Real-time generation of gap-free MODIS leaf area index product from 2000 to 2024 using a deep learning method

• A novel TSCGF framework generates gap-free MODIS LAI products in real-time. • Generated spatiotemporally continuous LAI products from 2000 to 2024 globally. • BiTCN model effectively reconstructs missing LAI using historical temporal patterns. • Temporal stability improved by 50–85% while preserving authentic vegetation dynamics. • Enhanced spatiotemporal continuity in cloud-prone regions with complex terrain. Leaf area index (LAI) is a crucial biophysical parameter to characterize the canopy structure and energy absorption capacity of vegetation in Earth system science. While MODIS LAI products are widely used in ecological research, their retrieval algorithms process each pixel and data independently, resulting in high noise levels and time series discontinuities that limit their application. In this study, a Three-step Spatio-temporal Constrained Gap Filling (TSCGF) method for generating high-quality MODIS LAI products in real time is proposed. Our approach consists of three main steps: first, the Spatial Similarity-based Gap Filling (SSGF) uses local neighborhood information within a 50 km × 50 km window to identify similar pixels based on phenological patterns and land cover types, and implements a correlated weighted average scheme to generate a complete LAI time series. Next, globally representative sampling pixels are generated by performing time-series K-means clustering on the pre-filled LAI data within each MODIS tile. Finally, a bidirectional time Convolutional network (BiTCN) model is trained on globally distributed samples, taking advantage of these rich time series to reconstruct high-quality LAI values by capturing long-term time dependencies while maintaining computational efficiency. Since 2000, this method has implemented real-time MODIS LAI datasets on a global scale. Validation of DIRECT2.1 (100 sites, 231 measurements) and GBOV (52 sites, 8423 measurements) ground measurements for different biomes showed that our approach achieved better performance compared to the original MODIS LAI product, with a decrease in root mean square error from 0.95 to 0.88. In areas with frequent cloud cover and complex terrain, TSCGF significantly enhanced spatio-temporal continuity while maintaining true vegetation dynamics, further demonstrating its effectiveness. This approach provides a robust framework for real-time generation of high-quality LAI products that can better support a variety of ecological applications.