Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Scaling inference across tiles and comparing results visually

Everything so far has processed one 128-by-128 patch. This page scales the same workflow to larger regions, and adds a qualitative comparison against genuine high-resolution basemap imagery, with an important caveat about what that comparison can and cannot prove.

1. Motivation

Real projects rarely need just one small patch enhanced. Knowing how tiling works here, and knowing how to sanity-check output against real high-resolution imagery without over-trusting that check, is what makes this workflow usable beyond a single demonstration patch.

2. Core idea

For any region larger than LDSR-S2’s native 128-by-128 input size, geoai.super_resolution() automatically tiles the input into overlapping patches, runs the encode-denoise-decode pipeline on each independently, and stitches the results back together using linear blending across the overlap regions, the same tiled inference idea you have now seen across detection, segmentation, and here, image translation.

3. Workflow

A. Run tiled super-resolution

sr_large = "sr_large.tif"
sr_large_img, _ = geoai.super_resolution(
    input_lr_path=s2_path,
    output_sr_path=sr_large,
    rgb_nir_bands=[1, 2, 3, 4],
    window=(700, 1300, 256, 256),
    patch_size=128,
    overlap=16,
    sampling_steps=100,
)

print(f"Input shape: (4, 256, 256) at 10 m")
print(f"Output shape: {sr_large_img.shape} at 2.5 m")

patch_size=128 matches the model’s native input size from two pages ago; overlap=16 controls how many pixels adjacent patches share. Larger overlap values produce smoother transitions between patches at the seams, at the cost of more total patches to process and more computation time, the familiar quality-versus-cost trade-off from earlier pages.

B. Compare the larger region

geoai.plot_sr_comparison(s2_path, sr_large, bands=[1, 2, 3])
plt.show()

This is the same comparison function from the single-patch workflow, now applied to the full stitched output, letting you visually confirm the tiling did not introduce obvious seams or discontinuities at patch boundaries.

C. Compare against a basemap

geoai.create_split_map(
    left_layer=sr_large, right_layer="Esri.WorldImagery", left_args={"vmax": 0.3}
)

create_split_map() produces an interactive split-panel map: dragging the slider reveals how well the super-resolved output’s fine-scale detail matches what a genuinely high-resolution basemap shows in the same location.

D. This comparison is qualitative only

4. Python reactivation

Nothing new syntactically here; this page reuses the same function-call and dictionary patterns from earlier pages, applied to a larger window and a couple of new parameters (patch_size, overlap).

5. Common pitfalls

6. Mini task

You compare your super-resolved output against the Esri basemap and notice a building present in the basemap but missing from your Sentinel-2-derived output. List two possible explanations, and how you would distinguish between them.

Sample solution

One possibility is that the building was constructed after the Sentinel-2 scene’s acquisition date but before the basemap’s, a genuine temporal difference rather than a model error. Another possibility is that the building is small enough, or has low enough contrast in the original 10-meter input, that the super-resolution model failed to reconstruct it. To distinguish between these, you would check the acquisition dates of both sources, and, if a genuine discrepancy remains after accounting for timing, examine the uncertainty map from the previous page at that specific location, since a model that failed to reconstruct a real feature would likely show elevated uncertainty there.

7. Key takeaways

Spatial Data Analytics
Interpreting results
Spatial Data Analytics
Project transfer