Monitoring of the Rehabilitation of the Historic World War II US Air Force Base in Greenland

Abstract

After the end of World War II, many military air bases in Greenland were abandoned with all the material left in place. One of these sites was the Bluie East Two military air base. A specific feature of this area is that it contained thousands of old barrels formerly used for fuel storage. In 2019, a rehabilitation of this area began. A few months prior to the rehabilitation, our expedition visited the area and mapped it using an unmanned aerial vehicle (UAV). This made it probably the latest documentation before the start of the sanitation. The aim of such mapping was to estimate the number of barrels in a given location. The second objective was to monitor the progress of the rehabilitation over the years. For this purpose, satellite data were acquired for the years 2019 through 2022. A supervised classification was performed to automatically detect sites with barrel occurrences, which enabled subsequent change detection. We conclude that a total of 33,786 old barrels were located in the investigated area in 2019. However, we suggest this number is a lower estimate of the actual number of barrels due to the factors we mention in our paper. The results further indicate that between the years 2019 and 2022 the barrels were removed from more than half of the area.

1. Introduction

Despite its inhospitable environment, numerous remains of settlement can be found in the area of the Arctic Circle. Apart from human settlements, these are mainly mining sites and military installations [1]. Many of these were established during and after World War II (WWII) throughout Alaska, the Canadian Arctic Circle, Greenland, and Iceland as defense installations, most of which are now abandoned [2].

As an example, Camp Century is an abandoned military base established beneath the surface of the northwestern Greenland Ice Sheet in 1959. During its abandonment, it was assumed that due to its location beneath the surface and ongoing snowfall the remains of chemicals and other waste would be conserved. However, as climate change progresses, modeling of the ice sheet suggests that this may not be the case [3].

It is also believed that up to 20 aircraft (many from WWII) are buried under the Greenland ice sheet [4]. Search and recovery work at these sites is important for both glaciological aspects (ice velocity, ice-flow modeling) and glacial archaeology [1]. In addition, most of the previously mentioned studies are consistent about the environmental hazard which these sites represent. Many of the abandoned sites were left behind with a considerable amount of solid waste which might affect local ecosystems [2]. It can be said that the military sites represent the archaeology of the conflict between the United States (US) and Germany, such as graves, fire sites, etc. These cultural remains should be documented and studied as such [1].

For its tactical location between the United States and Europe, the US built some of their air military bases in Greenland during WWII. Their purposes were meteorological and logistical, as well as for refueling aircraft heading from America towards Europe [5]. Figure 1 displays the location of such air bases. On the east coast there were Bluie East One (Torgilsbu), Bluie East Two (in the Ikateq canal near Kulusuk), Bluie East Three, Bluie East Four (Ella Island), and Bluie East Five, which was occupied by German forces in 1943 and destroyed by an American air raid, similarly to the German meteorological base on Sabine Island [6]. The name ’Bluie East’ is a combination of English and Greenlandic. Bluie East was the code name for a string of planned United States Army Air Corps (USAAC) stations in 1941. The primary purpose was to report current weather during the war.

Due to its remoteness and challenging weather conditions, the research in the Arctic area does not appear very extensive. Generally, Greenland research using remote sensing data is focused mostly on glacier and ice monitoring, specifically ablation zones of the Greenland Ice Sheet [7]. Some studies use data from UAVs [8,9] or data from the MODIS satellite [10]. A relatively large number of publications deal with the use of synthetic-aperture radar (SAR) systems, either for measuring ice velocity or for surveillance, change detection, moving target identification, and high-resolution imaging in Arctic regions [11,12,13,14]. A combination of Sentinel-2 and Landsat-8 data were used to monitor supraglacial lakes in Greenland [15], and hyperspectral data were used to map minerals [16]. Lastly, some of the historic features were documented by means of aerial photogrammetry [17]. The transformations of historical sites were assessed based on seven photo-couples of selected features in Antarctica and Svalbard.

As it appears, there is only a small number of publications mentioning the Bluie East Two air base. Those are mostly related to the military and its role during the Cold War. Some of these even mention the existence of “the thousands of old oil barrels” [18]. Another publication from 2018 comments on the planned “clean up” of the air bases Bluie East Two and Bluie West Four [19]. However, none of these are interested in the change detection happening in the last few years or in using remote sensing data for this monitoring, which happens to be the focus of our research.

The Czech Technical University in Prague, Faculty of Civil Engineering (CTU FCE) team has been focused for a long time on historical monument and landscape documentation using digital photogrammetry, laser scanning, and geophysical methods. In 2003, a cooperation was established between German universities and CTU FCE, department of Geomatics, focused on 3D documentation. This cooperation was newly targeted on Arctic research in 2014. The last expedition to Greenland was realized in 2019. During this international expedition, the research was focused on glaciology, global warming monitoring, biology, and geography. The CTU FCE part was focused on glacier movement monitoring and the documentation of the Bluie East Two air base using aerial photogrammetry [20].

With regards to the documentation of the base, there is little information on the internet and searching the archives takes a long time. We focused on the Danish archive of historical aerial photographs, and we failed here, as until now we have not found historical relevant data in cooperation with the Danish side in this area. We also tried to find historical information on the internet and by querying the Air Force Historical Research Agency. We can, therefore, say that our documentation shows the latest state before the commencement of its liquidation.

In this paper, we focus on the remnants of WWII from a different point of view. We describe the Bluie East Two air base using a combination of UAV and satellite imagery. Based on the satellite data classifications, we aim to monitor the development of the rehabilitation and to calculate the area occupied by the barrels. Further, we focus on a precise estimation of the barrel count by comparison of the results of the satellite data classification with the UAV data.

2. Materials and Methods

2.1. Site Description

Some of the former US air bases in Greenland, constructed during WWII, were rebuilt as Danish/Greenland’s international airports, but most of them were abandoned, which is also the case for Bluie East Two. As the name suggests, the air base is located on the east coast of Greenland on the Ikateq Channel, whose name translates to shallow water. It is approximately 60 km north-east from Tasiilaq, the nearest settlement, one of the largest on the east coast. Tasiilaq lies on an island and the nearest airport is Kulusuk, again one of the former military airports (Figure 2). The abandoned US base consists of a runway and surrounding areas with warehouses and hangars.

In April 1941, the US assumed responsibility for the defense of Greenland and, therefore, planned to build air bases there. This was not easy given the surface topography and meteorological conditions. In the south-eastern region, in 1941, an expedition led by Frederick E. Crockett managed to find a suitable site for an air base in the high mountains around the village of Tasiilaq (formerly Angmagssalik). The base served as an alternate airfield as well as for meteorological, navigational, and rescue purposes. The Bluie East Two base was in operation between 1942 and 1947 [21].

After leaving the area at the end of WWII, the US handed over the base to Denmark. The site was left with a lot of automotive equipment, a hangar, living and storage areas and, most importantly, a huge amount of metal aviation fuel barrels (Figure 3). Some of these remains have been documented using close-range photogrammetry (Figure 4). Estimates of the number of barrels vary—from tens of thousands [22] to 200 thousand [23]. More recently, the site has been an attraction for infrequent tourists, who usually visit the nearby Rasmussen and Karale glaciers. For a long time, there has been a question of who will take care of the removal of the material left behind after many years.

It is just a coincidence that after a long period of negotiations, the removal of the remnants of the base began in September 2019, just after our expedition and measurements. This means that our detailed on-site condition measurements were the last to document the condition of the base.

The area was documented using UAV data in July 2019. These data were preprocessed and classified to distinguish areas covered by the barrels. In addition, we decided to perform a similar classification with the use of satellite data. This analysis was carried out in the same and the following years to observe the changes in the surroundings of the base.

2.2. UAV Detailed Mapping

The UAV measurement was carried out from 18 July to 19 July 2019. During our two-day survey, the winged eBee UAV was used for the documentation of the abandoned US base and nearby neighborhood. Two flights were carried out, both approximately 35 min long at an altitude of 130 m. About 650 overlapping images were captured (Figure 5). The total documented area was 2.2 square km. Unfortunately, based on a dysfunction of the primary camera, the area was captured by a modified backup compact camera Canon PowerShot ELPH 110 HS NIR, 4608 × 3456 pixels with an unusual combination of blue (B), green (G), and near-infrared (NIR) bands (Figure 6).

2.2.1. Preprocessing

Data processing was performed using the Agisoft Metashape software and Pix4D. The absolute camera position and orientation uncertainties are displayed in

Table 1. Absolute camera position and orientation uncertainties (adapted from Pix4D software project report).

	X (m)	Y (m)	Z (m)	Omega (Degree)	Phi (Degree)	Kappa (Degree)
Mean	0.043	0.042	0.053	0.012	0.011	0.005
Sigma	0.012	0.012	0.014	0.003	0.003	0.002

. The absolute geolocation variance of image positions is displayed in

Table 2. Absolute Geolocation variance. The geolocation error is the difference between the initial and computed image positions. Note, the image geolocation errors do not correspond to the accuracy of the observed 3D points (adapted from Pix4D software project report).

	Geolocation Error X (%)	Geolocation Error Y (%)	Geolocation Error Z (%)
Mean (m)	0.03	0.04	−0.03
Sigma (m)	0.60	0.48	0.75
RMS error (m)	0.60	0.48	0.75

. Based on our UAV measurements, we obtained a very detailed orthophoto with ground sample distance (GSD) 5 cm and digital surface model (DSM) with GSD 10 cm (Figure 7). Both products were georeferenced based on an on-board global navigation satellite system (GNSS). Ground control points (GCP) were not measured because of a lack of time and precise instruments such as real-time kinematic (RTK) GNSS, which does not work without correction and which was not at our disposal in Greenland. Therefore, the georeferencing accuracy is given by conventional GNSS without corrections and it can achieve an absolute positional accuracy of around 3 m. However, the relative accuracy is much higher, around a decimeter inside the block (

Table 3. Relative geolocation variance. The percentage of images with a relative geolocation error in X, Y, Z (adapted from Pix4D software project report).

Relative Geolocation Error	Images X (%)	Images Y (%)	Images Z (%)
Point (−1.00, 1.00)	98.77	98.03	95.09
Point (−2.00, 2.00)	99.75	99.75	100.00
Point (−3.00, 3.00)	100.00	100.00	100.00
Mean of geolocation accuracy (m)	1.11	1.11	1.49
Sigma of geolocation accuracy (m)	0.15	0.15	0.12

) and about twice that on the edges.

Training Data

Training data were collected manually, as the land use land cover (LULC) product with suitable spatial resolution was not available. Four basic classes have been identified in the area:

• Barrels—the most relevant class for the classification, but also the least represented in the imagery data. This class was located mainly in the coast area, along the shore.
• Vegetation—general vegetation cover, where low vegetation such as shrubs, lichens, and mosses dominate.
• Bare land—parts of the area not covered by vegetation, most likely used for transportation (road-like features, runway), possibly stone beaches and rocks.
• Water—margin of the Ikateq canal, lake, and possibly streams.

The training areas were created with the emphasis on being as homogeneous as possible. Furthermore, sufficient counts of areas per class were obtained.

Ancillary Data

Due to the unavailability of the red (R) band, it was not possible to calculate the normalized difference vegetation index (NDVI) channel [24]. Instead, green normalized difference vegetation index (GNDVI) was used (Equation (1)) [25]. Further, the normalized difference water index (NDWI) was also incorporated as it can also be beneficial for distinguishing barrels from shrubs (Equation (2)) [26].

$$\mathrm{G}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}\mathrm{=}(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{G}),$$

(1)

$$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}=(\mathrm{G}-\mathrm{N}\mathrm{I}\mathrm{R})/(\mathrm{G}+\mathrm{N}\mathrm{I}\mathrm{R}),$$

(2)

where NIR is the reflectance of the near-infrared band and G is the reflectance of the green band.

2.2.2. Classification

The data were classified using a machine learning technique. The random forest classifier was selected as it had proven itself as a time efficient and accurate tool [27,28]. The five optical bands and channels (B, G, NIR, GNDVI, and NDWI) were used for the classification. The classification was processed using the GRASS GIS software [29], specifically, using the r.learn.ml tool with default settings.

2.3. Satellite Imagery Monitoring

The 2019 UAV mapping results were used for comparison with high and very high resolution satellite data. In addition to 2019, data for the years 2020 through 2022 were also acquired to further analyze the situation after the beginning of the rehabilitation of the territory.

2.3.1. Preprocessing

Satellite Data

The available satellite data included three very high resolution scenes (for the years 2019 to 2021), and one high-resolution scene for 2022. Due to the unavailability of data from the same sensor for each year, whether due to cloud cover or other factors, imagery data from a total of three satellites were used: WorldView3, GeoEye1, and SPOT-6. The image data consisted of both multispectral data (B, G, R, and NIR bands) and panchromatic data, which was used to create a pansharpened image. In this way, we were able to use the advantages of multiple bands with the more accurate resolution of a panchromatic band. The resolution from 2019 to 2021 ranged from 0.3 m to 0.5 m. Unfortunately, such high-resolution data were not available for 2022; either no data were acquired or the area was covered by snow or cloud cover. Therefore, we had to work with data with a resolution of 1.6 m, which unfortunately does not capture such detail. A summary of the imagery data used is provided in

Table 4. Comparison of individual scenes.

Sensing Date	Satellite	Spatial Resolution (Panchromatic Band)
31 May 2019	WorldView3	0.3 m
1 August 2020	GeoEye1	0.4 m
30 August 2021	GeoEye1	0.5 m
18 July 2022	SPOT-6	1.6 m

.

To simplify the processing, we selected only the areas that contained the barrels. The areas of interest (AOIs) were chosen using a combination of visual reconnaissance of imagery data and preliminary classification of the entire scene. Based on this process, three AOIs were identified (Figure 8).

Further processing was performed in order to detect changes between the years. The imagery of 2020 was determined as a reference year, as the data were in the best condition (illumination, resolution, distortion, etc.). Firstly, resampling was performed to a spatial resolution of 0.4 m for the 2019 and 2021 scenes. The nearest neighbors algorithm was used for this purpose. Secondly, a georeferencing task was carried out. The georeferencing turned out to be challenging, as there were mainly natural features in the area and not many reference points. Therefore, mostly larger stones were used as the reference points. Due to the lower resolution, the 2022 data were only georeferenced and not resampled. The year 2022 was not subsequently included in the change detection, because it could not be clearly determined whether the barrels were present. This input served only as an approximate assessment of the current status.

Training Data

The training data preparation was a similar process as for the UAV data with a few differences. In addition to four classes, the satellite data also included class snow and ice, which was present in the 2019 imagery. The training areas were iteratively improved based on the classification results (visual evaluation). Because of the different satellite data for each year, training areas were collected for each scene separately.

Ancillary Data

Similarly to the UAV data, some indices were computed from the satellite data. In contrast to the UAV data, the red channel was available (Figure 9). Therefore, the NDVI was calculated (Equation (3)). We also created the previously mentioned NDWI channel. The GNDVI was not necessary, as the red channel was available.

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}),$$

(3)

where NIR is the reflectance of the near-infrared band and R is the reflectance of the red band.

The entire preprocessing of the 2019 to 2021 satellite data, including the following classification, is shown in Figure 10. The 2022 data were processed using the same methodology, except the resampling was omitted as it was not planned to use these data for change detection due to their lower resolution.

2.3.2. Classification

A similar classification method was applied as in the case of the UAV data. This time, six optical bands and channels were used (B, G, R, NIR, NDVI, and NDWI). Each year was classified separately using corresponding training areas. Due to its size, AOI1 was classified first as the largest of the selected AOIs. The random forest model used for classification was saved and subsequently used to classify AOI 2 and AOI 3.

According to the results obtained, it turned out that transferring one model between years was not possible. The different radiometric values contained in the pixels across the scenes were the reason for classifying each year separately. Thus, the same classification process was applied to all four years.

2.3.3. Post-Processing

The post-processing part involved comparison of the classification results for various years. As mentioned earlier, only the data from 2019 to 2021 was used for change detection, and this analysis was based purely on satellite data. We chose a quantitative approach of comparing raster pairs. Specifically, the classification-based change detection method [30,31] was used for this purpose. However, before comparing classification results, a sieve filter was run on each classified raster to remove isolated and obviously misclassified pixels or pixel chunks, as similar works suggest [32]. The number of ten pixels was chosen as the threshold.

Difference maps were created using raster algebra. Each pair of raster images was compared pixel by pixel. Since we were only interested in changes related to barrels, the classified and filtered results were reclassified to values 1 and 0, where 1 represented the barrels and 0 represented everything else. These reclassified rasters were then subtracted from each other, making it easy to track where sanitation was occurring and where no change happened.

2.4. Barrel Count Estimation

In addition, the estimation of the barrel count was carried out based on the photogrammetry documentation from the year 2019. The specific estimate of the number of barrels was based on previous work [33], where piles of barrels were manually traced over the created orthophoto. Parameters such as the fill rate of barrel piles, percentage of lying and standing barrels, and height of each pile in multiples of barrels were estimated, from which the total number of barrels was calculated. The calculation showed that the total area of 14,206.1 m² contained approximately 30,162 barrels, which is equal to about 2.12 barrels per square meter. Since there was no point in performing the same work a second time, and the UAV data were only available for AOI 1, we used this ratio to estimate the number of barrels in the total area classified as barrel sites.

The satellite image processing workflow was developed using open source software. We used QGIS v. 3.18 [34] for the pre- and post-processing parts and GRASS GIS v. 7.8.5 for the classification. The entire post-processing methodology is illustrated in the following flowchart (Figure 11).

3. Results

Here, we summarize the results obtained from the proposed methodology. Firstly, we describe the results from the UAV measurements. Secondly, the results of the satellite data classification are presented. Lastly, change detection and barrel count findings are shown.

3.1. UAV Data Classification

The classification of the UAV data did not provide a satisfactory result. Although the classification accuracy was relatively high (

Table 5. Classification accuracy of UAV data (%).

		Reference
		Barrels	Water	Bare Land	Vegetation	Precision	Recall
Classification	Barrels	88.21	0.00	1.14	7.98	85.08	88.21
	Water	0.00	100.00	0.00	0.00	100.00	100.00
	Bare land	2.28	0.00	98.17	1.48	98.20	98.17
	Vegetation	9.51	0.00	0.69	90.54	92.66	90.54

), the resulting map contained a significant amount of noise, and a number of sparsely vegetated areas classified as “barrels” according to the visual inspection (Figure 12). Despite several modifications of the training areas, we were not able to eliminate this noise. The reasons for the failure of the UAV data classification are described in detail in Section 3.2 and Section 4.1.

3.2. Satellite Data Classification

The proposed methodology was applied for classification of multiple AOIs. The barrels were located mainly in the surroundings of the former air base and generally near the water with respect to its accessibility. Further, they were located along the paths at various distances from the base (Figure 13).

The classification evaluation for the years 2019 through 2021 is summarized in

Table 6. Mean classification accuracy, year 2019 (%).

		Reference
		Barrels	Water	Bare Land	Vegetation	Precision	Recall
Classification	Barrels	96.13	0.01	0.33	0.00	95.06	96.13
	Water	0.05	99.70	0.07	0.00	99.79	99.70
	Bare land	3.82	0.29	99.59	0.09	99.62	99.59
	Vegetation	0.00	0.00	0.01	99.91	99.98	99.91

,

Table 7. Mean classification accuracy, year 2020 (%).

		Reference
		Barrels	Water	Bare Land	Vegetation	Precision	Recall
Classification	Barrels	95.99	0.00	0.25	0.00	95.29	95.99
	Water	0.11	99.72	0.05	0.00	99.84	99.72
	Bare land	3.89	0.27	99.69	0.06	99.68	99.69
	Vegetation	0.00	0.00	0.00	99.94	99.98	99.94

and

Table 8. Mean classification accuracy, year 2021(%).

		Reference
		Barrels	Water	Bare Land	Vegetation	Precision	Recall
Classification	Barrels	90.08	0.01	0.15	0.00	94.76	90.08
	Water	0.69	99.44	0.09	0.00	99.70	99.44
	Bare land	9.16	0.49	99.63	0.73	99.30	99.63
	Vegetation	0.08	0.06	0.13	99.27	99.54	99.27

. For the purpose of evaluation, a part of the reference data were left out to conduct this assessment. Here, we display the classification results across all AOIs (mean values). Overall, the results over different years are comparable, with the precision exceeding 99% for most of the classes. In the barrel domain, we can see misclassification of between almost 4 and 9%, mostly with regards to the bare land.

Based on the visual comparison, the classification of the satellite data performed significantly better than the classification of the UAV data. As there was a difference in the analysis between the satellite and UAV data based on bands, we inspected the feature importance for satellite data classification across years (Figure 14).

As shown in Figure 14, for 2019, the NDVI band has the largest impact on classification, while the R band is the second least important. Once the NDVI band was removed from the classification, the R band became the most important. For the following years, the effect of the R band is not as significant, but it is still influential, either in its own form or within the NDVI. The reason why in 2020 and 2021 the R band or the NDVI did not have as much impact on classification as in 2019 is probably because of the sensing period of each scene. While the 2019 scene was taken in late May, the 2020 and 2021 scenes were taken in early and late August, respectively, when the vegetation period in Greenland is already coming to an end. However, the importance of the R band can also be observed in Figure 15, which shows the results of the 2019 satellite data classification. Firstly, all available bands were used, along with NDVI and NDWI. Secondly, we included only the bands and indices that were available when processing the UAV data. The results show that the absence of the R band caused the apparent misclassification of a number of sites into the ’barrels’ class. That made any subsequent statistical evaluation impossible without including the misclassified sites in the results.

The third fact confirming the importance of the R band comes from the analysis of the spectral reflectance. Figure 16 shows a comparison of the spectral curves of the barrels and three randomly selected sites misclassified as barrels in the 2019 satellite data. As can be seen, the largest difference in spectral reflectance is for the R band, while in the other bands the reflectance values are relatively close, and, therefore, the R band is crucial for this type of analysis.

3.3. Barrel Sites and Change Detection

Based on the classification of the satellite imagery, we were able to estimate the area occupied by the barrels. However, the fact that other rusty metal objects, such as an old hangar and trucks, were in the same class as the barrels also had to be taken into account. These areas were, therefore, masked out prior to statistical evaluation. Figure 17 displays the area occupied by the barrels in individual AOIs (the first three column groups from the left) as well as the total area in the entire satellite imagery (the three columns on the right). It is apparent that the largest amount of barrels was located in AOI 1. The barrels covered more than 1.5 ha in AOI 1, whereas in AOIs 2 and 3 that was only a few acres. However, AOIs 2 and 3 were smaller in area as well. Overall, there were nearly 1.6 ha occupied by the barrels in the entire study area in May 2019. The figure also suggests that by August 2021 the amount of barrels was reduced to approximately 0.7 ha (orange column). From this we can conclude that more than half of the barrels were removed.

There were various types of changes across the AOIs, where removing the barrels was the most noteworthy one (Figure 18). Apart from that, there were also changes which were more likely a transfer of the barrels from one place to the other, than a total removal.

Figure 19 summarizes the interannual changes with regards to the area occupied by the barrels. During the initial stages of our analysis, the interpretation was focused mostly on the area of the air base. Visually there was a much larger difference between 2020 and 2021 than between the previous years. Therefore, it was a surprising finding that quantitatively a larger change happened between the years 2019 and 2020.

Over 0.6 ha of barrels were removed in this period. Based on the visual inspection, the majority of the removed barrels were located along the roads and distributed in smaller chunks (Figure 20). They were not apparent at first, because they were located further from the base, and were partly blending in with the vegetation. In addition, a small portion of the stack at the air base was removed as well. In the following year (2020–2021) the most notable difference was observed with regards to the large stack at the air base. A removal of more than 0.2 ha of barrels was captured during that time.

With regards to AOIs 2 and 3, the numerical analysis revealed that the area of barrels expanded rather than reduced, by 68 m² and 81 m², respectively. From a visual inspection, this slight difference was more likely caused by a georeferencing error than an actual change in barrels. As mentioned before, there were not many ground reference points, which made the georeferencing task very challenging. Apart from that, the most visible change observed in AOI 2 was a replacement of one pile between the years 2020 and 2021 (Figure 21). The barrels occupied the smallest area in AOI 3 and almost no changes were observed there.

Overall, the most noteworthy and relevant differences were observed in AOI 1. Over the years 2019 and 2021, barrels were removed from an area of more than 0.8 hectares.

3.4. Barrel Count

A barrel count estimation for the year 2019 was performed in order to document the state of the area before the sanitation started. Firstly, we considered the information obtained from the UAV data. Based on manual tracing of the barrel piles and estimation of the parameters mentioned earlier, the number of barrels was estimated to be 30,162 with a total area of 14,206.1 m². Using the ratio of these values (2.12), we determined an approximate number of barrels based on our analysis. This resulted in an estimate of 33,028 barrels for AOI 1, which corresponds in location to the area covered by the UAV, and 33,786 barrels in total for all three study areas combined. Unfortunately, we did not have any material to perform similar analyses for the following years. Therefore, the analysis of the continued removal and estimating the number of remaining barrels could only be performed on the basis of satellite imagery. For this purpose the same “barrels per area” ratio was used for all years as in 2019. The numbers of barrels in each year and the comparison between manual and automatic classification are shown in

Table 9. Number of barrels in each year.

Classification	2019	2020	2021
Manual (UAV data)	30,162	—	—
Automatic (satellite data)	33,786	20,278	15,225

.

3.5. Assessment of the Current Status (2022)

As mentioned before, due to the lower spatial resolution, the 2022 satellite data were not used for change detection and they only served as an approximate assessment of the current status regarding barrel sanitation. As Figure 22 displays, complete sanitation has still not occurred and there were still barrel sites in the area as of 18 July 2022.

Based on the classification performed on all three AOIs, the total area covered by barrels and possibly other rusted metal objects was estimated to be 5196.5 m². However, this value should be taken with a reserve, as it was not possible to obtain reliable results with the available data due to the already mentioned lower spatial resolution. Nevertheless, it can be said that if there was a change between 2021 and 2022, it was not as significant as in previous years.

4. Discussion

4.1. UAV Detailed Mapping

The UAV detailed mapping was carried out using a backup camera with a spectral combination of B, G, and NIR bands. This did not turn out to be an ideal combination, as the missing R band caused a misclassification of a number of sites as barrels. Another reason for poor UAV classification results may be the spatial resolution of the data itself. With a pixel size of 5 cm, the details such as barrel shadows or shimmering sun reflection are captured. Because of this the barrels themselves often have different values of spectral reflectance. In land cover (LC) classification, this translates into noise. This problem can be partially solved by resampling the data to a lower spatial resolution. A test was performed with resampled data at 40 cm, but the absence of the R band still did not allow satisfactory results to be achieved. These facts suggest a more suitable approach for processing this data would be an object-based classification rather than pixel-based classification. A pilot experiment was carried out using neural networks (NNs) without satisfactory results. However, with more precise reference data, the NN classification might be applied as well. This domain might need more time and resources for further research. Despite the described drawbacks, the UAV data were used to answer another research question regarding the number of barrels in the area. With a pixel size of 5 cm, no better data are available.

4.2. Classification and Change Detection

The classification workflow of the satellite data and consequent analysis has some limitations. An inaccuracy in image preprocessing was observed when assessing the change detection in AOIs 2 and 3. However, the obtained results did not affect the final number as the values were very small. Therefore, most of the changes observed in AOIs 2 and 3 were probably due to georeferencing as the input data were of a different resolution and oblique angles. In addition, another challenging task was to find a sufficient number of reference points. This type of error is one of the disadvantages of the used method which the authors are aware of.

4.3. Barrel Count

Thanks to the UAV data, we were able to perform the barrel count estimate. We assume that in 2019 there were 33,786 barrels, which corresponds to the number obtained from manual classification (30,162). In previous sections, we mentioned multiple factors which influence the final value. The factor with the largest effect is that in some places the barrels lie in several layers on top of each other, as can be seen in Figure 23. This problem could be solved by using a digital surface model and a digital terrain model (DTM). The height of each pile could be estimated as a difference of these models and, therefore, the number of layers of barrels on top of each other. The DSM was collected based on the UAV data analysis. However, a DTM of sufficient resolution was unfortunately not available.

Given this fact, we assume that the number of barrels we obtained is rather a lower estimate of how many barrels were actually in the area. Other indicators of potential underestimation are the 4–9% uncertainty from the classification results, where there was a slight confusion between barrels and bare land, the fact that the statistical evaluation was performed on filtered classified data, and the uncertainties in the estimation of parameters such as the fill rate of barrel piles or the percentage of lying and standing barrels. On the other hand, based on our results, we do not believe that the number of barrels should have exceeded 60,000. Otherwise, the barrels would be stacked in two layers on top of each other all over the area, which was not observed.

5. Conclusions

A survey was conducted using photogrammetric and remote sensing methods to document and investigate the status of the rehabilitation of the historic WWII US air base Bluie East Two in Greenland. The main objectives were to find out approximately how many old barrels were in the area before the start of the remediation, which began in the autumn of 2019, and how this remediation has progressed over the following years, i.e., how many barrels have already been removed.

The UAV data obtained during the expedition, as well as available satellite data, were used to determine the area occupied by the barrels. The satellite data were classified automatically using supervised classification. The UAV data were primarily used to determine the number of barrels. For this purpose, parameters were estimated that, together with manually classified barrel locations, allowed us to determine how many barrels there are on average per square meter. Subsequently, this ratio was used to determine the number of barrels in the areas that were automatically classified using the satellite data. This process led us to a figure of 33,786 barrels before remediation began. However, we believe that due to multiple factors such as classification accuracy, georeferencing accuracy, and the fact that several layers of barrels lie on top of each other in some locations, our estimate is more of a lower limit of the actual number of barrels left at the former air base.

The UAV data were only available for 2019. Therefore, satellite data were used in order to determine the progress of remediation over the years 2019 through 2022. In 2019, the total area covered by barrels was more than 1.5 ha. Between 2019 and 2020, mainly smaller piles were removed. However, this totaled more than 0.6 hectares, the largest interannual change. Between the years 2020 and 2021, the most significant changes occurred in close proximity to the former air base, where the largest pile of barrels was removed, with a total of just over 0.2 ha removed. For 2022, data were not available with as high a spatial resolution as for the other years. Therefore, this year was not included in the change detection as such a comparison would not be of much value. However, a classification was carried out which showed that complete remediation had still not occurred and an area of approximately 0.5 ha was still covered with barrels or other rusted metal material.

In terms of change detection, major changes took place in the AOI 1. Some barrels were also detected in AOIs 2 and 3, but not many changes were present. Furthermore, a relocation of barrels could be observed instead of total removal of the barrels. However, it is important to note that georeferencing errors and possible misclassified pixels could have affected the change detection results obtained.

Overall, there were almost 1.6 ha occupied by the barrels in the entire study area in May 2019. By August 2021, the area had reduced to just over 0.7 hectares. This suggests that remediation occurred on more than half of the total area covered by the barrels prior to the start of rehabilitation in 2019. If the rehabilitation continues at a similar pace, Bluie East Two air base could be completely rehabilitated by 2024.