Landsat visual imagery at 3.75 m Fidelity Reconstruction (FR) from Landsat-8/9, and 7.5 m from Landsat-7 and the deeper archive.
Request accessLandsat's decades-deep record, reconstructed: 3.75 m visual from Landsat-8/9 and 7.5 m from Landsat-7 and the easier archive — evidence-bounded, never hallucinated.
Landsat is the longest continuous record of the Earth's surface. ETD applies Fidelity Reconstruction to its visual bands — 3.75 m GSD from Landsat-8 and 9, and 7.5 m from Landsat-7 and the deeper archive — recovering only the structure the source genuinely implies, so the reconstruction stays faithful to the historical evidence. It extends a consistent, comparable visual record back across decades for change and baseline work where the modern archive alone is too shallow.
Fidelity Reconstruction is a reconstructive, non-hallucinating approach to super-resolution. Rather than generating a plausible high-resolution image the way adversarial (GAN) upscalers do, an FR model performs a learned, information-bounded reconstruction: it recovers high-frequency structure that is genuinely implied by — aliased or blurred within — the input, and refines edges and fine geometry, but it does not synthesize novel content (textures, objects, or detail) that the source cannot support.
The defining, measurable property is downscale-consistency: when an FR output is reduced back to the input resolution, it returns the input almost exactly (in our tests ~56 dB PSNR), because every pixel is derived from real evidence. Generative upscalers fail this test (~30 dB) precisely because they add content that isn't in the source — content that looks sharp but is fabricated. That distinction is what makes FR appropriate for measurement and monitoring, where an invented building, road, or texture would be an analytical liability, not a cosmetic bonus.
The deliberate trade-off is that FR will not manufacture the “photographic crispness” a GAN conjures; its ceiling is the true information content of the input (and the resolution its training targets taught it to resolve). In exchange, its output is trustworthy and reproducible — sharper and higher-resolution than the source, but faithful to it — which is the correct guarantee for a satellite-analytics product.