Authors Benjamin Carter, Jeff Blackadar and Weston Conner joined each other in conversation for this new blog post for Advances in Archaeological Practice.

Ben: Back in 2018, Weston and I zigzagged across the Blue Mountain, climbing over rocks and dead trees, pushing through brambles, getting frightened by monstrous black rat snakes and brushing through knee high ferns that streaked our pants green—first heading down the mountain and then turning to summit again repeatedly. This type of landscape is what Bernhard Fernow labeled “waste places” in 1883. Now part of Pennsylvania’s State Game Land (SGL) 217, this was the very “waste place” that Fernow, one-time manager of the nearby Lehigh Furnace (and, later, third chief of the USDA’s Division of Forestry), would have been most familiar with. The mountain is a long, low rocky ridge stretching c. 250 miles across eastern and central Pennsylvania. It would have made poor farming—hence a “waste place”—but Fernow and many others recognized its value as a source for charcoal, an important fuel for iron furnaces. We were crisscrossing the mountain to check on small flat areas that we had previously identified on slope maps derived from LiDAR. We confirmed that these are relict charcoal hearths (RCHs)—locations where large mounds of wood were capped with earth and slowly carbonized. The charcoal was then transported to furnaces via a series of intricate roads and trails, one of which is now followed by the famous Appalachian Trail.

Weston: This led me to think, if this SGL was a “waste place,” undeveloped and preserved at least in part due to its charcoal producing past, why not others? Perhaps throughout Pennsylvania? Exploring this inkling of an idea seemed a quixotic pursuit given its scope, but as I adjusted to life post-college, I tinkered with it, and happily the next three years proved me wrong. From figuring out bash scripting to downloading 350 gigabytes of zipped lidar files, to connecting online with Moritz “Moe” Schiesser—a Swiss programmer who wrote programs to organize and process his files using LAStools (ballooning my data to 1.3 terabytes), to manually marking over 4,000 charcoal hearths across 62 SGLs, I got farther than I initially thought possible. I reconnected with Ben and we considered how to do this systematically.

Ben: A brief request for assistance with our project on Twitter, yielded a response from Shawn Graham (Carleton University, @electricarchaeo), who connected us with Jeff.

Jeff: In May 2020, I had just finished a project with lab partner Osama Javaid that used Deep Learning to recognize different types of Roman masonry construction in photographs of Pompeii. The model was based on a library of annotated photographs and data kindly provided by Eric Poehler and Cinzia Presti. When I saw Ben’s Tweet on Shawn’s feed, I wondered whether the technique could be applied to find objects in landscapes. I soon got in touch with Ben and Weston.

All: Quickly, the group saw potential with this approach, and Jeff got to work to train the model, refine it and test. As Covid-19 shut down our institutions and interfered with field work, we were able to continue this work virtually. Jeff retrained, tested and sampled to assure quality. Some training runs took several hours, as did predictions, so Jeff let the computer process them overnight and…dream of charcoal. The model was able to find thousands of objects that resembled charcoal hearths, but also produced a troubling number of false positives. Based upon a suggestion by John Clark (GIS Librarian at Lafayette College), we employed cluster analysis to refine the results yielding dramatically fewer false positives.

All: In the end, we were able to identify more than 20,000 RCHs across Pennsylvania located within and around SGLs especially in southern and eastern Pennsylvania. Identification of these industrial landscapes informs both the history of iron production and the nature of “waste places” and “wild” places in Pennsylvania. We are happy to share this article through open access, but also note that the article contains links to our data and code.

The associated open access article – When Computers Dream of Charcoal: Using Deep Learning, Open Tools, and Open Data to Identify Relict Charcoal Hearths in and around State Game Lands in Pennsylvania – is out now in Advances in Archaeological Practice.

References:
Fernow, Bernhard E. 1883. “Planting in Waste Places.” The American Journal of Forestry 8 (January): 153–55.