Single-Band NDVI Spatial Downscaling: Synthetic Benchmarking and Landsat-8/Sentinel-2 Cross-Sensor Evaluation

Medium-resolution NDVI products can smooth small vegetation patches, land-cover boundaries, and local gradients. This study presents a two-part evaluation of single-band NDVI spatial downscaling: a controlled synthetic Sentinel-2 benchmark and a separate transfer assessment on real Landsat-8/Sentinel-2 imagery. The contribution is the evaluation design and evidence; we assess, but do not propose, the published EDSR architecture. We compared EDSR with bicubic interpolation, SRCNN, and VDSR. Training used 2200 precomputed Sentinel-2 NDVI scenes; validation used 330 geographically separate Anhui patches; and testing used 300 Zhengzhou patches. Synthetic LR inputs were generated from NDVI using a fixed 3× bicubic operator. The cross-sensor set comprised 389 co-located image pairs acquired on 15 and 17 May 2023, with 117 pairs used for shared linear calibration and 272 held out. EDSR achieved 35.70 dB PSNR and 0.920 SSIM on the synthetic test set. On held-out cross-sensor pairs, its mean PSNR was 25.58 dB before and 26.40 dB after calibration, gains of 0.61 and 0.62 dB over bicubic, respectively. After calibration, the paired EDSR–VDSR PSNR difference was 0.232 dB (95% bootstrap CI, 0.203–0.263). These findings describe performance under the two evaluated protocols; the cross-sensor results apply to the available paired acquisitions.