Canopy Height Model (CHM)

Measuring the height, structure, and extent of a forest across vast landscapes presents a monumental challenge. Traditional field methods, while accurate, are limited in scale and cannot capture the continuous, intricate architecture of a forest canopy. This gap in our ability to see the forest in three dimensions hinders everything from ecological research to effective resource management. The Canopy Height Model (CHM), a product of modern remote sensing, provides a powerful solution to this problem. It offers a detailed, spatially explicit map of vegetation height above the ground, effectively creating a digital cast of the forest's surface.

This article delves into the science and utility of the Canopy Height Model. In the first chapter, Principles and Mechanisms​, we will explore how a CHM is created, journeying from the initial laser pulses of a LiDAR system to the sophisticated data processing required to sculpt raw point clouds into precise surface models. We will uncover the elegant yet simple mathematical foundation of the CHM and the critical considerations for ensuring its accuracy. Following this, the chapter on Applications and Interdisciplinary Connections will reveal the transformative power of this model, demonstrating how it is used to conduct a forest census, analyze ecological habitats, monitor disturbances, and even quantify the forest's role in the global carbon cycle.

Imagine you want to measure the height of a vast, impenetrable forest. You can't just walk in with a tape measure. So, how do you do it? The answer, as is often the case in science, is to be clever. Instead of touching the trees, we touch them with light. This is the essence of Light Detection and Ranging (LiDAR), a technology that has revolutionized how we see the world, and it is the foundation upon which the Canopy Height Model is built.

At its heart, LiDAR is breathtakingly simple. It's like shouting into a canyon and timing the echo to gauge its width. A LiDAR instrument, typically mounted on an airplane, fires a short, intense pulse of laser light towards the ground. This pulse travels down, hits something—a leaf, a branch, the ground itself—and a tiny fraction of that light scatters back to a detector on the airplane.

The instrument precisely measures the total time the light took for this round trip, a duration we can call $\Delta t$. Since we know light travels at a constant speed, $c$ (in the atmosphere), we can calculate the one-way distance, or range ($R$), to the object it hit. The total distance is $c \cdot \Delta t$, so the one-way range is simply half of that.

This equation is the heartbeat of LiDAR. The factor of $2$ is crucial; it accounts for the fact that our measurement is for a round trip, but we only want the distance to the target.

Of course, knowing the distance isn't enough. If you're on a moving train and shine a laser pointer, knowing only the distance to the illuminated spot doesn't tell you where that spot is on the landscape. You also need to know precisely where the train is, which way it's pointing, and the direction you're aiming the laser. A LiDAR system does the same: it uses a Global Navigation Satellite System (GNSS) to know its position ($\mathbf{x}_s$) and an Inertial Measurement Unit (IMU) to know its orientation ($\mathbf{C}_{ws}$). By combining this with the laser's precise pointing angle ($\theta, \phi$), it can convert each range measurement into a georeferenced three-dimensional coordinate $(x,y,z)$.

Repeating this process millions of times per second, the system generates a point cloud: a ghostly, three-dimensional digital replica of the landscape below, composed of billions of individual measurement points.

This point cloud is a magnificent but jumbled collection of data. To make it useful, we need to impose order. For measuring tree heights, we are interested in two primary surfaces.

First, we need to know where the "bare earth" is. This is the Digital Terrain Model (DTM), sometimes also called a Digital Elevation Model (DEM). Creating a DTM is a challenging art. It requires sophisticated algorithms to sift through the point cloud and classify which points are "ground" and which are "non-ground" (like vegetation or buildings). The DTM is then created by interpolating a continuous surface through only the ground points. In a dense forest, where very few laser pulses actually reach the forest floor, this can be particularly difficult.

Second, we need a map of the uppermost surface of the landscape—the very first thing the laser pulses would hit. This is the Digital Surface Model (DSM). Conceptually, this is simpler to create. For any given location on a grid, the DSM value is typically the elevation of the highest LiDAR point found in that area. It's the "skin" of the landscape, draping over tree canopies, rooftops, and everything in between.

Here comes the beautifully simple, core idea. If you have a map of the top surface (the DSM) and a map of the ground beneath it (the DTM), how do you find the height of a tree at any given spot? You just subtract!

This simple subtraction gives us the Canopy Height Model (CHM).

This elegant formula is the central principle of our topic. For a given coordinate $(x,y)$, the CHM value tells us the height of the object at that location above the ground. Over a vegetated area, a CHM value of $13.5 \text{ m}$ means the canopy is $13.5$ meters tall. Over bare ground, the DSM and DTM will have nearly the same elevation, so the CHM will be close to zero.

However, this subtraction is only meaningful under two strict conditions. First, the DSM and DTM must be perfectly aligned, or co-registered​, so that we are subtracting the ground elevation from the surface elevation at the exact same horizontal location. Second, both models must be referenced to the same vertical datum (the common reference for "zero" elevation, like mean sea level). Trying to subtract elevations from different datums would be like trying to measure a person's height by subtracting their shoe-sole elevation measured from the floor from their head elevation measured from the ceiling—the result would be meaningless.

It's also important to remember what this new model shows us. The CHM is the height of everything above the ground. In a mixed landscape, it will show the height of trees in the forest and the height of buildings in a city. For many applications, like calculating forest biomass, we must first use a land cover map to "mask out" the non-vegetated areas, ensuring our analysis is focused purely on the canopy.

Let's zoom in on the practical step of turning the point cloud into a gridded map like a CHM. A grid cell, say 2 meters by 2 meters, might contain dozens of LiDAR points. How do we distill all their heights into a single value for that cell? This choice of aggregation strategy has a profound impact on the final model.

• We could take the maximum height. This seems logical, as it should capture the very top of the highest tree. However, it's very sensitive to outliers. A single erroneous high point, or even a bird flying through the laser beam at the exact wrong moment, could give a falsely high value for that cell.

• We could take the mean (average) height. This is more robust against single outliers, but it has its own problem. The mean will be pulled down by the many returns from lower branches and the sides of the canopy, leading to a systematic underestimation of the true canopy top height.

• A clever compromise is to use a high percentile​, like the 95th percentile ($q = 0.95$). This strategy ignores the top 5% of points, making it robust to outliers, while still capturing a value very close to the true canopy top, unlike the mean.

The choice is not merely technical; it reflects a philosophical decision about what we want our CHM to represent: the absolute highest point, a measure of central tendency, or a robust estimate of the canopy's upper surface.

So far, we have a somewhat idealized picture. In reality, a forest canopy is not a solid surface. It's a complex, porous volume. When a laser pulse enters the canopy, it doesn't just hit the top and stop. Part of the pulse's energy might reflect off an upper leaf, generating the first return​. The rest of the pulse continues downward, possibly hitting a branch lower down to create an intermediate return​, and some energy might even make it all the way to the ground to generate the last return​.

The probability of a pulse reaching the ground decreases roughly exponentially as it travels through the canopy—much like sunlight fading as you dive deeper into water. This is described by a principle similar to the Beer-Lambert law. In a very dense forest, the chance of any given pulse reaching the ground might be very small, perhaps only 10-20%. This is why the last return is not guaranteed to be a ground return; it might just be the lowest branch the pulse managed to hit before its energy ran out. This makes creating an accurate DTM under dense canopy one of the greatest challenges in LiDAR processing.

This sampling issue also creates a major problem for the DSM: interpolation pits​. If, by chance, no laser pulse happens to strike the top of a tree within a certain area, the DSM in that spot will be erroneously low, created from nearby returns on lower branches. This creates an artificial "pit" in the CHM.

To combat this, a wonderfully elegant technique known as the pit-free CHM was developed. Instead of computing the DSM at just one scale (e.g., using the max height in a 1m radius), we compute several DSMs at multiple scales—say, using radii of 1m, 3m, 5m, and 10m. Then, for each pixel, we take the highest value from across all these scales. The small-radius search preserves the fine details of individual tree crowns where sampling is good. The large-radius searches act as a safety net, "bridging" across gaps where sampling was sparse, effectively filling the pits. This multi-scale approach perfectly balances the preservation of detail with the need for robust gap-filling.

This brings us to a fundamental trade-off in creating any gridded model from point data: the choice of cell size. We face two competing desires.

1. To capture fine details of the terrain or canopy, we want our grid cells to be as small as possible. The Nyquist-Shannon sampling theorem from signal processing gives us a hard limit: to faithfully capture a feature of a certain size (wavelength $\ell$), our cell size ($s$) must be no larger than half that size ($s \le \ell/2$).

2. To get a reliable estimate of height in each cell, we need to have enough LiDAR points falling within it. If our cells are too small and our LiDAR point density ($\lambda$) is too low, many cells will be empty, containing no data at all. The probability of a cell being empty can be modeled using a Poisson process, giving $P(\text{empty}) = \exp(-\lambda s^2)$.

Balancing these two constraints is the art of LiDAR data processing. One must choose a cell size $s$ that is small enough to capture the desired detail, but large enough to ensure that the number of empty cells is acceptably low given the point density of the survey. If the required detail level is too fine for the given point density, the only solutions are to accept data gaps or, better yet, to plan for a higher-density LiDAR acquisition from the start.

We have journeyed from a pulse of light to a beautiful, colored map of canopy height. But how good is this map? A map without a statement of its own uncertainty is scientifically incomplete.

First, we can assess the map's accuracy by comparing it to independent "ground truth" measurements. This is crucial. We can't use the data we built the model with to test it; that's like grading your own homework. By collecting field measurements of canopy height and comparing them to our CHM, we can compute metrics like:

• Bias​: Does our model systematically overestimate or underestimate height?
• Mean Absolute Error (MAE): On average, what is the magnitude of the error?
• Root Mean Square Error (RMSE): A metric similar to MAE, but it penalizes large errors more heavily.

Going even deeper, we can build a model of the uncertainty for every single pixel of our CHM. The total error in a CHM value comes from two main sources.

1. Systematic Bias​: The largest source of bias often comes from the DTM. If the ground-classification algorithm mistakenly labels low vegetation (with height $m_i$) as "ground", the DTM in that cell will be too high. If this happens with a certain probability ($q_i$), it introduces a positive bias in the DTM of roughly $b_{E,i} \approx q_i m_i$. Since CHM = DSM - DTM, this positive DTM bias creates a negative CHM bias, causing us to underestimate the true tree height.
2. Random Error​: The precision of our DSM and DTM estimates in a cell depends on the number of points used to create them. Just like in a political poll, the more data points (voters) you have, the smaller your margin of error. The standard deviation of the height estimate is typically proportional to $1/\sqrt{n}$, where $n$ is the number of points in the cell.

By combining these effects, we can calculate a total uncertainty, like the Mean Squared Error ($MSE = \text{Variance} + \text{Bias}^2$), for every pixel. This transforms the CHM from a simple picture into a true scientific data product: a map that not only shows us the height of the forest but also tells us exactly how confident we can be in every single value. This is the final, crucial step in turning a dance of light into a profound understanding of our world.

Now that we have a feel for what a Canopy Height Model (CHM) is—this remarkable plaster cast of the forest’s surface—we can ask the most exciting question: What is it good for​? A map of tree heights is beautiful, certainly, but its true power is revealed when we treat it not as a static picture, but as a key that unlocks a staggering array of secrets about the forest as a living, breathing system. Let us go on a journey, starting with the most tangible objects in the forest and ending with the forest’s role in the balance of the entire planet.

The most immediate and obvious thing to do with a CHM is to measure the trees. Looking at the CHM, we see a landscape of domes and spires, each one corresponding to the crown of a tree. The highest point of each dome is, quite simply, the treetop. Its height in the CHM is the tree’s height above the ground. The area of the dome's base gives us a measure of the crown's width. We can even define indices, such as the ratio of the crown's edge height to its peak height, to describe its shape or "compactness". For a single, isolated tree, this is wonderfully straightforward.

But a forest contains millions of trees. How can we perform a census for them all? We cannot simply sit and point at every dome. We must teach a computer to see the trees. This is where the real cleverness begins. The task is to design an algorithm that finds local maxima—the peaks—in the CHM. A naive approach might find dozens of "peaks" on the bumpy surface of a single large tree crown. The trick, it turns out, is a beautiful principle that echoes through so much of science: the scale of your measurement tool must match the scale of the object you are measuring.

To find treetops, we can use a digital filter that slides a "window" across the CHM, looking for the highest point within that window. If the window is too small, it will get lost in the texture of a single crown and report multiple spurious peaks. If it is too large, it might span two or three smaller trees and only report the tallest one, missing its neighbors. The ideal window size, therefore, is one whose radius is about the same as the radius of a typical tree crown in that forest. By matching the scale of our analysis to the scale of the trees themselves, we can reliably automate the process of locating and counting individual trees across vast landscapes, turning the continuous CHM into a discrete catalog of the forest's inhabitants.

Beyond individual trees, the CHM reveals the forest's collective architecture—its texture, its complexity, its "rugosity." A simple measure of this is the standard deviation of canopy heights within a plot. A high value suggests a complex canopy with tall trees and deep gaps, while a low value suggests a more uniform, smooth canopy.

But here we encounter a wonderful subtlety, a classic lesson in what numbers can and cannot tell us. Imagine two forest plots. One has a single, massive gap in the middle where a giant tree fell. The other is dotted with dozens of small gaps from smaller disturbances. Let's say that the total area of gaps is the same in both plots, and the remaining trees are identical. If you calculate the standard deviation of canopy height for both plots, you will get the exact same number! This simple statistic, for all its utility, is blind to the spatial pattern​. It cannot distinguish one big hole from many small ones.

To see the texture, we need more sophisticated tools—spatial statistics. Metrics like the semivariogram or lacunarity are designed to answer questions about patterns. A semivariogram, for instance, measures how different the canopy height tends to be for pairs of points as a function of the distance separating them. In the forest with many small gaps, height will become decorrelated over short distances, while in the forest with one large gap, the correlation will extend much farther. These tools allow us to quantify the characteristic size of gaps and crowns, providing a true measure of canopy texture that goes beyond a simple measure of height variation.

Where does this intricate architecture matter most? It turns out this structure is not just a curiosity; it is a map of ecological processes. Consider the edge of a forest fragment. A forest doesn't just stop at its boundary with a field; its character changes. Sunlight and wind penetrate from the side, stunting growth. The CHM allows us to see this "edge effect" with stunning clarity. If we take a transect from the open field deep into the woods, we see the canopy height gradually rise, recovering from the harsh conditions at the edge. Remarkably, this recovery often follows a simple, elegant mathematical law: an exponential curve. This is the same "approach-to-equilibrium" process that describes a cooling cup of coffee or a discharging capacitor. The CHM allows us to measure the characteristic length of this effect—the "structural edge depth"—and see how this zone of influence, a kind of ecological penumbra, varies from place to place.

Ultimately, this architecture is the stage upon which life plays out. A bird does not care about the "average height" of a forest; it cares about the availability of specific perches, nesting sites, and feeding grounds. A structurally simple plantation offers few niches, while a complex, multi-layered forest with gaps and snags offers many. This is the essence of the "habitat heterogeneity hypothesis." The CHM provides a direct, quantitative map of this heterogeneity. By incorporating metrics of canopy rugosity, height variance, and gap fraction into models of species distribution, ecologists can dramatically improve their ability to predict biodiversity. The physical structure, measured by LiDAR, becomes a powerful predictor of the biological richness the forest can support.

Perhaps the most powerful application of CHMs comes from comparing them over time. By flying a LiDAR survey over the same forest years apart, we can play "spot the difference" at a landscape scale.

The most dramatic use of this technique is in the aftermath of a natural disaster. Imagine a severe windstorm tears through a forest. By subtracting the post-storm CHM from the pre-storm CHM, a map of the destruction instantly materializes. Pixels with large negative values represent areas where the canopy has collapsed. We can precisely calculate the total area of canopy loss. Furthermore, we can see the pattern of the damage—often, a trail of destruction or a cluster of fallen trees where the failure of one exposed its neighbors in a catastrophic domino effect.

Of course, science demands rigor. How do we distinguish a small amount of real growth from the inherent "fuzziness" or noise in the LiDAR measurements? This is where the detective work becomes truly sophisticated. A change in measured height is a combination of true change and error from multiple sources: the vertical uncertainty of the first survey, the vertical uncertainty of the second, and the error from imperfectly aligning the two maps. Scientists can propagate these errors using statistical theory to calculate the total expected noise in the differenced CHM. Only then can they set a statistically defensible threshold and say with confidence, "This change is real." This allows them to create reliable maps classifying the entire forest into areas of significant growth, significant disturbance, or stability.

This ability to model structure and change opens the door to predictive modeling of hazards, most notably wildfire. The behavior of a fire is governed by fuel, weather, and topography. The CHM provides critical information about the three-dimensional structure of the fuel. It tells us not just the height of the canopy, but also its density and the height of its base. This structure, in turn, governs how wind flows through and over the forest. Using principles from fluid dynamics, such as the logarithmic wind profile, scientists can use CHM-derived metrics to estimate wind speed near the ground. This wind speed is a critical input for fire spread models. By linking the CHM to a chain of physical models, we can move from simply mapping the forest to predicting its potential fireline intensity, creating vital tools for risk assessment and management.

Finally, we zoom out to the planetary scale. Forests are one of the great engines of the global carbon cycle, absorbing vast quantities of carbon dioxide from the atmosphere and storing it as biomass. Quantifying this carbon stock is one of the most urgent tasks in climate science, and the CHM is an indispensable tool.

The method is a brilliant synergy of field work and remote sensing. We cannot weigh an entire forest, but we can weigh all the trees in a small field plot. Ecologists do this, and from these measurements, they calculate the total aboveground biomass (AGB) in the plot. Meanwhile, they extract structural metrics from the CHM for that exact same plot—metrics like the 95th percentile of canopy height, which is a robust indicator of the height of the dominant trees. By collecting this information from many plots, they can build a statistical model that relates the CHM's structure to the AGB on the ground.

This model becomes a "Rosetta Stone." It allows us to translate the language of canopy height into the language of biomass. We can then apply this model to the wall-to-wall CHM, converting the entire map of height into a map of carbon. Of course, to make these estimates credible enough for international climate mitigation reporting, the statistics must be impeccable. The models must account for the multiplicative nature of biological growth, the validation must be done on independent data to avoid self-deception, and subtle biases introduced by mathematical transformations must be corrected.

Even for a seemingly simple metric like "canopy cover"—the fraction of the ground covered by trees—the CHM provides a new level of sophistication. Instead of just a binary yes/no, scientists can use their knowledge of the LiDAR's measurement error to calculate the probability that any given pixel is truly covered by canopy. This gives us not just an estimate, but a map of our confidence in that estimate.

From a single tree to the global carbon budget, the journey is breathtaking. The Canopy Height Model, born from beams of laser light shot from a plane, is far more than a simple map of heights. It is a scientific instrument of profound power—a lens through which we can see the forest's architecture, monitor its pulse, predict its future, and quantify its vital role in the functioning of our world.