The data fusion of the OSM and SACHI datasets resulted in an infrastructure and human-impacted area map with higher spatial detail and coverage than the original datasets. While the polygonal features of SACHI only covered the coastal region with an area of 62 km2, the incorporation of OSM data extended the infrastructure map to encompass the entire state, now covering an expansive 640593 km2.

Furthermore, the integration allowed us to enhance the level of detail regarding the usage categories for various infrastructure features. While we were able to initially assign five LUCAS categories to the SACHI data, i.e., fishing; mining and quarrying; energy production; community services; and recreational, leisure and sport, the inclusion of OSM data expanded this categorization to include an additional eight categories: agriculture; commerce, finance and business; construction; forestry; industry and manufacturing; residential; transport and communication networks; and waste and water treatment (refer to Table  and Fig.  for a detailed breakdown).

This comprehensive categorization enhancement enabled us to refine the generalized approach. For example, we discovered that energy production sites, initially thought of as dominant with an area of 28 km2 in coastal regions, were, in fact, less extensive, covering only 17 km2 across the entire state (see Table  and Fig. ).

However, by incorporating the SACHI dataset, the map now also encompasses small and isolated elements like gravel pads and small paths, which were not mapped by the OSM community but which were successfully derived from the satellites (refer to Sect. ). On the other hand, the integration of OSM data provided a heightened level of detail, enabling clear identification and differentiation of roads and single buildings (see Fig. c).

Looking at the settlement of Nome using SACHI, we identified mining and quarrying as the primary land use category, apart from the transport network. These categories were determined by applying a buffer around each settlement (refer to Sect. ) and assigning it one predominant value (see Fig. b). Combining SACHI with the OSM data enhanced the quality of the transport network, where the streets are clearly defined even within areas with a high density of buildings and other human-impacted areas. It also improved the detail of these usage type categories (Fig. c). We learned that the majority of the settlement's area is actually residential, characterized by houses and recreational areas such as pitches and parks (see Fig. a and c). The OSM data also added detail, where the spatial resolution of the SACHI product derived from Sentinel satellites fell short. For example, the pier in the western area was not captured by the Sentinel satellites, but it was digitized by the OSM community. However, comparing the resulting human-impacted areas and the infrastructure map with aerial imagery from Bing (as accessed via the QGIS plugin OpenLayers; ) revealed that there is a second pier, which did not appear in OSM or in the SACHI dataset. Nonetheless, the true added value of the SACHI dataset is in its information on small features such as extraction pads and others, which only occasionally appear in OSM data.

A closer examination of the Prudhoe Bay area confirmed this observation. Once again, the SACHI dataset showed more human-impacted areas, probably from expanding exploration sites, while OSM offered more spatial detail. Furthermore, at both sites, we found that OSM exhibited higher quality in terms of linear infrastructure objects such as road and railway lines. As mentioned in Sect. , we compared the areas of the linear transport network between SACHI and OSM to evaluate the potential limitations of using OSM data. However, we discovered that the difference in area was only 5 km2 (or 6 % of the total SACHI linear infrastructure area), as shown in Table .

The resulting SIRIUS infrastructure and the human-impacted area inventory not only represents economic activities, but also incorporates fundamental functions for living , including agricultural areas, commercial and residential zones, recreational spaces, waste and water treatment, and community services. We also observed a significant decrease in the number of features with unknown land use types by internally overlaying OSM buildings with non-building OSM information (see Fig.  and Table ). Prior to the internal overlay, the area with unknown land use was 34 km2, whereas, after the overlay, it decreased to only 13 km2 (refer to Table ). This enhanced level of usage type detail allows for various applications, such as risk assessments for energy production facilities and transportation networks as well as evaluations of contaminated sites close to recreational or agricultural areas (refer to Sect. ).

The OA of the confusion matrix represents the ratio of correctly classified pixels to the total number of all pixels (positive and negative, true and false). The OA of the linear infrastructure data of SIRIUS is 0.5. While this value seems relatively low, we need to zoom in on a specific detail: of all 310 true road grid cells of the reference dataset showing a road infrastructure, 241 grid cells, i.e., 78 %, were accurately represented in the SIRIUS dataset (see Fig. a). A visual examination further revealed that, of the remaining 69 true road grid cells supposedly not represented by SIRIUS, 45 (65 %) were captured but with a slight spatial offset (see Fig. a), leading to a false negative when indeed it was only a positional inaccuracy. Taking into account these offset grid cells, the overall accuracy of the SIRIUS dataset improves to 0.69 and the true positive value increases from 0.78 to 0.92, indicating that 92 % of the road infrastructure is mapped in the SIRIUS inventory. All of the SIRIUS road grid cells not mapped in the reference dataset (false positives) were either small tracks, footways or narrow residential roads with widths of less than 10 m, and thus they were not mapped (see Fig. b).

The overall accuracy of the polygonal infrastructure and human-impacted areas of the SIRIUS dataset shows a similarly low value of 0.53. However, the true positive value, representing the ratio of the correctly classified values in SIRIUS to the actual positive values, is 94 % (686 of 731 true polygonal infrastructure grid cells) (Fig. b). Of the remaining 45 false negative grid cells, 13 % were indeed missing, another 18 % occurred again because of a spatial offset and 69 % appeared along the breakwater, protecting the shore (see Fig. a). OSM did not capture this structure, and due to the relatively coarse spatial resolution of the Sentinel sensors, the representation of the breakwater was sparse and patchy in SACHI, leading to an underestimation and a high number of false negatives.

However, substantially distorting the overall accuracy is the high number of false positives: 568 grid cells showed an intersection with the polygonal infrastructure in the SIRIUS dataset (Fig. b), which was not captured in the reference dataset; 23 % of these false positives stem from an overestimation of the airport area in SACHI and an altogether more generous mapping of the area in the OSM data. The eastern part of the runway, for instance, appears revegetated and allows the conclusion that it is no longer in use despite still being represented in the OSM data (refer to Fig. b). However, the highest number of false positives originates from areas affected by human activities represented in the SIRIUS dataset. These human-impacted areas posed a challenge to accurately mapping them for the reference dataset on the basis of the orthophotos alone. Some features, e.g., a playground, were either not visible or were difficult to delineate accurately. Figure c shows an example of a human-impacted area mapped as industrial land use by the OSM community. While the single storage structures are represented in the reference dataset, there was no indication of an enclosed area.

In summary, the low overall accuracy of the polygonal infrastructure data is distorted by a high number of false positives that originate from either an overestimation of the areas (e.g., airport) or a (conceptual) definition of land use (e.g., playground, industrial usage) that is difficult to reproduce with orthophotos alone. However, it is important to note that SIRIUS achieved representations of 78 % for the linear infrastructure and 94 % for the polygonal infrastructure, respectively, of the true infrastructure values.

With the text-mining approach, we successfully extracted additional information from the site reports of the Contaminated Sites Program. The use of regular expressions allowed us to identify dates, abbreviations and references to substances from the DEC glossary or any contaminated medium mentioned in the text. Consequently, we were able to calculate the total cleanup time at inactive sites and provide a comprehensive list of the substances mentioned in the site reports. To assess the accuracy, we retrieved a sample of 10 data entries from the dataset (see Table ). We confirm the successful extraction of dates, following the pattern described in Sect. . The expressions “sites added to database” and “sites closure approved/cleanup complete” were considered to be the first and last action dates, respectively. In the cases where no specific expressions were present, the first and last mentioned action dates were used instead. However, in 491 (6 %) entries, the cleanup duration was recorded as 0 d (for an example, see Hazard ID 2361 of the sample in Table ), and in 214 (2 %) cases, negative values were even reported. This again points to a heterogeneous approach or methodology used by the agency to input data into the database. In these cases, “site closure approved/cleanup complete” was entered on the same date or even before “sites added to database”. For the sample, the retrieval of the contaminants was highly successful, as all substances and containment structures listed in the DEC glossary (see Table ) were found. However, any substances not appearing in the glossary will not be retrieved with our approach. Also, the information regarding the contaminated medium was limited, as the DEC rarely provides details in the “contaminant information” section of the reports. Consequently, we were only able to derive the contaminated medium for 3321 (39 %) of the 8533 sites.

As a first application scenario, we wanted to retrieve the total length of the road and railway lines within Alaska's continuous permafrost zone. As GeoPackage uses an SQLite database container, we were easily able to query spatial information by using the “execute SQL” command in QGIS.

This query provided us with a length of 8456 km for the railway and road networks intersecting the continuous permafrost zone.

Another possible application is to determine the number of contaminated sites per watershed. To achieve this, the user can for example use the QGIS tool “count points in polygon”. We tested this and discovered that the Yukon watershed, which is also Alaska's largest watershed draining into the Arctic Ocean, contained the highest number of contaminated sites, totaling 2256. However, to account for the huge differences in watershed sizes and to normalize the number of sites per area, we further calculated the number of contaminated sites per square kilometer per watershed, showing that the watersheds along the coast of the Beaufort Sea (Fig. a and c) and Kotzebue (Fig. b) depict the highest density of contaminated sites per square kilometer (see Fig. ).

We further derived which land use category or infrastructure type shows the most contamination. For this analysis, we showcase the use of GeoPandas as a third processing option for our GeoPackage. By creating a spatial join between the SACHI_OSM_InfrastructureHIElements and SACHI_OSM_InfrastructureHIElements_RRNetwork (as a polygonal representation) layers along with DEC_ContaminatedSitesAK, we first derived all infrastructure and human-impacted areas and elements intersecting a contaminated site. Next, we dissolved these intersecting elements based on their LUCAS attribute. Subsequently, we counted the number of contaminated sites by examining the points within these dissolved polygons, representing the aggregated LUCAS attribute. For the Python code, see Appendix .

This application example showed that most of the contamination occurs in the land use categories “community services” (into which among others fall, e.g., military installations; see Table ), “transport, communication networks, storage and protective works”, “industry and manufacturing”, and “recreational, leisure and sport” (see Table ).

The resulting inventory on infrastructure and human-impacted areas and elements in Alaska provides a detailed and comprehensive overview of various human activities, encompassing not only economic functions, but also fundamental functions for living. Compared to the original SACHI dataset, we have achieved higher spatial detail and coverage throughout the entire state by incorporating OSM data (see Table  and Fig. ). On the other hand, the SACHI dataset has made a substantial contribution by capturing small elements that had been missed by the mapping efforts of the OSM community. This limitation may be attributed to the peripheral status of Arctic environments within the global OSM network, primarily due to their sparse population. report such a limitation for isolated, smaller communities and with regional variability (e.g., better coverage in North America and Eurasia compared to Asia). This deficit in the mapped regions underscores the necessity for infrastructure products derived from remote sensing images, such as SACHI, as the underlying algorithms used to retrieve these features remain unbiased in terms of area selection. However, as described in Sect. , the algorithms fell short in densely populated areas, which makes distinguishing between adjacent features of different classes – e.g., buildings, roads or extraction pads – challenging. To achieve the spatial detail of the OSM additions, the retrieval of infrastructure and human-impacted features could be enhanced by analyzing remote sensing data with sub-meter spatial resolution. However, this improvement would come at a significant cost, as most of these satellite images are commercial. On a pan-Arctic scale, this approach is nearly impossible due to the large spatial coverage necessary and the associated high-resolution imagery costs . A compromise could involve using satellite imagery from providers that offer educational programs or discounted rates for researchers, such as Planet's Planetscope with a spatial resolution of 4 m . Alternatively, deep-learning models could be leveraged to generate high-resolution images from lower-resolution sources like Sentinel-2 . Considering these challenges emphasizes the need to rely on crowd-sourced map data. These map data can also be generated remotely using accessible Web Map Servers or GIS plugins (e.g., Bing). Using OpenStreetMap as this data source serves as a gateway for this purpose. It establishes a low threshold for non-researchers, including citizen scientists, who can not only map various elements but can eventually also incorporate valuable information on contamination that has not been captured by official environmental agencies, highlighting the unique potential of OSM in this context. In addition, this approach allows the continuous development of suitable tags (the attribute fclass in our data). However, based on our own field visits, we have identified instances where certain areas and elements that contribute to the critical sector of “health and sanitation” are not accurately represented in OSM. For example, the Middle Salt Lagoon in Utqiaġvik (formerly Barrow), which is used for sewage purposes, is labeled “water” in OSM and is thus not included in our SIRIUS dataset. This underlines the need for a comprehensive review of the mapping tags before basing future inventories of critical infrastructure and human-impacted areas on OSM. Fortunately, due to OSM's open design and accessibility, these revisions can be easily implemented. Given that OSM undergoes daily updates through user contributions, the integration of OSM data also facilitates periodic updates within our inventory.

While the linear infrastructure data exhibited low overall accuracy in our Shishmaref test area, about two-thirds of the false negatives resulted from a spatial offset (see Fig. ). This indicates the presence of roads but with reduced positional accuracy, likely due to an image offset between the MACS imagery and the OSM data. All false positives were narrow residential roads or small paths visible in orthophotos but not digitized in the reference dataset, in order to comply with the 10 m Sentinel resolution. Including these narrow features in the reference dataset would have substantially improved the accuracy.

Nonetheless, 78 % of the true road grid cells were accurately represented in the SIRIUS dataset, increasing to 92 % when accounting for offset grid cells. This highlights the effectiveness of OSM in representing linear infrastructure compared to SACHI. OSM distinguishes between roads and adjacent infrastructure areas and includes narrow roads and footways. To improve the accuracy, it could be beneficial to integrate official data from local or federal agencies (e.g., the Alaska Department of Transportation) to evaluate the comprehensiveness of the OSM linear infrastructure data. Further, incorporating the Trans-Alaskan Pipeline would provide a spatial context for contamination, oil exploration and transportation data.

In the case of the polygonal infrastructure for the Shishmaref test area, the SIRIUS dataset achieves a representation of 94 % of all the true values. Distorting the overall accuracy are the false positives, approximately one-fourth of which belong to the area of Shishmaref's airport runway that is no longer in use. It is important to note that OSM encourages users to regularly update features. If a user finds that a feature no longer physically exists, they should delete it or tag it as “nonexistent” . If a feature still physically exists but is no longer in use, users are encouraged to tag it as “disused” . In this specific context, and considering the potentials of the contamination, it could be seen as an asset to have former land usage and industrial legacies represented in the SIRIUS dataset. An interesting approach might thus be to specifically filter for the OSM tags nonexistent and disused – in the regularly updated and historical OSM database – to highlight potential contamination sites.

The same applies to the human-impacted areas, such as playgrounds and industrial land use. While these features are important infrastructures critical to Arctic communities, they largely cannot be mapped on the basis of orthophotos alone. Accordingly, the polygonal infrastructure lacks this level of detail when derived from SACHI. As discussed in Sect. , the coarse spatial resolution of the Sentinel sensors poses a challenge in densely populated areas. In such regions, buildings and human-impacted areas become difficult to separate from adjacent roads. This challenge contributes to the high number of false positives, where roads are misclassified as buildings and areas of human activities are overestimated. However, this issue could be addressed using imagery with a higher spatial resolution.

It is important to note that the high mapping accuracies of 78 % (92 %) for the linear infrastructure and 94 % for the polygonal infrastructure in our test area of Shishmaref can likely be expected for most of the coastal regions (until 100 km inland). For inland areas (beyond the extent of SACHI), the infrastructure data rely solely on OSM, which may show the abovementioned limitations. Once again, integrating further official data sources could improve their quality.

We were able to successfully enhance the DEC Contaminated Sites dataset with complementary information regarding the substances, the affected medium and the duration of the cleanup measurements. However, the text-mining approach, using regular expressions to compare site reports to the DEC glossary to retrieve the contaminant and the affected medium, encountered limitations where data were entered heterogeneously into the database (see Sect. ). For instance, only 39 % of the site reports included information about the contaminated medium in the designated section “contaminant information”. In addition, in some cases, comparing the medium's keywords (soil, groundwater, etc.) to this section led to false positives as these terms are frequently used to describe the hazard level of the substances. The first entry (Hazard ID 26994) of our validation sample (refer to Table ) is one of these false positives. The site report actually lists “soil” as a contaminated medium, but the level description for the substances “benzo[a]anthracene and benzo[a]pyrene” is “between method 2 migration to groundwater and human health/ingestion/inhalation”. Consequently, our approach also lists groundwater as a contaminated medium, which is not accurate. If we were to compare the full report to these keywords, this would result in even more incorrect classifications, as these terms are also employed to describe a suspicion of contamination. Furthermore, using regular expressions for the retrieval of the polluting substances does not differentiate between the presence and absence of a contaminant, e.g., “PCB was found” vs. “PCB was not found”. Although we did not encounter statements of absent contaminants in the reports of our sample, we cannot rule out the possibility of false positives of this kind.

These shortfalls could be addressed by implementing advanced text classification approaches from natural language processing and text mining. This could provide a more comprehensive understanding of toxic substances, including those not mentioned in the DEC glossary. Furthermore, these methods would extract and classify information on the contaminated medium from the entire report, rather than solely relying on the sampled section labeled “contaminant information”. Another viable alternative would be the integration of large language models (LLMs). We tested our particular false positive case (Hazard ID 26994) with the LLM chatbot ChatGPT Version 3.5 by copying the full report into the prompt and requesting “reading this text, tell me what medium (soil, groundwater, river, lake, etc.) was contaminated:”, and it correctly classified the affected medium:

Based on the provided text, the medium that was contaminated is `Soil'. The text mentions that soil samples collected during site assessment activities showed elevated concentrations of contaminants, specifically `benzo(a)pyrene' and `benzo(a)anthracene', which exceeded certain cleanup levels. Therefore, the contamination occurred in the soil medium.

In this way, inconsistencies in data entries and false classifications could be easily addressed.

All the resulting datasets have been organized as individual layers within a single GeoPackage, which is available for download from our Zenodo repository (see the “Code and data availability” section). The GeoPackage does not have to be extracted (e.g., like a .zip archive), nor does it rely on the handling of multi-file data formats such as ESRI Shapefiles. The user can seamlessly integrate it by either opening it in a GIS application or importing it into a spatially enabled database like PostgreSQL/PostGIS. In this way, each layer can be analyzed independently or in conjunction with the others, facilitating easy querying of critical infrastructure and human-impacted areas and their interrelation with environmental parameters.

To achieve a more comprehensive understanding of the socioeconomic implications of permafrost degradation, we advocate incorporating additional environmental data, such as soil and waterbody databases, which are important for assessing the contamination severity and the significance of waterbodies as water resources. Additionally, incorporating demographic factors like age distribution, education, employment and income numbers can provide valuable insights into the impacts of permafrost degradation on the population's wellbeing.