrate the feasibility of the proposed approach. (2) to replace supervised learning with self-supervised learning to achieve gap-filling and time-series reconstruction of Hi-GLASS. At this stage, the model is designed to generate outputs with the same temporal length as the inputs while simultaneously retrieving multiple variables, ensuring the efficient and scalable production of high-level remote sensing products. (3) to integrate the above two strategies to complete the production of global surface reflectance and high-level remote sensing products, including Land Surface Temperature, Albedo, and Fractional Vegetation Cover, among others.