A field note on reconstruction, spatial information loss, and why AI-enhanced satellite imagery doesn’t always do what you expect.
TLDR
- Mapping urban vegetation is harder than it looks, especially in dense Arabian cities where greenery is fragmented and linear rather than broad and contiguous.
- The most important green features are very narrow. Palm-lined medians, roundabouts, and highway verges, the dominant form of Gulf urban greenery, are typically less than 10 metres wide.
- Sentinel-2’s 10-metre resolution can’t reliably capture them. These features are at best one pixel wide, meaning they are easily missed or blended with surrounding surfaces.
- Standard remote sensing tools are mismatched to Arabian urban green structure, creating systematic blind spots in vegetation mapping for the region.
The Setup
Mapping urban vegetation from satellite imagery is harder than it looks. In dense Arabian cities, the green that matters most, the palm-lined medians irrigated roundabouts, the highway verges, is packed into strips often less than 10 metres wide. At the standard 10-metre resolution of Sentinel-2, those features are at best one pixel wide. They blur. They merge with the asphalt on either side. Vegetation becomes invisible in remote sensing products.
The promise of deep learning super-resolution is precisely this: take a 10m image, feed it through a neural network trained on paired high-resolution and low-resolution datasets, and recover spatial detail that was never explicitly captured. The S2DR3 model, a commercial Sentinel-2 enhancement tool built on Google Earth Engine, claims to do exactly this, processing full Sentinel-2 MGRS tiles, integrating with local storage and Google Cloud, and delivering sub-pixel vegetation discrimination at scale.
We tested it on a stretch of Riyadh: the Eastern Ring Road and Makkah Road corridor, 24 tiles at 4×4 km each, using imagery from two acquisition dates (February 2022, February 2024 and August 2024) for the sensitivity analysis.
The experiment did not go as planned. And that, as it turns out, is worth writing about.
What Super-Resolution Was Supposed to Fix
Standard Sentinel-2 L2A imagery at 10m resolution captures vegetation as a fractional signal. A pixel flagged as “vegetated” in an NDVI-derived vegetation cover (VC) product might contain a palm tree, a shrub, a car bonnet, and a patch of sand, all mixed together. The NDVI value reflects a blended spectral response across that 10×10m footprint.
Road medians in Riyadh illustrate the problem sharply. Highways have raised medians typically 3–5 metres wide, planted with date palms and ornamental shrubs: real, maintained vegetation with genuine ecological and landscape value. At 10m resolution, those medians are sub-pixel features. Their spectral contribution bleeds into the adjacent road surface, where it is diluted. Depending on the NDVI threshold chosen, they either just barely register or disappear entirely.
Super-resolution was intended to upscale the image to an effective resolution of 1m, sharpen the spectral boundaries around these linear features, and enable the vegetation classifier to resolve them clearly. The idea was sound. The execution revealed a more fundamental problem, one that sits not with the model, but with the physics of what the original sensor actually captured.
Why the Original Signal Cannot Be Recovered
Single-image super-resolution is, formally, a severely ill-posed inverse problem. A coarse-resolution sensor pixel does not record a clean 10×10m box of the Earth’s surface. It records a spatially weighted integration of radiance across its footprint, described by the sensor’s Point Spread Function (PSF). This integration is irreversible: the within-pixel spatial distribution of surfaces, which sub-pixel is road, which is palm canopy, which is soil, is discarded entirely at the moment of acquisition. No algorithm can recover it from the output, because it was never encoded in the signal.
When a 1m super-resolution model produces 100 sub-pixels from a single 10m input pixel, it has no physical basis for deciding which of those 100 positions hold vegetation. The values it assigns are spectrally plausible given what the model has learned about how surfaces look, but they are geometrically synthesised rather than geometrically recovered. This is the distinction between reconstruction and hallucination, and it matters enormously for any application requiring spatial accuracy rather than visual plausibility.
The problem is compounded by scale. A 10m sensor and a 1m sensor do not capture the same scene at different resolutions: they capture fundamentally different datasets. In a homogeneous environment, such as a wheat field or a closed-canopy forest, a 10m pixel is a faithful representation of a larger, uniform patch, and the information loss from coarse sampling is limited. In a heterogeneous urban environment, a single 10m pixel may integrate road surface, tree canopy, building shadow, and irrigated grass into a single spectral value. The spatial frequencies encoding the boundaries between those materials sit well above the Nyquist limit of the 10m sensor and are simply not present in the data. No super-resolution algorithm, regardless of its sophistication, can reconstruct a spatial structure that was never sampled.
The Sub-Pixel Tree Problem
This theoretical limitation becomes tangible when the target feature is a tree.
Consider a date palm whose canopy covers 3m², occupying roughly 3% of a 10m pixel. The dominant spectral contribution to that pixel is non-vegetation: road, soil, or pavement. The pixel’s NDVI is suppressed well below any standard vegetation threshold. The SR model receives an input that contains no reliable signature of a tree, and it produces a 10×10 grid of 1m sub-pixels in which, most likely, none are classified as vegetation.
Even for a larger tree covering 40% of a pixel, the situation is not much better. The model knows the pixel has a mixed spectral composition, but it has no information about the spatial arrangement of materials within it. The 100 sub-pixels it generates might place a canopy on the left, a road on the right, or a canopy in the centre, or scatter them arbitrarily: all are consistent with the same input value. The correct spatial arrangement cannot be inferred from a single pixel, because there are no adjacent pixels carrying independent information about that tree’s boundaries.
For linear features that span multiple pixels, such as a road median extending across ten or twenty 10m pixels, the situation is somewhat better. Adjacent pixels provide cross-pixel context, and the model can use that context to constrain the boundary geometry. This is why SR performs better on continuous features than on isolated objects: there is a spatial signal to learn from. But for sub-pixel trees with no neighbours encoding their structure, the model interpolates based on global training priors, and the result is geometrically arbitrary.
The appropriate method for quantifying canopy cover in mixed pixels is not super-resolution but spectral unmixing: estimating the fractional abundance of end-member materials within each pixel. Unmixing cannot tell you where within a pixel the canopy sits, but it can tell you how much canopy is present, which is often the ecologically relevant quantity.
The Sensitivity Analysis: What the Numbers Say
To understand how much the NDVI threshold choice matters independently of the SR problem, we applied eight different VC thresholds to the same super-resolved scene (9 August 2024, tile T38RPN) and measured the resulting vegetation fraction across the study area. The results are stark:
Source: GDAL raster statistics from comparison outputs
Moving the threshold from 0.15 to 0.50 reduces the detected vegetation area more than sevenfold, from roughly 1-in-10 pixels to fewer than 1-in-73. This is not a marginal calibration question. It is a decision that fundamentally defines whether road medians, irrigated verges, and sparse desert shrubs exist in your dataset or not.
Results Table
| Threshold | 10m Median % | 10m Mean % | 10m SD | 1m Median % | 1m Mean % | 1m Mean % |
|---|---|---|---|---|---|---|
| 0.15 | 1.2545 | 1.7794 | 1.9333 | 3.2192 | 3.8334 | 2.3194 |
| 0.20 | 0.4776 | 0.8135 | 1.0532 | 1.9394 | 2.4255 | 1.6460 |
| 0.25 | 0.1706 | 0.3545 | 0.4902 | 1.2089 | 1.4880 | 1.0733 |
| 0.30 | 0.0458 | 0.1390 | 0.1946 | 0.6639 | 0.8075 | 0.5833 |
| 0.35 | 0.0088 | 0.0484 | 0.0852 | 0.2757 | 0.3614 | 0.2536 |
| 0.40 | 0.0003 | 0.0131 | 0.0274 | 0.0871 | 0.1252 | 0.0860 |
| 0.45 | 0.0000 | 0.0049 | 0.0152 | 0.0216 | 0.0341 | 0.0326 |
| 0.50 | 0.0000 | 0.0017 | 0.0054 | 0.0043 | 0.0086 | 0.0133 |
In an arid urban context like Riyadh, where genuine vegetation cover rarely exceeds 5–8% even in green corridors, the choice of 0.15 vs 0.30 is the difference between capturing the urban forest and erasing it. Any reported vegetation cover figure is only meaningful when the threshold used is explicitly stated. Reporting “vegetation cover” without specifying the NDVI cut-off is reporting an underdetermined quantity.
The Blurring Problem: Where Super-Resolution Failed
This is where the exercise produced its most instructive and frustrating result.
SR models are trained to maximise perceptual similarity between their outputs and real high-resolution images. In natural imagery, this means enhancing texture, recovering fine structure, and producing visually plausible detail. But the model has never seen a 3-metre road median in Riyadh. Its training distribution consists of globally distributed imagery in which vegetation appears as coherent patches: forests, agricultural parcels, parks. Not 1-pixel-wide strips of palms flanked by asphalt.
When the model encounters a sub-pixel linear feature with a weak NDVI signal, it does what any pattern-recognition system does: it interpolates based on its learned priors. The result is spectrally reasonable, but the spatial structure is hallucinated rather than resolved. The road medians, those narrow vegetated ribbons that were the entire motivation for the exercise, emerged from the SR model as wider, fuzzier features. The boundaries between vegetation and pavement were less defined in the 1m super-resolved output than in the original 10m image.
Applied to the VC mask, this produced a classic false-positive pattern: vegetation detected well into the road carriageway, aligned with the median axis but geometrically inflated. The detected vegetation polygons are spatially correct (aligned with the Eastern Ring Road median network) but geometrically incorrect, extending well beyond the actual planted area into the road surface.
This is not a failure of S2DR3 in general. The model appears to perform well on its intended use cases: large-area agricultural and forest monitoring where its training distribution is a reasonable match for the target. It is a failure of fit between the model’s training distribution and the target application, compounded by the fundamental information bottleneck described above.
Is Multi-Temporal Stacking a Path Forward?
A natural question follows: if a single 10m image lacks the spatial information for SR, could stacking multiple temporally close acquisitions help? The answer is: partially, and for specific feature types.
Multi-Image Super-Resolution (MISR) is a more physically rigorous approach than single-image SR. Successive Sentinel-2 acquisitions over the same tile are never perfectly co-registered: slight orbital variations and sub-pixel phase shifts between passes mean each image samples the scene from a marginally different geometric position. When these shifts are sub-pixel in magnitude, they encode genuinely distinct spatial information. Stacking and aligning them allows recovery of spatial frequencies that are aliased in any single acquisition, which is a real, physically grounded information gain.
Recent work on the Sen4x architecture, a hybrid combining MISR sub-pixel phase recovery with learned priors from Pléiades Neo training data, achieves 2.5m effective resolution from Sentinel-2 stacks, validated against WorldView-2. Separately, the L1BSR approach exploits sub-pixel shifts within Sentinel-2’s own overlapping CMOS detector segments, requiring no external acquisitions at all.
However, MISR has hard limits that are relevant here. For cross-pixel linear features such as road medians spanning multiple 10m pixels, MISR can meaningfully sharpen boundaries by exploiting phase shifts across adjacent pixels. For isolated sub-pixel trees, the Nyquist constraint still applies: there is no spatial frequency encoding a 3m canopy present in any single Sentinel-2 acquisition, and no number of stacked images can reconstruct what was never sampled by the sensor. MISR recovers aliased-but-present information; it cannot create information that was below the optical cut-off of the instrument.
Closely spaced acquisitions (5–10 cloud-free scenes within a single season) would minimise phenological change between images, which is the correct approach for vegetated urban scenes. This is a viable and worthwhile experiment for the road median mapping problem, with realistic expectations: sharper median boundaries are achievable; reliable sub-pixel tree detection is not.
What This Means for Vegetation Mapping Practice
The experiment produced four substantial lessons:
- Threshold choice is a substantive scientific decision, not a technical detail. The seven-fold range in detected VC across the threshold sweep means any vegetation cover figure is only meaningful when the threshold used is explicitly stated.
- Single-image SR does not improve vegetation detection accuracy for sub-pixel features. For isolated urban trees and narrow linear features, SR enhancement introduces geometric hallucination artefacts that inflate the apparent size and boundary of vegetation patches. The spatial resolution gain comes at the cost of geometric fidelity.
- The information bottleneck is at the sensor, not the algorithm. A 10m sensor and a 1m sensor collect fundamentally different datasets in heterogeneous urban environments. No post-processing algorithm can bridge that gap from the coarse side alone.
- Validation against known reference features is non-negotiable. The road median network served as an ideal validation dataset: linear features of known width, known vegetation type, and known location. When SR-derived VC polygons extend well into the road carriageway, the SR process has failed for this feature class, regardless of what aggregate statistics suggest.
Recommendation
The exercise surfaces a clearer path forward:
- Use 10m native resolution for area-level vegetation cover statistics and relative change detection (2022 vs 2024), where the blurring artefact affects all thresholds equally.
- Apply spectral unmixing to estimate fractional canopy cover within mixed pixels, the more appropriate method for isolated sub-pixel trees.
- Trial MISR with 5–10 temporally close, cloud-free Sentinel-2 acquisitions for road median boundary sharpening, where cross-pixel context makes phase recovery meaningful.
- Apply commercial high-resolution imagery (Pléiades or WorldView at 0.5–1.5m) for definitive road median mapping, where genuine sensor resolution resolves the feature
- Calibrate the VC threshold against field-validated reference plots: the road medians provide a spatially constrained, independently mappable calibration dataset
- Revisit single-image SR for larger patch contexts (parks, greenbelts, palm groves) where the model’s training distribution is a closer match for the target features
The S2DR3 model remains a promising tool at scale. At €0.8/km² and 13–14 hours of computing for full Riyadh coverage, the economics are viable for appropriate applications. The question is not whether to use super-resolution, but where its spatial enhancement claims are physically supportable.
Data sources: Sentinel-2 L2A, tile T38RPN, acquisition dates February 2022, February 2024, and August 2024. Super-resolution processing via S2DR3 (Gamma Earth). Vegetation cover extraction and threshold sensitivity analysis: 8 VC masks, thresholds 0.15–0.50. Validation: road network shapefile, Eastern Ring Road and Makkah Road corridor, Riyadh.
Leave a Reply