In the following article, find out how forest owners in Sweden are using aerial imagery, remote sensing and GIS to ensure the sustainability and health of their forests.

Simply put, forestry in Sweden is big business. Nearly 60 percent of the country — 23 million hectares (56.8 million acres) — is covered in forest. In comparable size, that’s an area bigger than the whole of Great Britain.

The world’s second largest exporter of paper, pulp and sawn wood products, Sweden’s forest industry is valued at around 90 billion euros ($100.4 million U.S.) — about one-fifth of Sweden’s GDP — and employs about 200,000 people.

That makes managing Sweden’s forests serious business. Ensuring the sustainability and health of this wealth-producing asset falls mainly to private land owners. Family enterprises own 50 percent of Swedish forests, and private companies own an additional 25 percent. These owners have been guided by management regulations since the first Forestry Act of 1903, a national forestry policy that, among other directives, requires owners to replant after harvesting.

Thus, on regular intervals, forest owners will assess and inventory their present-day holdings in order to develop long-term operational strategies focused on protection, production and regeneration. They have been quick to adopt technological tools that can help them do that. This includes aerial photography, optical satellite imagery and geographic information systems (GIS) software, all of which have been used as regularly as tree harvesters for about two decades.

However, the process of classifying forest land and delineating forest stands — a contiguous group of trees that are sufficiently homogeneous in species, density and size — has typically been a laborious combination of photogrammetry and drawing features by hand.

That began to change in 2008 when Erik Heimsdal Iversen, a chief remote sensing specialist with COWI, an international engineering, environmental and economics consulting group based in Lyngby, Denmark, endeavored to pair LiDAR data and other datasets with Trimble’s eCognition image analysis technology to develop an automated forest classification and mapping solution.

“I knew LiDAR data would provide key vegetation height detail and I knew the classification capabilities of eCognition, but I had never tried to classify and delineate large, heterogenous forest areas with the two together,” says Iversen, who works in COWI’s Norway office in Kongsberg. “It turned out that the combination is particularly well suited for rapidly classifying vegetation and delineating forest stands. And it’s 10 times faster than photogrammetry.”

Indeed, with the elevation component of LiDAR and the classifying intelligence of eCognition, Iversen says they’ve not only been able to cut the traditional forest mapping effort from several months to a few days, they’re able to offer a more precise, flexible, customizable and repeatable technique that visualizes only the forest detail the customer wants to see. It’s a new approach that’s been garnering a forest of interest.

The Right Mix

When Iversen first joined COWI’s GIS and IT department in 2008, he came with 20 years of photogrammetry experience in creating forest inventories, six years of experience with LiDAR forestry applications and about three years of experience with eCognition. The timing was opportune because COWI was interested in launching a forestry-mapping business but wanted to find a more efficient and flexible alternative to photogrammetry. Iversen, in particular, wanted to find a solution that would allow them to integrate valuable datasets such as old stand data to both update the stand data and create more detailed and informative maps.

“Many large forest owners maintain their own databases of forest stands, which contain details such as the stand shape and size, age, tree height, forest density and topography,” Iversen says. “Those attributes are collected from the field and are really important to retain because they help classify and map change in a forest, whether it be identifying clear cuts, defining productive forest, or distinguishing young from mature forest. And all of that detail directly feeds into planning and management operations. Those were the types of customized maps we wanted to be able to readily produce.”

Based on his experience with eCognition, Iversen thought the software would be the right tool to give them the flexibility to integrate diverse datasets and the automation to rapidly classify and map large swaths of forest.

“I don’t know of any other software that can seamlessly work with old data, new data, vector data and raster data, and automatically classify and map vegetation with such intelligence,” Iversen says.

Iversen built the first master rule set (the if-then processing tree that the software follows to determine specific vegetative types) in about two months. The eCognition rule set was designed to classify nine different vegetative types, identify clear cuts, delineate stand borders, classify micro stands — variations in tree species or height within the main stand — distinguish productive forest and, where warranted, re-categorize unproductive forest land as productive.

Soon after building the eCognition solution, Swedish forest company Bergvik Skog needed to assess its 2.4-million-hectare forest in order to better define its forest stands and map its productive forest. It was the test case Iversen needed to determine the validity of the new methodology, and it was a veritable success.

Since that debut project, Iversen has enhanced the rule set to classify more features such as individual trees and seed trees. To date, he and a small team have classified 5 million hectares (12.3 million acres) of Swedish forests, as well as swaths of forest in Finland and Uruguay.

One Forest, Multiple Perspectives

While the methodology can integrate diverse source data such as vector-based stand data, GIS data, raster imagery and LiDAR data to classify and map vegetation, Iversen designed the COWI solution to merge existing stand databases with LiDAR data to update the stand data, correct stand borders, classify the forest and precisely map the property.

Although classification results will depend on the source data provided, for a typical forest-mapping project, Iversen first prepares the existing stand data — identifying inaccuracies and correcting them — and uses the LiDAR data to create a normalized digital surface model (nDSM) of tree heights — detail that is integrated into the classification process. Then, eCognition takes those inputs and first segments the raster imagery into meaningful objects, based on both their spectral and spatial characteristics. Next, it flags abnormal stand borders and corrects them, identifies clear cuts and distinguishes between unproductive and productive forest. The software then classifies the tree stands and micro stands.

With that classification foundation, Iversen is able to produce maps of the entire forest and its stands, maps of just the micro stands, maps of just clear cuts or change-detection maps showing only the changes that have occurred in a specified number of years.

While eCognition’s speed, flexbility and accuracy make it a viable alternative to photogrammetry, Iversen says the ability to combine the historical stand data with the elevation detail of LiDAR is what really sets this approach apart.

“Updating old stand borders with manual methods is very time consuming and is prone to error,” he explains. “But the old data is quite valuable and contains detail not obtainable through remote sensing tools. Our eCognition method enables us to efficiently and accurately update the stand data, use it in combination with LiDAR-based data to perform a change detection analysis, and map the results. And I can repeat this process each time a customer wants an update. It’s quite a new approach but we already have repeat customers.”

One of those latest repeat customers is Swedish forest company Orsa Besparingsskog.

Revisiting the Forest

Based in the town of Orsa, Orsa Besparingsskog (Orsa) is responsible for 80,000 hectares (198,000 acres) of forest in central Sweden, property that is located about 15 kilometers north of the city of Mora. The company was one of the early recipients of the COWI eCognition method, having received a forest inventory and maps of its holdings in 2010. In January 2016, it returned to COWI with a request to assess its forest volumes and to update the stand borders based on those volumes and new clear cuts.

For the project, Iversen had Orsa’s existing vector database of stand polygons and a 2010 LiDAR-based nDSM that he prepared for the 2010 analysis. To prepare the data for processing, Iversen performed quality control checks on Orsa’s existing vector database of stand polygons and he created five different nDSMs based on 2015 LiDAR data. All of them offer different features for analyzing sparse and non-sparse swaths of forests. He also developed an additional LiDAR-derived vegetation density raster index to provide a detailed view of the forest density.

With the datasets prepared, Iversen began customizing the eCognition rule set. In two days, he had developed the algorithms to identify and delineate stand borders, distinguish vegetative types based on their height, spectral qualities and textural features — a process similar to how the human brain distinguishes specific objects — and classify the vegetation.

Working in 20,000-hectare (50,000-acre) batches, the software automatically recognized stand border changes and delineated them, defined existing and new clear cuts, detected and delineated differences in the forest between 2010 and 2015 and divided unproductive areas into two classes, trees shorter than 4 meters and trees taller than 4 meters — a workflow that took 30 minutes. Iversen then spent about two days performing a supervised classification on those initial results.

After the first round of processing, Iversen ran eCognition again to determine if stands needed to be further divided, to pinpoint the micro stands and to perform a final classification of the forest cover, down to the position and heights of individual trees — a second analysis that took 30 minutes to complete. It’s a process he repeated four times, allowing him to automatically classify Orsa’s 80,000 hectares of forest in four hours.

“To produce this amount of forest detail with photogrammetry would require hundreds of hours and several image analysts,” Iversen says. “With this approach nearly all of the processing and analysis is completely automatic. I only need to click ‘run’ and we get 10 times the information in minutes, with little to no effort.”

After the second round of classification, Iversen again performed supervised classifications and quality control to ensure the accuracy. Once the classification process was complete, the data was automatically integrated into ArcGIS to further analyze the data and produce customized maps. By March, Iversen had prepared three map layers showing the forest volume, the stands and micro stands, and nine different forest classes, including new clear cuts. Plotted across a standardized grid, each stand and micro stand had an individual ID, along with its corresponding attributes such as its size, age and tree heights. Upon receiving the forest inventory from Iversen, Orsa field tested the information and reported the maps were 95 percent accurate.

“Now Orsa not only has a detailed comparison of its forest between 2010 and 2015, they have a completely new, informative tool that can help them develop cost-effective planning and management strategies,” Iversen says. “And we can repeat the process and customize it to extract a host of other details, even ones we haven’t produced before.”