Suitability analysis of human activities over Antarctic ice shelves: an integrated assessment of natural conditions based on machine learning algorithms

ABSTRACT

Human activities increase significantly over Antarctic ice shelves. However, they are constantly faced with danger posed by the harsh environment. For decision-making, it is a prerequisite to have the macro-scale suitability information about human activity site selection. Here, we define a new index, the human activity suitability (HAS) index, to quantitatively analyze the locational suitability of human activity sites over Antarctic ice shelves. Two multi-criteria decision analysis methods (AHP + Entropy, TOPSIS) and three machine learning methods (support vector machine, random forest and logistic regression) are tested to develop HAS maps. Nine conditioning factors about ice surface features, ice shelf stability, meteorology and topography are generated as input parameters. The accuracy of the proposed models is evaluated using metrics such as the area under curve (AUC-ROC), root mean square error, overall accuracy and kappa index. The results indicate that the Random Forest performs best. The HAS map exhibits great heterogeneity driven by the synergistic influence of multiple factors. Areas in low HAS classes are concentrated at Crosson, Brunt, Thwaites and the edge of some ice shelves, implying the complex environment in these regions. The findings can provide a new insight for forecasting the potential human footprint and support sustainability research in Antarctica.

KEYWORDS:

• Human activity suitability
• Antarctic ice shelf
• machine learning
• remote sensing

1. Introduction

Widely considered as the last unexplored continent, Antarctica contains plentiful resources and scientific mysteries. Since the Antarctic Treaty, Antarctica has been targeted as a high-profile strategic region for scientific research (Hanessian 1960). Antarctic ice shelves are the large floating ice masses attached to the land-based ice. They not only slow the flow of grounded ice, but also provide buttress to the ice sheet and affect the sea levels (Liang et al. 2022). As climate change is exacerbated, the Antarctic ice shelves experience drastic and profound transitions in the twenty-first century (Fürst et al. 2016). Such a complex system has attracted increasing attention from researchers worldwide.

As ice shelves provide a natural test site for researchers, many focused and long-term field observations in the Antarctic ice shelves are conducted. Also, the flat and relatively crevasse-free surfaces of Antarctic ice shelves can provide convenient highways for surface access to the cope part of Antarctica (Müller and Schøyen 2021). Consequently, human activities in the Antarctic ice shelves are currently in a phase of rapid expansion. There are a certain number of research stations that support various experiment activities such as the Halley station built on the Brunt ice shelf and McMurdo Station on the Ross ice shelf. Moreover, many international organizations are conducting large-scale ice shelf observation experiments (Müller and Schøyen 2021). For example, the iSTAR – NERC ice sheet stability program features a two-year plan drawn up to gather data and find out how Pine Island glaciers are changing (Joughin et al. 2021). With the development of technology and infrastructure, the frequency and intensity of human activities around ice shelves are expected to further increase in the future.

However, conducting in-situ explorations over Antarctic ice shelves is extremely dangerous due to the harsh physical environments. The extreme conditions can cause serious damage to station equipment, injuries to researchers and even death. They are attributed mainly to the influence of ice surface features, ice shelf stability, extreme meteorology and topography such as severely low temperatures (Schutz, Zak, and Holmes 2014; Smith, Kinnafick, and Saunders 2017). Not least, the crevasses at the ice shelf surface, which are usually covered by snow bridges, can impede the movement of personnel and snowmobiles greatly. Victims of crevasse falls without a rope may lose their lives due to the extreme cold and wind or may sink further into the ice gap. As reported, a helicopter pilot from the Australian Antarctic Division fell into a crevasse and lost his life during a fuel delivery on the West ice shelf in 2016 (Giesbrecht and Brock 2022). It is also recorded that numerous stations are destroyed and abandoned because of the ice shelf motion, snow cover and so on, such as Belgrano I station in the Ronne-Filchner ice shelf and Little America base in the Ross ice shelf (Byrd 2015). These catastrophes raise an important question, that is, how can we mitigate the natural hazards and minimize the unnecessary losses caused by the unreasonable site selection for human activities?

Conducting comprehensive environmental assessment and investigations is the primary task for planners to perform before any activities on the Antarctica. The environmental investigations usually involve two parts, general remote-sensing surveys and detailed field investigations. Detailed field investigations are usually carried out through unmanned aerial vehicles, ground penetrating radar and discussions among decision-makers (Cui et al. 2019), which are accurate but also dangerous, time-consuming, and difficult to get large-scale information, especially working in the harsh environment on the Antarctic. General surveys based on satellite remote sensing data can help to solve these issues because the satellite remote sensing technology enables the efficient collection of spatiotemporal data over vast spatial areas at a low cost (Liu et al. 2023a; 2023b; Yang et al. 2022). The general survey results can serve as an early warning in the planning phase and reduce the safety risk posed by unmanned aerial vehicle deployment and dangerous field investigations. Currently, remote sensing data has been extensively used in the practice of selecting airfields (Cui et al. 2019), exploring meteorite collection sites (Tollenaar et al. 2022) and mapping human footprints (Brooks et al. 2019) in Antarctica. Therefore, it is feasible to collect the macro-scale suitability information in human activity site selection through general surveys based on remote sensing data on the Antarctica.

Despite the foundational role that macro-decision-making information played for human activity site selection, only a few studies have focused on the locational suitability analysis of human activities. For example, Pang, Liu, and Zhao (2014) and Yavaşoğlu, Karaman, and Özsoy (2019) adopted the analytical hierarchy process method to identify suitable research station sites in inland Antarctica and the Antarctic Peninsula, respectively, by considering environmental constraints, logistics and human factors. Nevertheless, there is still a lack of suitability analysis conducted on the activities for human beings over the whole Antarctic ice shelf. Although these previous studies focus on some meteorological and topographic factors, some vital natural conditions have not yet been considered such as ice surface crevasses. On the other hand, the previous study focuses mainly on the development of assessment models through traditional multi-criteria decision analysis (MCDA) methods. Recently, various machine learning methods, such as regression analysis and deep learning, have shown successful applications in the field of the cryosphere and other geographical research (Chen et al. 2023; Wu et al. 2023; Xiao et al. 2021). Mehravar et al. (2023) has also concluded that the hybridized models, combining machine learning methods with metaheuristic algorithms, can optimize the model performance in flood susceptibility mapping tasks. Thereby, the advantages of machine learning in solving complex problems give us access to new viewpoints. In spite of this, it remains unclear whether the machine learning methods are applicable to the suitability analysis of Antarctic human activities remain unclear.

The objective of this study is to define a novel index, namely the Human Activity Suitability (HAS) index, and develop a comprehensive framework to assess and map its distribution over Antarctic ice shelves. The HAS index allows for analyzing the suitability of human activities quantitatively over the Antarctic ice shelf region, which can better serve the scientific expedition planning and research station site selection. To achieve this, we firstly investigate the reliable conditioning factors that are associated with the HAS. Then, the performance of five well-known models based on MCDA techniques and machine learning techniques is compared for the spatial mapping of HAS. The MCDA methods include the Analytic Hierarchy Process and Entropy Weight (AHP + Entropy) and Technique for Order Preference by Similarity to Ideal Solution (TOPSIS). The three machine learning techniques applied are random forest (RF), support vector machine (SVM) and logistic regression (LR). Finally, the HAS index over Antarctic ice shelves is mapped and its spatial distribution is analyzed in depth.

2. Study area and datasets

2.1. Study area

The Antarctic ice sheet is the largest mass of ice on Earth. It is generally divided into the Eastern Antarctic Ice Sheet, the Western Antarctic Ice Sheet and the Antarctic Peninsula. Its environment is extremely dry and cold, with an annual precipitation of about 151 mm and an average temperature of about −62 °C (Drewry, Jordan, and Jankowski 1982). Surrounding the Antarctic ice sheet, there are approximately 300 ice shelves. They are the floating extensions of the land ice and flow slowly towards the ocean, which is essential for the dynamics of the Antarctic ice sheet. The thickest ice shelves in Antarctica can be over 700 meters thick such as the Ross ice shelf. The elevation and topography of Antarctic ice shelves can vary widely. Some portions closer to the interior of the continent may have elevations well over 1,000 meters above sea level. Ice shelves along the coast and on the Antarctic Peninsula can have much lower elevations, often less than 100 meters above sea level. In this study, we select 53 Antarctic ice shelves as the study objects covering an area greater than 1,000 km², counting for 70% of the Antarctic coastline (Figure 1). The total area of our study is 2.63 × 10⁶ km². So far, Antarctica ice shelves have been deployed dozens of stations for glaciologists, meteorologists and climatologists. Figure 1 exhibits the locations of all the research stations and field sites used in this study. The Ross and Ronne-Filchner ice shelves host the most human activity sites, accounting for 40% of the sample dataset. The ice shelves in Queen Maud Land and West Antarctica are also hot spots for scientific explorations.

Figure 1. The location of ice shelves, stations, camps and field sites used in this study.

2.2. Datasets

2.2.1. Indicators and data sources

The suitability analysis of human activity locations is assessed based on the natural factors affecting human activities in four categories, including ice shelf surface features, ice shelf stability, meteorology and topography. Distance from crevasses, distance from the blue ice area, strain rate, wind speed, snow accumulation, slope, maximum buttressing, air temperature and ice velocity indicators are selected (Table 1 and Figure 2). The distance from the blue ice area and maximum buttressing indicators are positive indicators, which means the larger the indicator value, the more beneficial to human activities. Other indicators are the opposite. The selection of these indicators is based on several research studies (Pang, Liu, and Zhao 2014; Yavaşoğlu, Karaman, and Özsoy 2019; Shimada et al. 2020).

Figure 2. The indicators applied in the HAS map: (a) crevasses, (b) blue ice, (c) slope, (d) wind, (e) strain rate, (f) max buttressing, (g) snow accumulation, (h) ice velocity and (i) air temperature.

Table 1. Summary of the indicators used in our study.

Indicators	Source	Period	Spatial resolution	Correlation with HAS
Distance from Crevasses	Sentinel-1, MODIS	2021, 2009	500 m	Negative
Distance from blue ice area	MODIS	2014	500 m	Positive
Strain rate	MEaSUREs	2007–2018	450 m	Negative
Wind speed	ERA-5	2017–2021	0.10°	Negative
Snow accumulation	BAS	2001	1 km	Negative
Slope	ICESat-2	2020	500 m	Negative
Maximum buttressing	MODIS	2003	125 km	Positive
Air temperature	ERA-5	2017–2021	0.10°	Negative
Ice velocity	MEaSUREs	2007–2018	450 m	Negative

Crevasses are deep, narrow cracks or fissures developing in the surface of Antarctic ice shelves, which can be several meters wide and extend down to depths of hundreds of meters or more (Colgan et al. 2016). Numerous crevasses are concealed by snow bridges and posed a huge danger for workers. If they unknowingly cross over a hidden crevasse, they can fall through and suffer serious injuries or even death (Qiao et al. 2023). We first obtained the crevasses distribution based on previous work (Zhao et al. 2022), which used Sentinel-1 imagery and U-Net semantic segmentation algorithm. For the regions that are not covered by Sentinel-1 data because of its near-polar orbit, we used crevasse distribution products from Lai et al. (2020) to supplement, which exacted crevasses used the Moderate Resolution Imaging Spectroradiometer (MODIS)-based Mosaic of Antarctica (MOA) datasets at a 125 m resolution and U-Net semantic segmentation algorithm.

The blue ice areas in Antarctica are snow-free zones with blue, smooth and flat surfaces (Bintanja 1999). They provide essential sites for transportation as the hard surface can support the heavy-wheeled aircraft landing (Jawak et al. 2023). Additionally, melting blue ice can be a potential source of water supply for some research stations. Furthermore, they are closely related to research location selection, such as for ice core drilling experiments or meteorite seeking. In this study, the MOA datasets during 2013–2014 was employed to map the blue ice, which is compiled by the National Snow and Ice Data Center (Scambos et al. 2007). A threshold-based method provided by Hui et al. (2014) was applied to identify blue ice areas where snow grain sizes were greater than 400 μm.

Strain is a measure of how much stretches, compression, and deformation forces received by the ice shelf in all directions during flow (Alley et al. 2018). Strain rate is defined as the rate at the ice deforms occur. Strain rate is included in our analysis due to its known relationship with the emergence of crevasses in previous research (Colgan et al. 2016). In some cases, there are no visible fractures but they may still be present with a high strain rate. The first principal strain rate ϵ was calculated from the MEaSUREs Antarctica ice velocity data product (Rignot, Mouginot, and Scheuchl 2017) by Eqs. (1)(1) $ϵ=\frac{1}{2}({ϵ}_x+{ϵ}_y)+\sqrt{\frac{1}{4}({ϵ}_x-{ϵ}_y{)}^2+{ϵ}_{xy}^2}$(1) –(2)(2) $\left[\begin{array}{cc}{ϵ}_x& {ϵ}_{xy}\\ {ϵ}_{xy}& {ϵ}_y\end{array}\right]=\left[\begin{array}{cc}{\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }x}}& {\displaystyle \frac{1}{2}({\displaystyle \frac{\mathrm{\partial }v}{\mathrm{\partial }x}+{\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }y})}}}\\ {\displaystyle \frac{1}{2}({\displaystyle \frac{\mathrm{\partial }v}{\mathrm{\partial }x}+{\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }y})}}}& {\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }y}}\end{array}\right]$(2) .

(1) $ϵ=\frac{1}{2}({ϵ}_x+{ϵ}_y)+\sqrt{\frac{1}{4}({ϵ}_x-{ϵ}_y{)}^2+{ϵ}_{xy}^2}$(1) (2) $\left[\begin{array}{cc}{ϵ}_x& {ϵ}_{xy}\\ {ϵ}_{xy}& {ϵ}_y\end{array}\right]=\left[\begin{array}{cc}{\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }x}}& {\displaystyle \frac{1}{2}({\displaystyle \frac{\mathrm{\partial }v}{\mathrm{\partial }x}+{\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }y})}}}\\ {\displaystyle \frac{1}{2}({\displaystyle \frac{\mathrm{\partial }v}{\mathrm{\partial }x}+{\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }y})}}}& {\displaystyle \frac{\mathrm{\partial }u}{\mathrm{\partial }y}}\end{array}\right]$(2) where u and v are the velocity components in the x and y directions, respectively.

Heavy wind is also a dangerous source of accidents. When the wind speed exceeds 28 knots, there is a possibility for loose snow to drift, thus reducing the visibility. In some scenarios, it can decrease to zero. It is exactly the case when the McDonnell Douglas DC10 aircraft crashed into Mount Erebus in Antarctica, resulting in the loss of 257 lives (Auburn 1983). The wind speed of monthly average data from 2017 to 2021 was extracted using ERA-5 monthly average reanalysis dataset (Hersbach et al., 2020).

The low level of snow accumulation can prevent humans and buildings from being covered by snow. Such as the Asuka station, it was constructed on the snow surface at the base of the Sør Rondane Mountains and only remain in operation for six years due to the high accumulation rates of snow and powerful katabatic winds (AoKI and Yamanouchi 1992). The snow accumulation map dataset used is provided by British Antarctic Survey (BAS), which was derived from field measurements and Advanced Microwave Scanning Radiometer for EOS and Advanced Very High Resolution Radiometer satellite observations (Arthern, Winebrenner, and Vaughan 2006).

The slope of a region is an indicator of how steep the terrain is. A steeper slope implies a more challenging area that requires more advanced physical strength and outdoor abilities of the participants (Swain 2018). We utilized a digital elevation model dataset created by Shen et al. (2022) using ICESat-2 data to calculate the ice snow surface slope indicator.

The maximum buttressing of an ice shelf is quantified by Fürst et al. (2016) by comparing the normal force it exerts on upstream ice and vertical hydrostatic pressure. It is the reflection of the buttressing potential effect of the ice shelves. This dataset has been widely used in the stability study of Antarctica ice shelves (Lai et al. 2020; Miles et al. 2023). Lower maximum buttressing values suggest that the ice shelf provides insufficient support and can be considered as weakly buttressed (Gudmundsson 2013). It is therefore hazardous to conduct activities in these regions due to an increased risk of ice shelf collapse.

Ice velocity is an essential reflection of glacier dynamics and instability (Rignot, Mouginot, and Scheuchl 2011). The coastal regions are much more active than inland regions in Antarctica. For this reason, it incurs more time and financial costs to design and test the construction of stations in these active regions (Siegfried et al. 2016). Human activities are more likely to be carried out in the regions with a lower velocity, especially for long-period activities because fast ice flows can cause damage to some perennially human-inhabited facilities.

In some scenarios, the air temperature along the Antarctic coast can fall below −40°C during winter (Drewry, Jordan, and Jankowski 1982). Extreme low temperatures can increase the cost of constructing a research station and negatively impact the living conditions of staff, thus inconveniencing their normal life and work. For our study, we selected the 2 m temperature data from the monthly averaged data of ERA5-Land (Hersbach et al., 2020).

2.2.2. Auxiliary dataset

In order to model the HAS index, it is necessary to obtain a sample dataset that represents the locations suitable for human activities. It is assumed that suitable locations are more likely in the zones with the same characteristics. So, the sample dataset can help us understand the relationship between the HAS index and conditioning factors. We believe that successful and suitable exploration locations in Antarctic ice shelves could operate for a long time. In this case, 36 active stations, camps, airports and refuges from governmental databases were included in the suitable sample dataset (Figure 1). To expand the sample dataset, we then collect 244 history records of large-scale scientific investigations during 2016–2022 period from the published literatures including hot water drilling, sediment coring, ApRES radar and GPS measurements (Figure 1 and Table S1). The research stations are established by numerous national administrations. Some of the largest research stations built in Antarctica can accommodate over 1,000 people during summer (Davis 2017). Field camps are the permanent structures or tents used during the annual Antarctic summer to support various activities such as airport, topographic surveys and logistics. All these locations are selected to support large-scale human activities. Therefore, they are considered representative in the process of HAS index modelling.

3. Methodology for HAS index

A HAS assessment framework is developed to qualify the HAS index over Antarctic ice shelves, which includes four steps (Figure 3). Firstly, we collect nine indicators to establish a reliable HAS indicator system. Secondly, we construct a sample dataset for model training and validating. Subsequently, two types of models (MCDA and machine learning) are compared for modelling the HAS. The MCDA models evaluate the nine indicators according to a predefined set of decision criteria. The machine learning models use sample dataset as the target variable to be modelled and the conditioning factors as the explanatory variables for exploring the associations between them. Each developed model is validated using our sample dataset to determine the optimal model. Finally, the HAS map is generated by this optimal model for an in-depth spatial pattern analysis. More details are presented as follows.

Figure 3. The methodological flowchart.

3.1. Valid indicator selection

The selection of appropriate conditioning factors is crucial for HAS mapping. To eliminate irrelevant and redundant factors, we apply a multicollinearity diagnosis test to the nine selected indicators. Multicollinearity is a statistical phenomenon in which a high-level relationship exists between variables in a multiple regression model. Factors exhibiting high multicollinearity should be removed to minimize the bias of the model and enhance performance. The main metrics for the multicollinearity diagnosis test are Variance Inflation Factors (VIF) and Tolerance (Shabanpour et al. 2022). Values of VIF greater than 10 and Tolerance less than 0.1 indicate the presence of multicollinearity issues within the indicators. To mitigate estimating errors, we employed a 10-fold cross-validation technique to compute VIF and Tolerance.

3.2. MCDA and machine learning methods

One of our objectives is to compare the performance of multiple MCDA and machine learning algorithms for aggregating HAS indicators. The detailed descriptions and steps of each algorithm are referred to supplementary information (Text S1). MCDA methods can assist decision-makers in analyzing probable decision alternatives based on trade-offs between multiple criteria. AHP + Entropy and TOPSIS methods are both popular comprehensive methods and have been used in previous studies on Antarctic decision support (Yavaşoğlu, Karaman, and Özsoy 2019). AHP determines the subjective weights of evaluation indicators based on a two-by-two comparison matrix. Entropy calculates the objective weight based on the notion of information entropy. AHP + Entropy is an approach to combine the subjective and objective weighing processes (Macharis et al. 2004). TOPSIS designates solutions based on a distance calculation that assumes the shortest distances to positive ideal solutions as the best solutions (Hwang and Yoon 1981).

Machine learning algorithms can learn from previous data and adapt iteratively without being explicitly programmed to describe data or predict outcomes. We utilize SVM, RF and LR methods to establish the models and compute the HAS index. SVM creates an ideal linear hyperplane to classify training data and distinguish two groups based on the principle of structural error minimization (Mehravar et al. 2023). RF is an important ensemble model that generates multiple decision tree classifiers to execute a classification (Breiman 2001; Xiao et al. 2018). LR is a statistical model that uses a logistic function to describe the relationship between one binary dependent variable and multiple independent variables (Shabanpour et al. 2022). Generally, the models are successfully established using the sample dataset and the data of influencing indicators for each pixel. In this study, a total of 12,682 pixels are considered as suitable samples and denoted as ‘1’. As Section 2.2.2 mentioned, 280 pieces of records about sites, camps and in-situ investigation records are included to construct the suitable sample dataset. To mitigate the impact of point scale contingency on the results, the pixels within 2 km from the collected record sites are considered as the suitable samples. Unsuitable samples are randomly selected in the regions outside the 10 km buffer zone of the collected points and denoted as ‘0’, with the same number of suitable samples. The random sampling procedure is conducted to ensure objectivity and avoid the negative impact of subjective sampling. 70% of the sample points are used in the training phase of machine learning models, while the remaining 30% of the sample points are testing dataset, that has not been employed previously, are executed to validate the reliability of the results.

3.3. Method performance metrics

To evaluate the performance of all the methods, this study utilizes the Receiver Operating Characteristic (ROC) curve to be an overall performance statistic index for models. It is a graph depicting the false positive rate on the x-axis and the true positive rate on the y-axis given different thresholds (Shabanpour et al. 2022). The area value under the ROC curve (AUC) is estimated to be from 0 to 1. The higher the AUC value, the better performance of the developed models.

Additionally, the performance of the machine learning models is also evaluated by calculating the statistical criteria. The training dataset’s statistical results show the degree of fitting, whereas the testing dataset’s results can demonstrate the accuracy of the proposed model's predictions (Lv et al. 2022). Overall accuracy, kappa index and Root Mean Squared Error (RMSE) are the statistical criteria adopted, as specified in Eqs. (3)(3) $Overallaccuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$(3) –(5)(5) $RMSE=\sqrt{{\displaystyle \frac{1}{n}\sum _{i=1}^{i=n}{\left({\mathrm{X}}_i-{\hat{X}}_i\right)}^2}}$(5) .(3) $Overallaccuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$(3) (4) $Kappaindex={\displaystyle \frac{P_{obs}-P_{\exp }}{1-P_{\exp }}}$(4) (5) $RMSE=\sqrt{{\displaystyle \frac{1}{n}\sum _{i=1}^{i=n}{\left({\mathrm{X}}_i-{\hat{X}}_i\right)}^2}}$(5) where TP (True Positive) and TN (True Negative) represent the correctly classified pixels’ number. FP (False Positive) and FN (False Negative) represent the falsely classified pixels’ number.$P_{obs}$ is the observed proportion of agreement; $P_{\exp }$ is the proportion of agreement expected; X_i and ${\hat{X}}_i$ are the observed and predicted values, respectively. n means the total amount of sample dataset.

4. Results

4.1. Predictive ability of conditioning factors

The HAS index can be significantly affected by multiple factors. Before establishing the model, it is necessary to analyze the predictive ability of the selected factors and remove the invalid factors for the models. We first utilize multicollinearity diagnosis (Section 3.1) to demonstrate the validity of the selected indicators. The results obtained from the multicollinearity diagnosis reveal that the VIF values are between 1.13 and 2.61 and the values of Tolerance are from 0.38 to 0.88, indicating the absence of multicollinearity issues among the selected indicators (Table 2). Then, we use the built-in feature importance of the RF algorithm to rank the importance of the selected indicators (Xiao et al. 2022) (Figure 4). According to the results, the distance from crevasses (0.24) is of the most importance, followed by the ice velocity (0.22) and the maximum buttress (0.16). By contrast, slope (0.01) and wind speed (0.03) have less significant value in estimating the HAS index.

Figure 4. Factor importance estimates obtained by the developed RF model.

Table 2. Tolerance and VIF values in multicollinearity diagnosis test for the selected indicators.

Indicator layer	Tolerance	VIF
Distance from Crevasses	0.79	1.26
Distance from blue ice area	0.65	1.54
Strain rate	0.88	1.14
Wind	0.72	1.38
Snow accumulation	0.49	2.05
Slope	0.88	1.13
Maximum buttressing	0.69	1.45
Air temperature	0.38	2.61
Ice velocity	0.58	1.72

4.2. Performance evaluation

Five HAS results over Antarctic ice shelves are calculated based on the AHP + Entropy, TOPSIS, SVM, RF and LR models. Afterwards, the AUC statistics are generated to investigate the performance of each established model based on the training and testing dataset, respectively (Figure 5). The results of the ROC curve method reveal that all five models demonstrated adequate performance (AUC > 80%). Among the models considered, the RF model demonstrates superior performance, indicated by the AUC of 0.9834 in the training process and 0.9773 in the testing process. However, high AUC values sometimes cannot be served as the only performance evaluation metric of high accuracy models. For our machine learning results, the correlation coefficients among them are from 0.74 to 0.82, implying a high level of spatial agreement. To further evaluate the accuracy and effectiveness of the machine learning models during training and testing, we apply three statistical metrics as additional criteria. The results indicate that in the training phase, the RF model performs best in terms of higher overall accuracy (0.9780), kappa index (0.9560), and lower RMSE (0.1520) as shown in Table 3. Also, the testing phase shows a substantial agreement among RF, SVM and LR models. Judging from the above analysis of AUC and three statistical metrics, the RF model is clearly advantageous.

Figure 5. The five developed models’ ROC curves with training dataset (a) and testing dataset (b).

Table 3. Three statistical metrics of the machine learning model.

Statistical metrics	Training dataset			Testing dataset
	SVM	RF	LR	SVM	RF	LR
Overall accuracy	0.9516	0.9780	0.9438	0.9458	0.9639	0.9363
Kappa index	0.9032	0.9560	0.8877	0.8916	0.9479	0.8726
RMSE	0.1627	0.1520	0.1806	0.1739	0.1612	0.1889

4.3. Results of HAS index map

As a result, the suitability index for human activity locations is calculated based on the natural conditions and the developed RF model, which is from 0 to 1. They are reclassified into five different HAS classes based on the natural break method. The natural breaks classification method is available in ArcGIS software and can find the best ranges by minimizing variation as closely as possible. To specify the relation between HAS index with the physical environment, we calculate the average values of the selected indicators at different HAS classes (Table 4). Different HAS index classes represent different degrees to which the location of human activities is recommended. Classes below moderate represent a greater threat posed by natural conditions to the security of human activities, deserving more attention from the planners. Figure 6 shows the final distribution of HAS index classes. The conducted sample dataset (see Section 2.2.2) is used to evaluate the final HAS map. It demonstrates that 88.24% of the suitable training pixels are located in the very high and high HAS areas, and 74.07% of the suitable validating pixels are located in these areas as well. These results illustrate the credibility of the HAS map.

Figure 6. The final HAS map by RF model. We divide Antarctica ice shelves into seven regions, which are labeled and delimited by line segments in dark blue. For each region, we compute area percentages of five HAS classes (pie charts). Yellow represents the testing dataset, green represents the training dataset. Triangle represents the research station sites, and circle represents the human activity records.

Table 4. Mean values of the selected indicators for the five HAS classes.

Parameter	HAS classes
	Very high	High	Moderate	Low	Very low
HAS index	1 - 0.83	0.83 - 0.63	0.63 - 0.43	0.43 - 0.22	0.22 - 0
Distance to crevasses (Km)	16361.09	15069.99	12724.17	10104.06	3446.96
Distance to blue ice (Km)	135674.68	146677.67	198421.92	206022.09	177121.63
Slope (°)	0.09	0.10	0.16	0.32	0.37
Wind speed (m/s)	3.23	3.18	3.21	3.99	3.76
Strain rate (1/yr)	0.0030	0.0052	0.0065	0.0062	0.0066
Buttressing (dimensionless)	0.51	0.25	0.14	0.09	0.08
Snow accumulation (kg/m2/yr)	218.54	203.00	209.60	238.10	248.42
Ice velocity (m/yr)	129.45	261.35	523.23	686.87	857.57
Air temperature (°C)	−18.44	−20.34	−22.15	−22.45	−19.57

For the whole of Antarctica ice shelves, we find that 20.76%, 16.34% and 9.52% of the study area in Antarctic ice shelves fall within the moderate, low and very low HAS classes, respectively. Figure 7 shows the proportion of area for each HAS level within different ice shelves. A total of 16 ice shelves have low and very low HAS area, accounting for more than 30% of their own areas. There are 30 ice shelves with high and very high HAS areas, accounting for more than 50% of the total. At a regional level, ∼70% of the West Antarctica ice shelf areas belongs to the high or very high HAS classes, notably in Abbot and Getz ice shelves. Ice shelves around the Queen Maud Land coast display approximately 60% high HAS regions yet their values are more homogeneous. Conversely, we find more low and moderate HAS area proportions (∼50%) along the Wilkes Land coast.

Figure 7. Proportion of area for each HAS class within different ice shelves.

There is also a significant discrepancy of HAS among different ice shelves. As for the Antarctic Peninsula, there are ice shelves with the highest and lowest area proportions of high and very high HAS classes. Located on the western side, Wilkins, Bach and George VI ice shelves are the most suitable areas that may be used as ice shelf landing sites. Conversely, Larsen B – E ice shelves on the eastern side have the largest proportion of areas with low and very low HAS classes, which poses a risk of collapse as suggested in prior studies(Mitcham, Gudmundsson, and Bamber 2022). In the West Antarctica, the area proportions of high HAS class in Thwaites and Crosson ice shelves are even less than 10% and Pine Island is less than 50%. These areas are where long-term field activities are not recommended. Because it is very likely that the collapse here will directly cause the departure of life and facilities. We suggest establishing research stations in nearby ice shelves such as Abbot and Dotson ice shelves or inland areas, so as to better support short-term scientific research activities over the very low HAS ice shelves. It should be noted that our guidance discussion is based on natural conditions. In practical applications, it is still necessary to consider the scientific interest. For example, the Halley VI station is situated in the low HAS class region on Brunt ice shelf. Although it is still operating normally, our results can provide a warning for such equipment maintenance personnel.

4.4. Local HAS index analysis for different ice shelves

Large ice shelves occupy most of the total area classified as very low and low HAS classes, which exhibit significant heterogeneity in terms of the HAS index. The Ross, Ronne-Filchner and Amery ice shelves are the three largest ice shelves over Antarctica, collectively representing around two-thirds of the total area of Antarctic ice shelves. These ice shelves are of great interest to researchers and serve as hosts to numerous research stations and exploration activities. Combined with the conditioning factors, we analyze their HAS index distribution patterns in detail. Using the testing dataset, we further verify the rationality of our HAS index results.

4.4.1. Ross ice shelf

In the Ross ice shelf, areas with low HAS index scores are consistently located at the front edge of the ice shelf and near the Transantarctic Mountains (Figure 8). The buttress function of the front part of the Ross ice shelf is relatively weak and the ice flow is fast, resulting in the distribution of many large transverse crevasses. As mentioned in Section 2.2.1, such regions are unsuitable for human activities. Despite the strong buttress function in the areas near the Transantarctic Mountains, dense crevasses also lead to the low HAS values. Furthermore, it is noteworthy that many outlet glaciers over Transantarctic Mountains exhibit large blue ice area distribution, so some fixed landing sites have been established. However, these areas still feature high ice velocity and numerous crevasses, requiring further detailed field exploration for human activity security. In contrast, the central-eastern and southern regions of the Ross Ice Shelf, as well as the area near Ross Island have high HAS index scores characterized by the low strain rate, moderated ice velocity and flat ice surface terrain. This pattern is consistent with our sample dataset. Near the Ross Island, there are several research stations and airports distributed, such as the McMurdo station and Phoenix airfield. South Pole Traverse is a 1,601 km-long route that stretches from McMurdo Station across the Ross ice shelf and leads to the South Pole. This remains to the present day as crucial locations for numerous scientific explorations such as hot water drilling.

Figure 8. Crevasse and blue ice areas (a), HAS index (b), ice velocity (c), maximum buttressing (d) and strain rate (e) of the Ross ice shelf.

4.4.2. Ronne-Filchner ice shelf

The Ronne ice shelf is located immediately west of the Filchner ice shelf, from which it is divided in part by the Berkner Island. Similar to the Ross ice shelf, there are large transverse rifts distributed at the front edge of the Ronne-Filchner ice shelf, creating an adverse environment to human activities (Figure 9). Besides, the Filchner ice shelf shows higher ice velocity and strain rate, resulting in more crevasses and longitudinal structures. Consequently, the Filchner ice shelf shows a lower HAS index score than the Ronne ice shelf. Afterwards, the presence of the Berkner Island, Henry ice rise and Korff ice rise block the movement of the ice shelf. On their southern side, the condition is more favorable for scientific activities. Accordingly, 53 on-site records in two scientific activities for base melt are gathered there. Across the Ronne-Filchner ice shelf, there are 40 collected ApRES sites. However, there are 15 moderate-HAS and 3 low-HAS points, indicating a higher level of uncertainty compared to other ice shelves.

Figure 9. Crevasse and blue ice areas (a), HAS index (b), ice velocity (c), maximum buttressing (d) and strain rate (e) of the Ronne-Filchner ice shelf.

4.4.3. Amery ice shelf

In the front of Amery, the two large crevasses regions are very low HAS regions. Such areas are characterized by more hidden crevasses, fast ice flow and little buttress function, meaning that facilities can quickly flow into the ocean with the collapse of the ice shelf (Figure 10). The upstream region is at a lower HAS level, which is converged by outlet glaciers. This area is characterized by high strain rates and dense longitudinal structures. But it is undeniable that this region contains blue ice areas, melting lakes, and some ice-free areas, making it a region of great scientific interest. Judging from the results, we suggest paying more attention to equipment maintenance and safety protection here. In contrast, we recommend that the location for long-term activities be chosen in the middle of the ice shelf area and near the Single Island on the west side. This is because of the flatter terrain in these regions and the proximity to the blue ice areas and water sources.

Figure 10. Crevasse and blue ice areas (a), HAS index (b), ice velocity (c), maximum buttressing (d) and strain rate (e) of the Amery ice shelf.

5. Discussion

In our HAS assessment framework, the optimal model is determined by comparing the performance of several MCDA methods and machine learning methods. To our knowledge, it is the first time to map the HAS over Antarctic ice shelves. In this section, the first subsection compares the applicability of MCDA and machine learning models for HAS index mapping. In the second subsection, we further evaluate the rationality of our HAS index results. The third subsection discusses the implications of our study and future work.

5.1. Comparison of MCDA and machine learning models

The five HAS maps derived from MCDA (AHP + Entropy and TOPSIS) and machine learning (RF, SVM, and LR) methods are shown in Figure 11. To assess the spatial agreement among these five maps, pixel-wise Pearson’s correlation coefficients are calculated (Figure 11(a)). The correlation coefficients range from 0.61 to 0.82, all of which are statistically significant (p-value < 0.05). Overall, the five maps demonstrate consistent spatial patterns. Specially, the regions near Wilkins, George VI, Baudouin, the southern side of Ronne-Filchner and the southern side of Amery show a relatively consistence in high and very high HAS class among the five maps. The regions falling under the low and very low HAS classes are mainly distributed in the edge of some ice shelves, such as Fimbul, Larsen C and Amery. Nevertheless, some variations in spatial pattern between the five maps are evident. The AHP + Entropy and SVM models tend to show lower HAS levels for ice shelves located in the West Antarctica sector as compared to other models. Additionally, the MCDA models show a lower HAS level for the Ronne-Filchner ice shelf as compared to the machine learning models. The RF model displays a lower HAS level for ice shelves located close to the Ross Sea, but a higher HAS level in the Queen Maud Land region.

Figure 11. Comparison of results based on MCDA and machine learning models. (a) Generated HAS maps based on five models and the pairwise correlation coefficients for these maps, (b) Area and relative percentage for different HAS classes.

The area and relative percentage of five HAS classes between the MCDA and machine learning models are exhibited in Figure 11(b). It reveals that their proportion is comparable. In the MCDA results, about 45% of Antarctic ice shelf area belongs to the high and very high HAS classes, while they occupied nearly 55% in the machine learning results, representing a difference of around 2.14 × 10⁵ km². The TOPSIS result shows lower HAS levels on the whole, with a total area of moderate to very low HAS classes measuring 9.58 × 10⁵ km². This contrasts with the LR result which shows the smallest area of moderate to very low HAS classes, measuring 6.66 × 10⁵ km².

It is shown that different techniques can affect the HAS index results and exhibit specific advantages and disadvantages. In this study, the performance of machine learning techniques is better than the MCDA techniques. Although MCDA methods are simple to develop and do not rely on a large sample database, they are difficult to deal with a large number of criteria and alternatives. Some of the MCDA methods, such as AHP, rely on subjective judgments when assigning weights to criteria and scores to alternatives, which may introduce bias and uncertainty. The machine learning approaches (SVM, RF and LR) require a training phase and are dependent on the goodness-of-fits, but they can identify patterns and relationships of variables automatically, leading to more accurate predictions and better decision-making (Zlaugotne et al. 2020). SVM can effectively handle a high number of HAS conditioning factors without being impacted by data dimensionality (Mehravar et al. 2023). RF uses numerous decision trees for ensemble learning. LR fits a linear location and is less complex (Shabanpour et al. 2022). This study also shows that the RF model outperforms the others. This ensemble technique generally outperforms individual decision trees in predicting accuracy and overfitting. The averaging effect of numerous decision trees makes it robust to outliers and noisy data (Rafiei-Sardooi et al. 2021). However, LR and SVM may need additional data preparation due to outlier sensitivity. SVM requires specific feature engineering and kernel selection, but RF may model complex nonlinear relationships without them. Thus, RF is observed as the best model for HAS mapping in the study area compared with other models.

5.2. Application discussion of HAS index map

In our results, we obtain the HAS index map based on the developed RF model. The training and testing dataset are used to evaluate the model performance against the metrics in Section 3.3. In this part, we further demonstrate the spatial consistency of our results with the actual environmental conditions and discuss how much application value the HAS index map with the resolution of 500 m can bring for human activities.

We use the Sentinel-2 imagery with a resolution of 10 m to compare our HAS class results with the ice surface features. We collect 12 test areas, each containing 16 (4 × 4) HAS map pixels. Meanwhile, Sentinel-2 data in January of 2021 for each test area is also collected and shown in Figure 12. Apparently, very high HAS class areas are around the crevasse-free blue ice zones, sometimes with some melt lakes as water supplies for human activities. Very low HAS class areas always contain harsh ice surfaces for human activities such as dense and large crevasses, shear zones. Moderate HAS class areas have more flat ice surfaces, but they are accompanied by obvious crevasses, longitudinal structures and meltwater. Therefore, the regions with moderate and lower HAS levels need more attention and detailed investigations.

Figure 12. Characteristics of HAS index and Sentinel-2 data over Antarctic ice shelves. (a) Larsen C. (b) Larsen D. (c) Brunt. (d) Baudouin. (e) Shackleton. (f) Moscow University. (g, h) Nansen. (i) Getz. (j) Dotson. (k) Crosson. (l) Pine Island.

By comparing the HAS index with the Sentinel-2 images, it can be seen that our results can reflect a comprehensive assessment of the actual environment. The 500 m locational suitability analysis for human activities from remote sensing data may not be as detailed as required for practical applications such as field investigations or tourism planning due to its limited spatial resolution. Nonetheless, it still offers scientific guidance during the initial planning phase of field experiments and provide valuable information for further explorations.

5.3. Implications of the present study and the future work

Following our research objectives, the findings of this paper would be useful in three aspects. First, all the conditioning factors are crucial variables that impact the HAS over Antarctic ice shelves. Our study can help the selection of the most appropriate HAS factors in this study region and beyond. Second, our HAS map results provide a comprehensive assessment based on national conditions. It identifies potential suitable locations for scientific explorations and research stations. Our findings can be useful in developing targeted management strategies for general planning and assessment purposes, taking into account different suitability characteristics. Third, the insights provided by this study demonstrate that our approaches are practical for analyzing the locational suitability and can be effectively applied in other regions and settings. Given the limited number of literatures on HAS mapping over Antarctica, the information presented in this paper provides baseline data based on well-known methods, which will be very useful in future research.

However, the main limitations of the present study arise from the data sources used and the lack of information on temporal changes. In future studies, more factors that influence human activity site selection can be considered in the evaluation models. Due to the complexity of the integrated suitability assessment, the HAS results are closely related to the whole environmental conditions and even human factors. For example, the ice sheet mass change and the fresh water sources of human activities may be also the potential factors that impact the scientific interest and logistics in Antarctic explorations. The Gravity Recovery and Climate Experiment (GRACE) satellite can observe the small temporal fluctuations in the Earth's gravity field and provide great information about the water storage change on a global scale, ice and snow mass change in polar regions and oceanic mass change (Mohamed et al. 2022). The observation results of GRACE have been applied to multiple fields to help solve social problems (Mohamed et al. 2023; Zhou et al. 2021). Therefore, the GRACE monitoring data can also be included in the future HAS assessment work. In addition, our future work will focus more on incorporating the results with temporal factors and further exploring the temporal trends with the climate change. As discussed in Section 5.1, it is also recommended to develop new individual or hybrid models and combine the strengths of different models to further improve the model accuracy.

6. Conclusions

This study raises a novel index named the HAS index to quantitatively evaluate the locational suitability of human activities over Antarctic ice shelves. Nine corresponding natural conditions about ice surface features, ice shelf stability, meteorology and topography are integrated into the suitability analysis. Two MCDA methods (AHP + Entropy, TOPSIS) and three machine learning methods (SVM, RF and LR) are tested for HAS modelling. The AUC, overall accuracy, kappa index and RMSE results of the training and testing samples indicate that the RF model outperforms the other models. Comparing with the high-resolution Sentinel-2 satellite images, we further discuss the application potential of our HAS index map for human activities. It shows that more than 30% of Antarctic ice shelf area is in the low and very low locational suitability classes. It is important to note that this does not mean that human activities can’t be carried out in these areas, but rather that more detailed field investigations and more attention are required. In the local analysis for Ronne-Filchner, Amery and Ross ice shelves, the HAS index shows a high level of heterogeneity. The peripheral parts and the upstream areas are usually at a lower level of HAS. Yet the middle parts are more suitable areas. Our research can provide a realistic view to decision-makers for human activity management over Antarctic ice shelves. It is useful for Antarctic environmental assessment and sustainable development.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data are openly available in a public repository. Antarctic coastline and groundling data used in this study are from National Ice Center (https://usicecenter.gov/Resources/AntarcticShelf). The ERA5 reanalysis data were downloaded from the website (https://cds.climate.copernicus.eu/). The Antarctic ice velocity data were downloaded from the website (https://nsidc.org/data/nsidc-0484/versions/2). The MODIS Mosaic of Antarctica data were downloaded from the website (https://nsidc.org/data/nsidc-0730/versions/1). The fracture location map of Lai et al. (2020) is available at https://doi.org/10.15784/601335. The snow accumulation data were downloaded from the website (https://legacy.bas.ac.uk/bas_research/data/online_resources/snow_accumulation/index.php). The ICESat-2 DEM data were downloaded from the website (https://data.tpdc.ac.cn/en/disallow/9427069c-117e-4ff8-96e0-4b18eb7782cb/). Sentinel-1 imagery is available in Google Earth Engine (https://earthengine.google.com/).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

This work was supported by International Research Center of Big Data for Sustainable Development Goals: [Grant Number No. CBAS2022IRP03]; National Natural Science Foundation of China: [Grant Number Nos. 42276252, 42201148 and 42101124]; the Joint Project of the Chinese Academy of Science (CAS) entitled Using Earth Observations to Address Ecology and Environment Change in the Pan-Antarctic Cryosphere: [Grant Number No. 183611KYSB20200059].