ARTIFICIAL INTELLIGENCE-DRIVEN INTEGRATION OF SATELLITE, IN-SITU, AND SIMULATION DATA TO ENHANCE UNI-TEMPORAL AND MULTI-TEMPORAL VARIABLE ESTIMATION

Abstract

Reliable estimation of environmental variables is critical for understanding land–atmosphere interactions, ecosystem functioning, and the global carbon cycle, and for supporting applications such as natural hazard prediction, resource management, and sustainable development. However, current environmental datasets derived from satellite observations, in-situ measurements, and process-based ecosystem models have inherent limitations. Satellite observations provide broad spatial coverage but may be affected by observational constraints and the indirect nature of satellite measurements. In-situ observations provide highly accurate measurements but are spatially sparse and often temporally incomplete. Process-based models simulate ecosystem dynamics with physical interpretability but often depend on simplified assumptions. Therefore, effectively integrating these complementary data sources remains a key challenge for Earth system monitoring.This dissertation develops artificial intelligence (AI)-driven integration frameworks that combine satellite, in-situ, and simulation data, particularly through knowledge-guided machine learning (KGML), to improve environmental variable estimation. The proposed frameworks address both uni-temporal and multi-temporal variable estimation tasks, with two representative case studies: land surface temperature (LST) and gross primary productivity (GPP), where LST serves as a concrete example of uni-temporal (snapshot) variable estimation, while GPP serves as a concrete example of multi-temporal (time-series) variable estimation. The first study conducts a systematic evaluation of existing satellite-based LST products and ecosystem model GPP simulations using satellite observations, model outputs, and in-situ measurements. The analysis identifies key limitations in current environmental datasets, including the strong dependence of LST retrieval on surface properties and the sensitivity of GPP estimation to vegetation composition. The second study applies a machine learning (ML)-based framework for improving snapshot retrieval of LST. A transfer learning (TL) approach integrates radiative transfer simulations (RTMs) with real-world observations. By combining shortwave and longwave top-of-atmosphere (TOA) observations, the proposed TL method reduces reliance on external emissivity inputs and improves retrieval accuracy compared with conventional single-channel (SC) and split-window (SW) algorithms. The third study proposes a KGML framework for time-series estimation of GPP. The framework introduces a decomposition-and-resembling (DERE) strategy to link process-based ecosystem simulations with satellite-derived vegetation composition observations. In addition, a probabilistic label expansion module based on diffusion modeling is introduced to address the limited temporal coverage of flux tower observations. Experiments across multiple observational networks demonstrate that the proposed DERE framework improves estimation performance for GPP compared with conventional baseline and KGML algorithms. Overall, results show that the proposed methods consistently improve estimation accuracy compared with existing approaches across both tasks, highlighting the effectiveness of AI-driven and knowledge-guided data integration. By integrating satellite, in-situ, and simulation data, the proposed approaches improve environmental variable estimation and contribute to more reliable Earth system monitoring.