Deep Learning Architectures for Landsat 9 Pan-Sharpening

The 30 m sampling of the Landsat 9 Operational Land Imager-2 (OLI-2) multispectral channels (B2, B3, B4) and the 15 m sampling of the panchromatic channel (B8) bound the spatial resolution at which surface reflectance can be quantitatively retrieved, including over topographically complex, mountainous, and agricultural landscapes where field- and slopescale features are often smaller than the native multispectral pixel. Five convolutional architectures—DCNN, PanNet, ResASPP, DenseNet, and Attention-Sharpen—were benchmarked for reconstructing the multispectral cube at panchromatic resolution, following the reduced-resolution variant of Wald's protocol, with the multispectral stack degraded by 1/4 and the panchromatic channel by 1/2. A 2048×2048 extent of scene LC09_L1TP_175034_20251109 was partitioned into 1024 non-overlapping 64×64 patches. Each network was trained for 20 epochs with the L1 loss, Adam optimisation (lr = 1×10⁻³), and a batch size of 8. Quality was assessed with four full-reference metrics (PSNR, SSIM, SAM, ERGAS) and three noreference metrics (QNR, Dλ, Ds). PanNet, by restricting its learning manifold to the high-frequency residual of the upsampled multispectral cube, attained PSNR = 46.47 dB, SSIM = 0.98, SAM = 0.617°, ERGAS = 1.394, and QNR = 0.9921, outperforming all baselines. Pairwise Wilcoxon signed-rank tests rejected the null hypothesis of equal performance against each baseline at p < 0.05. A timing analysis exposed an approximately sevenfold gap between the fastest network (DenseNet, 1.49 s/epoch) and the heaviest (AttentionSharpen, 10.98 s/epoch), motivating an application-conditioned modelselection guideline for operational Earth-observation pipelines serving highresolution land and landscape monitoring.