Detecting Pitfall Systems in the Suomenselka ¨ Watershed, Finland
Introduction
Artificial intelligence (AI) is increasingly used to locate and identify archaeological sites.
Airborne laser scanning (ALS) data and satellite images can be analyzed automatically or semi-automatically for faster and more accurate discovery of features like ancient ruins.
Deep learning algorithms can identify patterns such as building remains and human-dug pits.
AI can help classify discovered remains, facilitating dating and valuation of cultural significance.
The LIDARK project (2021–2022) is an example of AI in Finnish archaeology, which developed an algorithm for detecting ancient pitfalls.
Pitfalls in Finland have been mainly studied in Lapland, associated with the Sami people's hunting of mountain reindeer.
Pitfalls in southern Finland have received less attention, often only briefly mentioned in regional histories or archaeological reports.
There is a need for archaeological examination of pitfalls south of the Sami area using updated research methods.
This article aims to use geographic information systems, airborne laser scanning data, and semi-automatic algorithms to examine pitfalls in the Suomenselkä region.
The Suomenselkä watershed separates the lake area of central Finland from the river valleys of Ostrobothnia.
The area features barren lichen heaths on moraine ridges and extensive bog areas, a habitat favored by the Finnish forest reindeer (Rangifer \tarandus \fennicus).
Forest reindeer have been an important game animal in the northern boreal forest zone for millennia but were hunted to extinction by the end of the 19th century.
The forest reindeer was successfully reintroduced in the early 1980s, and the area now has the largest population in Finland, about 1,100 individuals (as of the time of the study).
The presence of pitfall sites recorded in the Ancient Relics Register demonstrates the importance of cervids in local food procurement throughout history.
The research area is limited to the northernmost part of the Suomenselkä region, practically covering the parts located in the province of Northern Ostrobothnia.
The study investigates the possibility of using a semi-automatic identification algorithm with airborne laser scanning data to find new pitfall sites.
It examines how well previously known sites can be distinguished in airborne laser scanning data.
The topographic environment of old and possible new pitfall sites and their possible differences are investigated.
The paper sketches an overall picture of Suomenselkä’s pitfall sites based on the results.
Materials and methods
The research area was delimited by combining geospatial data from governmental interfaces like
Landscape Region classification from the Finnish Environment Institute
Regional Division dataset from Statistics Finland.
The research area covers 7,435.6 square kilometers.
By the spring of 2023, 5 points/m2 laser scanning data covered 6,778.9 square kilometers (91.2%) of this area.
As of March 2023, 731 archaeological sites had been registered in the Ancient Relics Register, including 25 “pure” pitfall sites.
Previous studies show that prehistoric settlements may have a spatial connection to pitfall sites according to Paulaharju 1970: 123.
The number of pits varies across sites.
Fifteen sites have less than eight pits.
Seven sites have 15–34 pits.
Three largest sites:
51 pits at Sievi Lietejarvi
93 pits at Pyh¨ ant¨ a Lohipuro
469 pits at Sievi Maansyd¨ an, which is the largest pitfall site in Finland covering 1.24 square kilometers.
Predictive model and polygon filtering
The predictive model for pitfalls, created in the LIDARK project, is based on convolutional neural networks (CNN) and applies semantic segmentation in the Python Keras deep learning environment with the TensorFlow machine learning platform.
The model uses a 512 × 512 -pixel Topographic Position Index (TPI) image tiles with 25 cm pixel resolution and a 30-cell radius for detection.
These were derived via DEMs from the National Land Survey of Finland’s 5 points/m2 pre-classified laser scanning data distributed as 1 × 1 km tiles using a 50 m overlap with a neighboring tile.
Intersect over union (IoU) was used to measure the model’s performance, while the training was ceased after the IoU-value had not improved in the ten previous epochs or after a total of 200 epochs was reached.
The model was trained with two iterations.
The first iteration used 1,666 pitfalls and dugouts from an area of approximately 78 square kilometers in Finnish Lapland.
The predictive model (#20220131-110831) produced a total of 73,254 detection polygons from the Ropijarvenper area.
The second iteration included 3,337 actual pitfalls together with 9,019 false detections located in areas such as swamps, rivers, and lakes
The result was the second development version (#20220329-151958) of the predictive model.
By the end of 2022, the locations of pit features had been predicted in 37 laser scanning production areas in Northern Finland using this development version of the model.
This resulted in 2.31 million detection polygons, which would take approximately seven months to verify manually.
There are 154,289 detection polygons, located within the limits of the Suomenselka research area.
The data should be filtered in a way that leaves as many potential pitfall sites as possible, especially pitfall rows.
There are two strategies for filtering the polygons that predict the location of pitfalls: filtering based on (1) the characteristics of the data or (2) the associated spatial information such as topographic and landscape context.
The goal is to efficiently identify archaeological pitfall sites that contain multiple remains.
The surface area and the probability value assigned by the algorithm for each detection polygon are very useful parameters for filtering the prediction when the goal is to find entire rows of pitfalls instead of locating individual pitfalls.
The suitable filtering values for this study were obtained by comparing the values from approximately 300 previously known and algorithm-detected features.
a suitable probability threshold value is p > 0.9,
while the lower limit for the surface area can be an integer of 20 m2 or a precise value of 19.63 m2
all polygons with a diameter of less than 5 m are filtered out of the data.
In this study, the filtering values used were p > 0.9 and a > 20 m^2.
Human activity has reshaped the terrain in a way that can lead to false detections. The bases of forest roads and paths are often leveled with soil extracted from the surrounding area, which can lead the algorithm to interpret the resulting pits as pitfalls.
These predictions were removed by cutting the polygons with a 30-meter-wide buffer created around the “Tieviiva (Road Line)” vector line level of the National Land Survey of Finland’s Topographic database.
Ditches and other anomalies were removed with a 10-meter buffer crated around the National Land Survey of Finland’s “Virtavesikapea (Narrow Stream)” vector line level in the Topographic database.
The accuracy of ALS data has been reduced to 0.3 points/m2 by the National Land Survey of Finland in these areas, which significantly complicates the use of identification algorithms.
The 83.9 square-kilometre area around the Haapajarvi depot serving the Finnish Defence Forces’ ammunition and explosives production is a good example of this.
The algorithm is quite keen to identify terrain features that are located on swamps, cliff faces, and boulder fields as pitfalls.
Such erroneous predictions can be easily removed using the “Kallio (Cliff face)”, “Kivikko (Boulder field)”, and “Suo (Bog)” polygon layers of the National Land Survey of Finland’s Topographic database.
Filtering prediction polygons can also be approached in reverse by prioritizing areas where pitfalls are known to occur.
The Archaeological Cultural Heritage Guide (Ranta 2017) states that pitfalls are usually dug in sandy soil in narrow passes delimited by natural barriers such as isthmuses between lakes or areas delimited by rivers and moraine ridges.
The corresponding spatial data set to this description can be found in the Geological Survey of Finland’s Soil Database layer: “Glacier stream-generated and moraine formations”, which covers a total area of 1,157 km2 in the study area.
Locating and processing sites
The initial analysis was based on the same Topographic Position Index (TPI) visualizations made from the digital terrain model (DTM), which the detection algorithm uses to search for sites.
In these visualizations each pixel’s elevation is compared to the mean elevation of surrounding pixels within a 30-pixel or 7.5-meter radius.
Three types of area were examined:
Previously known pitfall sites with their surroundings.
Square-kilometre-sized blocks of laser scanning data where the algorithm identified more than five polygons—practically, this meant going through 431 blocks—and
Glacial moraine ridges and other related soil formations.
All pits detected in the preliminary examination were digitized into the same point file.
Digitization stopped at the edge of the 5p laser scanning data production area, but the possible continuity of the pitfall systems beyond it was evaluated using the 0.5p data-based LiDAR visualization available at the Museokartta website.
A boundary for each identified site was determined with the digitized points, and an index of the UTM1 positioning grid squares intersecting with them was created.
Using the resulting index, new digital terrain models (DTMs) were generated from the original laser scanning blocks with a pixel size of 0.2 × 0.2 m and three new visualizations were produced for each of them.
A new TPI visualization was calculated using the algorithm with a search radius of 15 pixels or 3 m.
In addition, the Relief Visualization Toolbox v. 2.2.1 software was used to perform a sky-view-factor (SVF) and local dominance (LD) analysis, the latter using a search radius of 15–25 pixels or 3–5 m.
Results
A total of 1,194 pitfalls were digitized from the study area, forming 66 sites using the commonly applied 200-meter buffer.
This represents about a 2.5x increase compared to the initial situation.
Their spatial distribution reveals three larger northwest-southeast oriented systems: the Reisjärvi, Piippola, and Tavastkenkä systems.
The Reisjärvi and Piippola systems are closely related to esker formations.
The Reisjärvi system is the largest, most coherent, and has the most pitfalls.
The total length of the Reisjärvi system is approximately 37 km.
A six-kilometre stretch with 80 pitfalls is in the province of Central Finland.
Several rows totalling 538 pitfalls are found in Northern Ostrobothnia.
The previously known nine pitfall sites make up only about 7 km of the total length of the system.
The Sievi Maansydan is located near the northern end of the Reisjärvi system.
The remaining pitfalls located in the municipality of Reisjarvi are evenly divided into the northwestern row (N = 187) and the southeastern row (N = 226) of Lake Reisjarvi.
Two other larger systems are in the northeast part of the research area.
The Piippola system contains 117 pitfalls over about 30 km.
The Tavastkenkä system has 305 pitfalls over 50 km.
The pitfalls of the Piippola ridge may have been created by different communities at different times utilizing the same geological formation.
The Tavastkenkä system is not located on a single esker formation but consists of five pitfall rows, of which some have been dug into fault ridges and shore deposits.
Two pitfall sites enlisted in the Ancient Relics Register were virtually invisible not only to the algorithm but also in the visualizations.
At the Pyhant a Suksikangas site, a total of 15 pitfalls have been reported from a combined result of three field surveys, over about 1,700 m.
Only six detection polygons falling into three different laser scanning blocks had been preserved after filtering within the Suksikangas site, and the presumed pitfalls also appear vague in the various visualizations.
Therefore, the site was excluded from the calculations presented above.
In the case of the Haapajarvi Lohijoki site, which is reported to contain 25 hunting pits, the main body of the site falls into the Haapajarvi depot area regulated by the Territorial Surveillance Act.
Discussion
The reliability of the site and pitfall identifications can be assessed by examining the Suomutunturi (-Morottaja) pitfall system.
Previously, 156 possible pitfalls had been identified in the area based on the LIDARK algorithm predictions and basic TPI visualizations.
In the summer of 2022, the archaeologist at the Lapland Provincial Museum, Mr. Jari-Matti Kuusela, determined their ground truth in the field.
The results show that the reliability of the virtual identifications was good, i.e., almost 80%.
Based on the results from the Suomutunturi area, it can be inferred that most pitfall identifications as well as observations regarding the continuity of pitfall systems over water bodies and bogs are real, even if about one-fifth of individual sub-target identifications would be incorrect.
The results also show that the surroundings of known pitfall rows and systems should be surveyed using various visualizations created from ALS-data as soon as such data becomes available.
Discoveries made from visualizations can also link together multiple sites that were earlier considered to be separate entities.
Evidence on the productivity of such re-evaluation can also be pointed out immediately southwest of the research area, where the Yli-Lesti/Kasalankangas pitfall site on the southern shore of Lake Lestijärvi has hundreds of pitfalls along a several-kilometre-long esker ridge.
Contextualized with visualizations created from ALS-data, the site turns out to be in the mid-section of a 34-kilometer-long and 593-pit system.
This strengthens the view that large pitfall systems are not only a characteristic of Upper Lapland but systems containing several hundreds of pitfalls can be found even further south in Finland.
More important than the number of pitfalls, pitfall rows, and systems is understanding how the new findings affect the overall picture regarding them in the northern part of the Suomenselka watershed.
Excluding the Piippola system, these discoveries mainly complement what was already known.
The Reisjarvi system becomes more condensed and elongated with the addition of new sites, and the Tavastkenka system is now conceivable.
As pitfall systems are located on glaciofluvial ridges, many interfering factors make it difficult to understand the entire complex.
These formations are popular pathways for humans and have been utilized, for example, in the construction of Finland’s road network.
The soil of glaciofluvial ridges is also a suitable raw material for various purposes, especially when extensive lateral sand formations are present.
The central part of the study area, which belongs to the Pyhäjoki River catchment area, stands out in the analysis due to the low number of pitfall sites.
This can be explained by the lack of suitable habitats and hunting opportunities for cervids, especially forest deer, in the area.
The area lacks lakes and extensive bogs that significantly regulate the movements of forest deer or long glaciofluvial ridge formations that are central routes for them.
Based on previous research and the spatial analysis conducted here, ridges are also the most typical environment for the location of pitfalls.
The only exception to this is about a 20 km long esker formation west of Lake Pyhäjarvi, where pitfalls are almost completely absent.
The survey of toponyms executed with the Finnish Land Survey’s Topographic database further supports the observation that the central part of the study area stands out as an area less suitable for cervids.
The area lacks Finnish toponyms starting with both “peura-” (wild deer) and “hirvi-” (elk).
These names occur almost in equal numbers across the research area (31 for wild deer and 25 for elk), and their regional distribution is virtually identical.
However, these toponyms do not correlate spatially with pitfall sites, as there are no occurrences within a kilometer radius of any of them.
The distribution of elk names is more uniform than that of wild deer names.
The abundant (N = 20) “hangas-”names—a word signifying a trapping fence—do not relate to pitfall rows but appear rather separate from them in the research area.
Pitfalls are usually associated with populations that have earned their livelihood through cervid hunting, namely the Sámi people.
Toponyms indicate that the Sami-speaking population once inhabited even the southern regions of Finland.
Toponyms with a root in Sami words related to wild or domesticated reindeer (fi. poro) were sought in the Topographic database.
The results were limited to one occurrence of the word “reindeer” and to a cluster of toponyms with the prefix Tolvan-, derived from the Sámi word for the running of reindeer, on the southern shore of Lake Pyhajärvi.
Yet, the group of seven pitfalls located three kilometres east of Cape Tolvanniemi is the only site potentially associated with the Sami based toponyms in the research area.
If the environmental conditions have remained the same, wild reindeer have kept their routes from century to century.
Therefore, it is likely that the labor-intensive pitfall systems have been upkept in the same areas from prehistoric times to the early modern era.
Because of this assumed regional and temporal continuity, one ought to ask whether the pitfall systems outlined in this article are really “systems” that were used at the same time.
The location of the pitfalls may have varied regionally.
Alternatively, old pitfalls may have been renovated to be more functional, or a new, less labor-intensive pit has been dug next to a poorly performing one.
Archaeologists can only determine the temporal arc of a system with datable material obtained through invasive research.
Problems arise then, for example, regarding the association of dated organic material with the period of use of the pitfall.
In Finland, except for northernmost Lapland, another current problem is the small number of radiocarbon-dated pitfall sites, which does not allow for similar extensive statistical analysis of old results as in Sweden.
An alternative way to obtain indicative information about the dating of pitfall rows or systems would be to examine the age of nearby settlement sites, as they were often located close to each other for the maintenance and monitoring of the pitfall systems.
A total of 115 ancient dwelling sites are located within the 5-kilometer radius of the research area’s pitfall sites.
Nearly 2/3 of them are from the Stone Age, while 44 dwelling sites date to the historical period.
However, there are no known Bronze or Iron Age sites in the area due to their low detectability, while, at least in Sweden, these have been identified as periods of intensive use of pitfalls.
Conclusions
Using semi-automated detection algorithms that utilize ALS-data, it is possible to find new pitfall sites, even in areas that have recently been the subject of archaeological surveys.
Previously known sites are generally distinguishable from ALS-data, but there will be most certainly cases also in other research areas, where the site is either difficult or impossible to detect with algorithms or directly from visualizations.
As for the geographic location of the sites, no significant difference was observed between previously known and newly discovered ones, rather the location of the new sites confirmed previous interpretations of prehistoric and historical cervid hunting with pitfalls.
Looking ahead, the picture of pitfall sites in Finland will gradually become clearer with the yearly releases of new ALS-data from the National Land Survey of Finland.
Although ALS-data paired with artificial intelligence makes it relatively easy to locate pitfall sites, research should not focus solely on finding sites, as not all sites are likely to be found even with these methods.
There is still much to be explored regarding the use and temporal stratification of pitfall sites that are, at least superficially, similar in appearance but may upon closer examination provide insight into the social and cultural significance of cervid hunting for past societies.
While this research highlights the potential for new discoveries and insights into past cervid hunting practices through the combination of ALS-data and AI technologies, the addition of multi-sensor data fusion techniques may also prove to be useful in improving the accuracy and efficiency of pitfall site detection.
The potential to investigate pitfall rows and systems that extend over national borders, such as those between Finland and Sweden, or Norway, by combining ALS-data with AI should not be overlooked either, as such efforts could provide valuable insight into the prehistoric and historic practices of cervid hunting and human movement across the northern regions.
Artificial intelligence (AI) is increasingly used to locate and identify archaeological sites using airborne laser scanning (ALS) data and satellite images.
Deep learning algorithms identify patterns like building remains and human-dug pits, aiding in classification, dating, and valuation of cultural significance.
The LIDARK project (2021–2022) is an example of AI in Finnish archaeology, which developed an algorithm for detecting ancient pitfalls.
The study uses geographic information systems, airborne laser scanning data, and semi-automatic algorithms to examine pitfalls in the Suomenselkä region, focusing on the Finnish forest reindeer (Rangifer \tarandus \fennicus).
The research area is limited to the northernmost part of the Suomenselkä region, practically covering the parts located in the province of Northern Ostrobothnia.
A predictive model for pitfalls, based on convolutional neural networks (CNN), applies semantic segmentation, using Topographic Position Index (TPI) image tiles.
Filtering strategies involve using characteristics of the data or associated spatial information to identify archaeological pitfall sites efficiently.
The initial analysis uses TPI visualizations, comparing each pixel’s elevation to the mean elevation of surrounding pixels.
A total of 1,194 pitfalls were digitized, forming 66 sites, revealing three larger northwest-southeast oriented systems: the Reisjärvi, Piippola, and Tavastkenkä systems.
The reliability of virtual identifications was tested in the Suomutunturi area, showing almost 80% accuracy.
Discoveries made from visualizations can link together multiple sites that were earlier considered separate entities, highlighting large pitfall systems even further south in Finland.
Using semi-automated detection algorithms with ALS-data makes it possible to find new pitfall sites, confirming previous interpretations of cervid hunting practices.
Research should explore the use and temporal stratification of pitfall sites, examining the social and cultural significance of cervid hunting.