Results with
their context.
The same Sentinel-2 contexts, NEON references and masks for every model. The selected HyperSR v1 pipeline uses both fine-tuned SEN2SR fusion stages.
| Model | PSNR ↑ (dB) | SSIM ↑ | MAE ↓ | RMSE ↓ | SAM ↓ (rad) |
|---|---|---|---|---|---|
| Bicubic interpolationMeasured | 28.90 dB27.04 dB–31.29 dB | 0.7440.672–0.823 | 0.02750.0219–0.0322 | 0.03910.0309–0.0461 | 0.05140.0395–0.0625 |
| SEN2SR LiteMeasured | 28.23 dB26.33 dB–30.78 dB | 0.7480.685–0.821 | 0.02970.0235–0.0347 | 0.04240.0333–0.0495 | 0.05910.0444–0.0727 |
| SEN2SRMeasured | 27.63 dB25.71 dB–30.22 dB | 0.7370.677–0.811 | 0.03190.0251–0.0372 | 0.04540.0357–0.0530 | 0.06750.0510–0.0833 |
| SRGAN + SEN2SR LiteMeasured | 28.07 dB26.17 dB–30.65 dB | 0.7330.671–0.809 | 0.03030.0240–0.0353 | 0.04310.0339–0.0503 | 0.06090.0458–0.0747 |
| SRGAN + SEN2SR (released)Measured | 27.37 dB25.45 dB–29.98 dB | 0.7040.645–0.783 | 0.03290.0259–0.0383 | 0.04670.0368–0.0546 | 0.06880.0520–0.0849 |
| LDSR + SEN2SR LiteMeasured | 28.18 dB26.28 dB–30.74 dB | 0.7480.685–0.822 | 0.02990.0237–0.0349 | 0.04260.0334–0.0498 | 0.05940.0446–0.0732 |
| LDSR + SEN2SRMeasured | 27.61 dB25.69 dB–30.19 dB | 0.7340.673–0.808 | 0.03200.0252–0.0373 | 0.04550.0358–0.0531 | 0.06800.0514–0.0841 |
| HyperSR v1Measured | 30.36 dB28.39 dB–32.80 dB | 0.7730.705–0.847 | 0.02350.0186–0.0283 | 0.03340.0260–0.0406 | 0.04380.0342–0.0522 |
225 tiles. Eight independent groups.
One deterministic 128 × 128 Sentinel-2 context per test tile, compared with a 512 × 512 NEON reference. The score excludes a 32-pixel high-resolution border and invalid observations. The six evaluated bands are B05, B06, B07, B8A, B11 and B12.
Means weight the eight independent groups equally, across nine sites. Intervals use 2,000 group-bootstrap resamples. Eight groups limit the strength of uncertainty estimates. PSNR uses a reflectance range of one; MAE and RMSE use reflectance units, and SAM uses radians.
This is a local SEN2NEON crop protocol. It is not a reproduction of the published full-tile benchmark. F2 training checks describe 10 m outputs; their scores are not substituted for the combined pipeline's 2.5 m results. Shared research-GPU timings are not controlled speed benchmarks.
The JSON includes per-band, site and land-cover results. Qualitative example scenes carry their own split labels and acquisition dates. Output pixel spacing does not establish native observation.