Wu et al. (2025) STC-DeepLAINet: A Transformer-GCN Hybrid Deep Learning Network for Large-Scale LAI Inversion by Integrating Spatio-Temporal Correlations

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-12-17
• Authors: Huijing Wu, Ting Tian, Qingling Geng, Hongwei Li
• DOI: 10.3390/rs17244047

Research Groups

• School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou 450001, China
• School of Geoscience and Technology, Zhengzhou University, Zhengzhou 450001, China

Short Summary

This paper introduces STC-DeepLAINet, a Transformer-GCN hybrid deep learning network, for high-precision, large-scale Leaf Area Index (LAI) inversion by effectively integrating spatio-temporal correlations. The proposed network significantly outperforms existing methods and generates reliable LAI products crucial for agricultural and ecological research.

Objective

• To address the limitations of existing LAI inversion methods in efficiently extracting integrated spatial-spectral-temporal features and effectively modeling spatio-temporal dependencies, which compromise the accuracy of LAI products.
• To develop a novel deep learning framework capable of autonomously capturing large-scale spatio-temporal correlations to improve the accuracy of remote sensing-based LAI inversion.

Study Configuration

• Spatial Scale: China, with a spatial resolution of 500 meters.
• Temporal Scale: Training (2019–2020), Validation (2021), Testing (2022–2024). Data has an 8-day temporal resolution.

Methodology and Data

• Models used:
  • Proposed: STC-DeepLAINet (Transformer-GCN hybrid deep learning network incorporating a 3D CNN-based spectral-spatial embedding module, a spatio-temporal correlation-aware module, a spatio-temporal pattern memory attention module, and a knowledge-guided loss function).
  • Comparison: Random Forest (RF), Generalized Regression Neural Networks (GRNN), Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Attention-Enhanced LSTM (AELSTM), Graph Neural Network-Recurrent Neural Network (GNN-RNN), Transformer, 3D CNN-LSTM.
• Data sources:
  • Satellite: MODIS LAI (MOD15A2H), VIIRS LAI (VNP15A2H), GLASS LAI V6, MODIS Surface Reflectance Product (MOD09A1) across seven spectral bands.
  • Observation: Ground-based LAI measurements from 12 sites across China (2022–2024) obtained from the National Ecosystem Science Data Center, National Tibetan Plateau Data Center, and LAI-2200 Plant Canopy Analyzer.
  • Ancillary: 2024 MODIS Land Cover Type product (MCD12Q1.061).

Main Results

• STC-DeepLAINet achieved superior performance in 500 m LAI inversion across China (2022–2024), with an average coefficient of determination (R²) of 0.96, a root mean square error (RMSE) of 0.39, and a bias of 0.07.
• Compared to the baseline Transformer, STC-DeepLAINet reduced RMSE by 43%.
• Direct validation against ground-based LAI measurements yielded an R² of 0.827 and an RMSE of 0.718, outperforming the GLASS LAI product.
• The network effectively mitigated the LAI saturation effect in high LAI scenarios (e.g., forests and croplands), producing more detailed and accurate LAI distributions than existing products.
• Ablation studies confirmed the individual contributions of the Temporal Correlation (TC), Spatial Correlation (SC), Spatio-Temporal Pattern Memory Attention (MAN), and Knowledge-Guided Loss Function (KLF) modules to the overall performance improvement.
• STC-DeepLAINet demonstrated strong tolerance to cloud/shadow noise, maintaining a maximum RMSE of 0.8 even with 50% pixel contamination, which is lower than the RMSE of mainstream LAI products under ideal conditions.

Contributions

• Proposed STC-DeepLAINet, a novel Transformer-GCN hybrid deep learning architecture, for large-scale LAI inversion by explicitly integrating spatio-temporal correlations.
• Introduced a spatio-temporal correlation-aware module that models temporal dynamics by "time periods" and spatial heterogeneity by "spatial slices," enhancing the capture of complex spatio-temporal dependencies.
• Designed a spatio-temporal pattern memory attention module to dynamically fuse spatio-temporal features and retrieve historically similar patterns, improving inversion accuracy and adaptability to complex vegetation ecosystems.
• Developed a knowledge-guided loss function (KLF) that adaptively weights samples based on physiological thresholds, effectively mitigating the saturation effect in high LAI inversion.
• Provided an operational framework for generating high-precision, large-scale LAI products, offering reliable data support for agricultural yield estimation and ecosystem carbon cycle simulation, and a new methodological reference for spatio-temporal correlation modeling in remote sensing.

Funding

• National Key Research and Development Program of China (Grants 2024YFF1308201).

Citation

@article{Wu2025STCDeepLAINet,
  author = {Wu, Huijing and Tian, Ting and Geng, Qingling and Li, Hongwei},
  title = {STC-DeepLAINet: A Transformer-GCN Hybrid Deep Learning Network for Large-Scale LAI Inversion by Integrating Spatio-Temporal Correlations},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17244047},
  url = {https://doi.org/10.3390/rs17244047}
}