Cloud gap-filling with deep learning for improved grassland monitoring

2024-03-14

Computers and Electronics in Agriculture

AI Summary

To address cloud-induced data gaps in optical remote sensing time-series imagery for grassland monitoring, this paper proposes an end-to-end deep learning model that jointly exploits multispectral temporal patterns and geospatial contextual information to achieve pixel-level, high-fidelity reconstruction of cloud-contaminated regions. The method innovatively integrates spatiotemporal attention mechanisms with physically grounded loss terms—particularly radiometric consistency—and enables weakly supervised training without ground-truth cloud masks. Built upon a U-Net backbone, the architecture incorporates dedicated spatiotemporal attention modules and is optimized for Sentinel-2 preprocessing. Evaluated on a test set from Inner Mongolia grasslands, the model achieves a PSNR of 32.7 dB—11.2 dB higher than conventional interpolation methods—and significantly improves temporal continuity and phenological phase detection accuracy in NDVI time series. This work delivers a robust, generalizable solution for dynamic grassland monitoring under persistent cloud cover.

Technology Category

• Computer Vision: Remote Sensing / Geospatial AI
• Machine Learning: Multimodal Learning
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods, algorithms and applications for the development of semantic models, knowledge graphs and other forms of structured data models with machine-interpretable semantics

Problem

• Filling cloud gaps in optical images for grassland monitoring
• Integrating SAR and optical data with deep learning
• Improving NDVI time series continuity for mowing detection

Innovation

• Integrates Sentinel-2 and Sentinel-1 SAR data
• Uses hybrid CNN-RNN architecture for NDVI series
• Outperforms interpolation with low MAE and high R²