Generating a 1 km nationwide long-term NDVI dataset for China (1982–2015) via an enhanced vegetation monitoring super-resolution network

The Normalized Difference Vegetation Index (NDVI) serves as a crucial indicator for assessing extended vegetation trends. Existing 1 km MODIS NDVI products start from 2000, leaving a gap in high‑resolution observations for 1982–2000 and limiting systematic studies of vegetation change across China. To fill this spatio-temporal void, we built a 1 km nationwide NDVI time series dataset covering 1982–2015 by downscaling the 5 km AVHRR NDVI with a High‑Performance Super‑Resolution (HPSR) method. The proposed network architecture is composed of three key submodules that work sequentially and synergistically: (i) A multi-scale feature extraction network first extracts texture information across different receptive fields through parallel convolutions at varying scales, achieving multi-scale fusion via feature splicing. (ii) The resulting fused features are then passed to an attention refinement network that utilizes the Convolutional Block Attention Module (CBAM), incorporating channel and spatial attention mechanisms, to adaptively weight the fusedu features, thereby emphasizing significant vegetation regions while reducing irrelevant noise. (iii) Finally, the refined features are fed into a super-resolution network that utilizes up-sampling modules and residual blocks to reconstruct the feature maps to a high resolution. The model spatially downscales AVHRR NDVI data from an approximately 5 km resolution to 1 km, utilizing high-quality 1 km MODIS NDVI as a reference dataset, and integrating DEM terrain factors and the China Land Cover Database (CLCD) as ancillary data. The experimental findings indicate that the HPSR framework effectively mitigates the numerical centralization common in standard convolutional models through an integrated attention mechanism, achieving a significant reduction in MAE (0.050) and RMSE (0.061) while maintaining a PSNR above 27 dB. By ensuring a holistic enhancement of NDVI sequences from high spatial fidelity to long-term physical and logical consistency, the model successfully preserves the original signal entropy and statistical distribution characteristics across complex landscapes. Consequently, by resolving the traditional trade-off between spatial resolution and temporal coverage, this study produced a 1 km nationwide NDVI dataset for China (1982–2015), furnishing a consistent, high-quality baseline for decadal-scale ecological monitoring and biophysical analysis.