Imagery in Forest Inventory - Platforms, Sensors, and Analysis

Foresters were some of the earliest users of remote sensing - the US Forest Service was using aerial imagery as a tool for assessing forests back in the 1930's. Over the past few decades, there has been an accelerating development of new sensors and platforms for remote sensing, from orthophotography and LiDAR to drones and microsatellites. Modern foresters are now faced with an overwhelming array of options for imaging their forests. What is the right tool for the job?

Remote sensing complements a traditional timber cruise rather than replacing it - it "fills in the gaps" between sample plots. What makes one "filler" better than the other is how strongly it is related to the underlying population parameters (species, diameter distribution, stem count, volume, etc.) being estimated. The power of remote sensing is that it delivers a wall-to-wall measurement of every part of the forest - a true census. If there is a strong relationship between the remotely sensed data and the field measurements (the things you actually care about), it is possible to significantly improve the statistical precision and confidence level of a forest inventory while reducing the amount of fieldwork.

When assessing a new approach to remote sensing, the key question is always "do the reduction in fieldwork and improvements in precision and management outcomes outweigh the cost of imagery acquisition and analysis?" And as is often the case in forestry, the answer is "it depends!" There are many factors including forest type, plot measurement costs, acceptable silvicultural practices, market prices, etc. that influence the costs and benefits of adopting a new approach to forest sampling. But using the "cost + loss" approach we have developed over our last few "Biometrics Bits" articles, we now have a quantitative framework for evaluating different imaging options. We will conduct in-depth quantitative analyses of individual technologies in future articles, but first let's survey how new imagery platforms, sensors, and analysis techniques are enabling better decision-making in forestry.

Many foresters are excited about drones. Drones can deliver beautiful, high resolution images of individual stands and new software is making the rectification and georeferencing of drone imagery faster and easier. There remain significant logistical and cost challenges (weather, flight ceiling, battery life, line-of-sight regulations, training, maintenance, etc.) to using drones at scale, but some foresters are finding drone imagery to be a useful tool for assessing stands and communicating with landowners and community stakeholders.

Of course, plane-flown aerial imagery has been in use in forestry for decades. The USDA's 1-meter resolution National Agricultural Imagery Program (NAIP) imagery is available for most counties in the US for free online. But the high costs and technical knowledge required to commission planes, pilots, and imaging equipment puts other plane-flown imagery out of reach for most foresters outside of large companies and government agencies. However, several states have paid for state-wide LiDAR flights and some of this data is now available for free. It is unclear if or when new statewide flights will be conducted - especially since one of the primary justifications for the LiDAR flights was to develop more precise digital elevation maps which will not need to be refreshed any time soon.

The venerable Landsat satellite mission is another useful (and now free - thanks NASA!) platform for forest imaging. With a revisit frequency of 16 days, the latest Landsat 8 satellite typically has a few good images a year for most forested areas within the United States. Many foresters use imagery from Landsat because it delivers images of vast swathes of remote forest effortlessly and for free. There are a variety of other government and commercial satellites as well that provide imagery useful in monitoring forests.

One emerging imaging platform that doesn't get enough attention in forestry is the "microsatellite." Unlike traditional, bus-sized satellites like Landsat, microsatellites are about the size of your arm. They are essentially glorified smartphones with solar panels strapped on. But even so, their imagery can be used for many forestry applications and because there are now hundreds of these microsatellites in space, the revisit frequency for the continental US is nearing 24 hours. This not only ensures a wide selection of imagery to choose from (important in cloudy areas like the Pacific Northwest), but also creates new possibilities for detecting timber theft, beetle kill, fire damage, etc.

1) A conventional grid of plots (green) in a stand in the Pacific Northwest yielded an estimate of basal area with an error of +/- 17% at 90% confidence 2) Remotely sensed imagery adds a lot more information about the forest and helps fill in the gaps between plots 3) Total basal area distribution across the stand (redder means higher BA). Statistical analysis of a stack of remotely sensed imagery tightened up the total BA confidence interval to +/- 10% 4) Western Hemlock BA 5) Douglas Fir BA 6) Noble Fir BA. Aerial imagery courtesy of Bing Maps

Visible spectrum and near infrared imagery have been used in forestry for decades. When timed appropriately, they can help differentiate species and reveal elements of forest structure. These sensors are similar to those found in modern consumer cameras and smartphones and can be flown on drones, aircraft, or satellites.

Hyperspectral sensors are similar to visible spectrum sensors, except that they detect many more wavelengths than just red, green, and blue. Researchers are finding many uses for hyperspectral imagery and there is work underway to try to detect disease and insect infestation. However, the high cost of hyperspectral sensors has limited their deployment in the forest industry so far.

LiDAR is another imaging technology that has been the subject of much discussion over the last decade. A LiDAR sensor emits bursts of laser pulses that penetrate forest canopy to the middle and understory as well as the underlying ground. The resulting high-resolution point cloud enables measurements of tree height which can be used to determine the site index. Expensive and technically intricate, LiDAR sensors are usually flown from planes with dedicated support staff. Recently though, people have been (nervously!) strapping them onto drones and the GEDI LiDAR sensor will be mounted on the International Space Station in 2019.

Radar is a type of sensor that many people don't immediately associate with forestry. In many ways, however, radar is similar to LiDAR. It penetrates the forest canopy and different forest structures reflect radar pulses differently. Individual canopy segmentation is not possible with radar, but it is strongly correlated to forest structure and total volume. Most commonly flown on satellites, radar data is often a bit tricky to process but there are now several free and paid sources of radar imagery.

Imagery acquisition is just the first step. To inform quantitative management decisions, the imagery must be computationally analyzed. Easier said than done! As humans, we can look at an image and understand it instantly. It is much more difficult to teach a computer how to do so.

This is particularly the case for analyzing point clouds - a common starting point for remote-sensing analysis. Point clouds can be derived from LiDAR sensors or from multiple normal spectral sensors using a technique called "structure from motion." Deriving a statistically sound forest inventory from point cloud data is still an open question that is attracting much research. There are often considerable difficulties with the systematically unbiased segmentation of individual trees, particularly in mixed or mature forests with canopy closure and overlap.

An emerging trend in remote sensing analysis is to incorporate statistical techniques from the field of "machine learning" in order to design more efficient inventory systems. In many significant respects, these techniques resemble elements of the classical forest stratification approach and the more advanced 3-P sampling methods. A stack of imagery from many different sensor types creates a "fingerprint" for each area of the forest. When paired with conventional cruise data, this census-level information can be used to inform well-understood model-assisted / model-based techniques for accurate estimation of stand-level forest inventory. One of the advantages of these statistical sampling techniques is that they use (and can fall back on) the same type of cruises that foresters are accustomed to. The imagery "fills in the gaps" between a traditional grid of sample plots. This is the approach that we take at SilviaTerra and have found that it enables foresters to incrementally adopt this new technology by gradually reducing their plot density over time while getting more informative data.

It is a very exciting time to be a forester. With all sorts of new imagery platforms, sensors and analysis techniques, we are getting closer to a future where every landowner has a good forest inventory and a data-driven management plan. Stay tuned for future articles where we'll conduct a "cost + loss" analysis for these technologies in different forest types so that you can determine which technology makes the most sense for your forest.