10 SEP 2025 (MON) 11:35 – 11:55

Long-term, High-spatiotemporal Resolution Monitoring of Global Land Cover Dynamics via Multi-satellite Data Fusion and Deep Learning

Mr CHEN Shuang 

( Supervisor: Prof Peng Gong )

Abstract:

Accurate, timely, and high-resolution monitoring of global land cover change is vital for climate modeling, biodiversity assessment, and sustainable resource management. Yet, achieving both fine spatial detail and frequent temporal coverage remains challenging due to limitations of individual satellite sensors.

This study develops a suite of data fusion and deep learning frameworks to overcome these challenges, enabling global land cover monitoring at 10–30 m resolution with near-daily frequency.

First, a Robust Optimization-Based Temporal fusion (ROBOT) model was introduced to integrate Landsat (30 m) and MODIS (500 m) data, generating the Seamless Data Cube (SDC30) of daily global 30-m reflectance (2000–2024). ROBOT reduced reconstruction errors by >20% and achieved over 1000× computational efficiency gains, with strong validation across 425 global sites (MAE = 0.014).

To harmonize higher-resolution sensors, the Spectral and Spatial Harmonization Model (SSHM) aligned Landsat and Sentinel-2 observations, reducing cross-sensor spectral discrepancies by 5–22% and reflectance errors by 25–71% across spectral bands.

Building on these data foundations, a Surface Water Integrated Monitoring (SWIM) framework was developed, applying weakly supervised deep learning to monitor near-daily surface water dynamics. SWIM achieved user’s/producer’s accuracies of 96.8–96.9% and 95.7–97.9%, successfully capturing transient events such as floods and droughts.

Finally, this study produced FROMGLC30, a global annual 30-m land cover dataset spanning 2000–2024, with an overall accuracy of 81.3%. This dataset provides unprecedented spatiotemporal consistency for tracking long-term global land cover change.

Overall, this research advances Earth observation by delivering innovative methods and high-quality datasets (SDC30, FROMGLC30), offering powerful tools for global change studies and decision-making.