10 SEP 2025 (MON) 11:05 - 11:25

Spatio-temporal Continuous LAI Estimation Based on the Remote Sensing Foundation Model

Ms GU Xiangfeng 

( Supervisor: Prof Shunlin Liang )

Abstract:

LAI is one of the terrestrial essential climate variables in the Global Climate Observing System and a significant indicator for model simulation of terrestrial ecosystems, estimation of crop yields, and monitoring of vegetation changes. Research on LAI estimation with high temporal resolution and spatial continuity is often restricted by clouds in remote sensing imagery. Based on extensive core features related to remote sensing imagery, Remote sensing foundation models could provide new methodology for the downstream tasks of imagery pixel reconstruction and spatio-temporal continuous LAI estimation, which can be achieved with only a small amount of data fine-tuning the pretrained model. Compared to other models, Prithvi foundation model offers advantages such as the comprehensive and unified pre-training dataset, a classic and stable model architecture, clearly defined downstream tasks, and suggested applicability in regression tasks.

In this research, we initially investigate the Prithvi ’s capabilities to estimate the numerical pixel values of images, take the LAI estimation task for example. Utilizing these capabilities, the fine-tuned Prithvi achieves reflectance reconstruction for cloud-affected pixels and subsequently enables spatial-temporal continuous LAI estimation. The Prithvi contains 100 million parameters and utilizes the 3D Vision Transformer downstream architecture, fine-tuned with 30 m resolution Harmonized Landsat and Sentinel-2 multi-band spectral data and Hi-GLASS LAI products in China region. Through three experiments: LAI estimation, cloudy pixel value reconstruction, and two-stage model verification, the downstream regression model of Prithvi is optimized to reconstruct the reflectance affected by clouds and estimate the LAI values with a resolution of 30 m and shorter temporal interval.

Research show that the fine-tuned downstream regression models can complete the pixel-level regression tasks, demonstrating the feasibility of the Prithvi foundation model on LAI estimation and cloudy pixel reflectance reconstruction tasks. The single-phase LAI estimation model can reconstruct the macroscopic spatial pattern, with a RMSE of 0.20 in validation. With multi-temporal phases inputs, the cloudy pixel value reconstruction model can restore reflectance features achieving a RMSE of 0.39, but lose the high-frequency details and grid artifacts. To optimizes detail restoration and grid artifacts, the loss function modification has been an effective path in the downstream model. The evaluation indicators of the combined LAI estimation and cloud pixel value reconstruction model are affected by the two-stage sequence, with the best optimized RMSE being 0.69. The two-stage model improve the spatial and temporal continuity of LAI estimation, achieve spatio-temporal reconstruction of 30m resolution LAI, effectively correct the numerical missing caused by cloud, detect temporal anomalies and provide a method for cloudy area LAI estimation and anomalies correction in vegetation dynamic monitoring. The optimization logic, dominated by loss function optimization, architecture adjustment and post-processing compensation enhances the model's detail capture ability and improves the LAI frequency and quality. This method provides an important reference for the construction of regression downstream models, helping researchers optimize models and obtain more accurate and reliable prediction results.