AI Forest Maps: Read the Pixel Size and the Height Error

A forest canopy-height map offers an inviting view of a complicated landscape. A coloured surface turns uneven crowns into something that can be explored, compared and discussed. Zoom in far enough and the detail can suggest that every tree has been individually measured.

Before using that map in an article, a woodland project or a business presentation, pause over the numbers in its description. A small pixel size tells you about spatial detail. A height-error figure tells you how predictions compared with reference measurements. Both matter, and they answer different questions.

How AI adds height to an image

In a study published on 28 September 2023, Nico Lang and colleagues produced a global canopy-top-height map for 2020 at a ground sampling distance of 10 metres. Their approach combined Sentinel-2 optical imagery with sampled height information from NASA’s GEDI spaceborne lidar, using probabilistic deep-learning models.

The authors distinguish the 10-metre grid from its effective resolution. Each pixel represents the largest canopy-top height within an approximately 25-metre GEDI footprint centred there. A finely spaced grid therefore should not be read as independent measurements of every tree.

Read two different kinds of metres

Think about a hypothetical map divided into one-metre cells. That spacing describes positions across the ground. It does not establish that the predicted canopy heights are accurate to one metre vertically. Likewise, adding more displayed detail would not, by itself, demonstrate a better height estimate.

Jamie Tolan and colleagues’ study, published in Remote Sensing of Environment in 2024, illustrates the distinction. It produced maps for California and São Paulo using satellite imagery with a ground sample distance below one metre, with models trained against aerial lidar and a GEDI-based correction. Table 2 reports a mean absolute error of 2.8 metres for the corrected self-supervised model, averaged across three validation datasets.

Mean absolute error averages the sizes of the prediction errors without letting positive and negative errors cancel. It is a summary of a particular evaluation. It neither promises that every location lies within that distance of its reference value nor sets a guaranteed maximum error.

Ask where the error was tested

The same table in Tolan and colleagues’ paper is more informative than the headline average alone. For that corrected model, mean absolute error was 2.7 metres on the NEON test dataset and 5.1 metres on the São Paulo dataset. The reported 2.8-metre overall figure was an unweighted average across the three datasets.

For a reader considering a particular woodland, this raises a useful question: how closely do the validation locations resemble the place being discussed? A result from one evaluation cannot automatically settle performance at a new site. Look for the geography, vegetation and reference data behind the score before adopting it as a local expectation.

Lang and colleagues also supplied uncertainty estimates whose calibration varied between biomes; regional applications might need recalibration. Ask what an uncertainty layer represents and how its reliability was checked.

Keep the observation dates attached

A research publication date, a data-release date and an observation date belong in separate boxes in your notes. The Lang map described above represents 2020, despite the paper’s 2023 publication. A recently downloaded file can still describe an earlier landscape.

Imagine a creator preparing an article about a community woodland in 2026. If a canopy map represents 2020, it can contribute historical context. It cannot alone establish what currently stands on the site. The sensible next step is to find appropriately dated evidence and explain which period each source describes.

Comparisons over time need another pause. Before labelling a difference as growth or loss, establish whether the products use comparable definitions, methods and coverage. A changed prediction is an observation about the outputs; explaining a real-world change requires supporting evidence.

Canopy height leaves more to establish

Height and biomass are different quantities. NASA’s ORNL DAAC guide to the GEDI Level 4A biomass-density product, revised on 2 September 2026, describes predictions of aboveground biomass density and associated prediction standard errors for sampled laser footprints. Its methodology links waveform height metrics with field-plot biomass estimates through calibrated models.

That product guide shows the additional modelling involved. A canopy-height map alone does not complete the job of estimating biomass. Any move from a forest’s visible structure to a statement about its mass needs an appropriate method, relevant calibration and an account of uncertainty.

For the hypothetical woodland article, a careful description might say that a historical map suggests where taller canopy was estimated. Claiming that the same map establishes the woodland’s current carbon stock would require substantially more evidence. Keep the public wording within what the analysis actually supports.

Give the map a specific job

A useful first job might be identifying places worth investigating. In our hypothetical project, the creator could use contrasting height estimates to organise questions for a woodland manager: which areas were planted at different times, which have changed, and which mapped features merit a closer look? These are proposed reporting steps, not findings about an actual woodland.

The map then becomes part of a documented investigation. Keep the dataset version, observation period, measurement definition and validation notes beside the exported image. If the story changes after a site visit or additional records, readers should be able to follow why.

The same discipline applies to other research-led stories. MaryChuks.com’s guide to critically reading a research paper offers a practical way to separate the study’s question, methods, findings and limits before building an argument around it.

Check these details before sharing

• The quantity: Is the map describing canopy-top height, another height metric or biomass density?
• The dates: Which period do the underlying observations and finished product represent?
• The scale: What does the pixel spacing mean, and what height errors were reported?
• The evaluation: Where was performance checked, against which references, and how were results averaged?
• The intended claim: Does the evidence support your sentence about this place, or suggest a question for further investigation?

AI mapping can make remote-sensing evidence easier to explore. Good reporting preserves the connection between a compelling image and the measurement process behind it. Give the resolution, dates and evaluation enough space in the explanation, and the map can help people ask more precise questions about a forest.

Sources

• Lang et al., A high-resolution canopy height model of the Earth (2023), with a figure-label correction (19 February 2024).
• Tolan et al., Very high resolution canopy height maps from RGB imagery (2024 journal article; linked author manuscript, revised 15 December 2023).
• NASA ORNL DAAC, GEDI L4A Footprint Level Aboveground Biomass Density, Version 3, documentation revised 2 September 2026.