Fusion of UAV–Sentinel‑2 images using convolutional neural networks for remote-sensing analysis

Remote sensing has consolidated the use of satellite imagery for agricultural applications such as crop monitoring, which drives the development of multisensor integration approaches aimed at improving spatial resolution and radiometric quality. In this study, a UAV–Sentinel‑2 fusion methodology was implemented through spectral compositions based on seven vegetation indices. This methodology enabled the evaluation of four target spatial resolutions (0.5, 1, 1.5 and 2 m) through three super‑resolution models: FSRCNN, VDSR, and ESRGAN, independently trained for four acquisition dates. Multiscale analysis allowed the identification of 2 m as the optimal fusion resolution, by simultaneously maximising radiometric fidelity and structural preservation metrics. At this scale, VDSR demonstrated the best performance (RMSE = 0.041; PSNR = 33.83 dB; SSIM = 0.983; Correlation = 0.981). The dispersion analysis of metrics across bands showed greater reconstruction uniformity at coarser resolutions, while the inter‑dates evaluation revealed high consistency of the three models across four acquisition dates at 2 m (CV < 12.9% for RMSE, <4.33% for PSNR, <1.24%, for SSIM and <1.25% for correlation). Overall, the results validate the proposed methodological approach and highlight the importance of balancing spatial enhancement and spectral coherence in UAV–Sentinel‑2 fusion processes for quantitative agricultural applications and precision biophysical monitoring.