A two-stage framework for enhancing MODIS Gross Primary Production (GPP) from 500 m to 20 m using Sentinel-2 vegetation information and environmental variables, without requiring spatially dense high-resolution GPP ground truth.
Gross Primary Production (GPP) is widely used to characterize ecosystem carbon uptake, but MODIS GPP is available at a relatively coarse 500 m spatial resolution. This study investigates whether Sentinel-2 vegetation information and environmental variables can reconstruct spatially detailed GPP at 20 m without requiring a spatially complete high-resolution GPP target.
Unlike conventional image super-resolution, fine-resolution GPP does not have a spatially dense ground-truth target. This research addresses the resulting learning and validation challenge by connecting coarse MODIS productivity estimates with fine-scale Sentinel-2 observations and ecological knowledge.
Formulates GPP downscaling as an estimation problem constrained by coarse productivity information and fine-resolution environmental observations, rather than relying on unavailable 20 m GPP labels.
Limited-label LearningIntroduces a progressive auxiliary-to-main learning framework designed to transfer vegetation-related fine-scale structure to GPP reconstruction while retaining information from observed MODIS GPP.
Cross-scale LearningIntegrates coarse-scale consistency and vegetation–productivity relationships to support plausible spatial patterns, rather than optimizing image sharpness alone.
Physical RelevanceThe approach combines MODIS GPP with Sentinel-2 vegetation information and environmental context to estimate 20 m productivity patterns. Its central research contribution is learning and evaluating spatial detail when pixel-wise 20 m GPP ground truth is unavailable. The architecture, supervisory relationships, and specific training objectives are restricted to the detailed methodology.
Independent validation uses three flux towers: US-Ro4, US-Ro5, and US-Ro6. Among the compared estimates, PR-Aux provides the strongest aggregate performance, while PR-Main also improves the overall relationship relative to the original MODIS GPP.
PR-Aux achieves RMSE 27.7095 and R² 0.6420, compared with RMSE 29.6649 and R² 0.5332 for MODIS GPP. The presentation reports a 6.59% reduction in RMSE and a 20.41% improvement in R² relative to MODIS GPP.
RMSE 27.71 · Validation across 3 flux towers
California validation uses seven flux-tower sites: US-Bi1, US-Bi2, US-Ton, US-Tw1, US-Tw4, US-Tw5, and US-Var. In this region, PR-Main provides the strongest aggregate performance among the evaluated methods.
PR-Main achieves RMSE 30.5925 and R² 0.5720, compared with RMSE 39.2453 and R² 0.5471 for MODIS GPP. The presentation reports a 22.04% reduction in RMSE and a 4.55% improvement in R² relative to MODIS GPP.
R² 0.572 · Validation across 7 flux towers
The spatial analysis focuses on limitations of coarse MODIS GPP, including missing or weakly represented productivity patterns in non-vegetated and heterogeneous areas. High-resolution Sentinel-2 reference information enables finer spatial variation, while the ecological relationship loss discourages implausible GPP values.
This work was presented at the 2025 IEEE International Geoscience and Remote Sensing Symposium (IGARSS).
The research overview, contributions, and preliminary evaluation are public. Model architecture, training procedure, supervisory design, and loss specifications remain password protected while the manuscript is in preparation.