Human activities are predicted to cause an increase of up to 5°C in the global average temperature by the end of the 21^st century (IPCC, 2021). In addition, the global rise in temperatures is melting glaciers and ice accumulations (Geyman et al., 2022; Sirmacek and Vinuesa, 2022), which may lead to a significant reduction of Arctic ice extension and duration in future decades. This rapid and marked increase in warming ocean temperatures and melting sea ice is also leading to seasonal shifts in habitat use of some marine mammals (Insley et al., 2021), which could exacerbate the extinction risk of ice-associated species by population depletion. Furthermore, as the speed of environmental change is higher in marine systems than in terrestrial systems (Poloczanska et al., 2016), some species may be more vulnerable to this new climatic scenario than others (Simmonds & Isaac, 2007; Moore & Huntington, 2008). Traditionally, cetaceans living in remote areas have been monitored through aerial line transect surveys from planes, among others (e.g., DeMaster et al., 2001; Heide-Jørgensen et al., 2010). Cetacean aerial surveys imply in a significant amount of time and money and can lead to population underestimations if researchers cannot cover the whole distributional range of the studied species (Kaschner et al., 2012). Classifying freely available satellite imagery using artificial intelligence allows for low-cost identification and monitoring of large cetacean populations over time (Guirado et al., 2019). In addition, traditional methods to forecast species distributions over different time periods after aerial surveys usually comprise species distribution models (SDMs) (Bouchet et al., 2019), although predictions from SDMs may be less accurate over long time frames or in periods and areas undergoing rapid warming (Franklin, 2010). By combining animal object detection through artificial intelligence and climate change projections from the Copernicus Climate Change Service (C3S), we can contribute both effectively and at low costs to identify and establish new high seas habitat protected areas, which may be essential in the conservation of beluga whales and other threatened species. Cetacean species with a range limited to non-tropical waters are more likely to be impacted by water temperature alterations (MacLeod, 2009). Beluga whales (Delphinapterus leucas) (Pallas, 1776) are strictly ice-dependent species that inhabit isolated Arctic Ocean waters (Suydam et al., 2001) and, thus, may comprise good bioindicators concerning how climate change can affect other ice-dependent Arctic species. They occur in coastal and offshore waters of western and northern Alaska and Canada (Richard et al., 2001; Suydam et al., 2001) and Russia (Shpak, et al., 2019), comprising the most geographically isolated cetacean (Laydre et al., 2015). To estimate their abundance, aerial surveys began in Cook Inlet and in the Gulf of Alaska in 1993, documenting a 42.6% decline (1993-2006) (Hobbs et al., 2006). Due to these critically low numbers, beluga whales were listed as endangered (Goetz et al., 2012), and their current situation seems somber as, out of 19 populations, 14 are data deficient, three are declining, one is stable and only one is increasing (Laydre et al., 2015). The combination of natural and anthropogenic factors, such as climate change and/or illegal harvesting, may be hindering beluga population recoveries (Carter and Nielsen, 2011).

Traditionally, long-term marine mammal monitoring techniques have comprised direct observation, and, during the last decades, the use of satellite tracking tools has efficiently revealed individual movement patterns and distributions (Kennedy et al., 2013; Scales et al., 2017). Object detection photo identification is significantly more accurate than visual observation (Chernetsky & Krasnova, 2018), with lower costs than traditional counting techniques in sampling animals living in remote areas (Taylor et al., 2007). Recently, remote sensing imagery has been used to identify both whale individuals and aggregations allowing for estimates concerning population abundance. This technique, however, still does not provide a high detection accuracy for marine mammals (Abileah, 2002; Fretwell et al., 2014; Fretwell et al., 2017), which makes it difficult to develop reliable estimates concerning spatial distribution or aggregation patterns.

In remote sensing methods, free Google Earth images have been used to train Deep Convolutional Neural Networks (CNN) for conservation goals (Guirado et al., 2017), which automatically learn the distinctive features of each object class from a large set of annotated images (LeCun et al., 2015). Previous studies in marine mammal detection and counting have resulted in a recall rate of more than 80% (Bogucki et al., 2019; Guirado et al., 2019; Gabaldon et al., 2022; Khan et al., 2022). This can have a big impact in reducing the effort required for manual verification, increasing the advantage of employing an automatic detector in long-term monitoring (Li et al., 2022; Marquez et al., 2022), estimating population abundances and projecting dynamics and fluctuations under future climate change scenarios. Therefore, this technology provides an essential tool in predicting, assessing and implementing conservation measures, which may, in turn, contribute to the protection of these and other marine and ice-dependent species.

Our aim here is to identify beluga whale individuals and aggregations in the Arctic Ocean from high-resolution satellite imagery employing Deep CNN to automatically detect beluga whales from open Google Earth images, and assess their correlation with sea surface temperature (SST) projected scenarios and with beluga-data to evaluate potential changes in beluga distribution. To this end, this study (i) detected beluga whale individuals and aggregations, (ii) assessed intra-population spatial patterns and their relationship to SST, and (iii) compiled SST future projections at beluga whale locations from 2015 to 2100 under three IPCC representative concentration pathway (RCP) scenarios.

We were able to detect six beluga whales’ aggregations for during the summer feeding period between 2007-2020 using open-access satellite imagery and deep learning. A total of 1,980 beluga whales were detected, grouped in eleven aggregations at the Cunningham Inlet, Nunavut land and Hudson Bay (Canada), Norton Sound (USA), and North and South Ojotsk sea (Russia) ( Figures 1 , 2 ; Supplementary Table 1SI ).

In a first test in the bay where the North Elwin River flows (73.554006396, -90.96751917) with Google Earth images, the model trained for the detection of the family Monodontidae based on Faster-RCNN Inception-Resnet v2 of Tensorflow API obtained 99% accuracy. In the test phase, we obtained a 98% F1 measurement of D. leucas ( Supplementary Table 2SI ) with images exclusively from Google Earth in the classification for object detection using bounding boxes ( Supplementary Figure 2SI ). For the iceberg class the measured F1 was 100%. However, in other bays further north and with more mixed icebergs the automatic results were not good (F1 < 60%). Therefore, we proceeded to manually photo-interpret the other bays with high probability of beluga whale presence.

The intra-population spatial patterns (L-function) of many individual beluga whales displayed low minimum Nearest Neighbour distances, following a quadratic relationship with the Clark-Evans R index (R^2 = 0.65) ( Supplementary Figure 3SI ).

To assess the relationship between SST and these aggregations, we calculated the correlation between the SST (°C), detecting a positive correlation between SST and Nearest Neighbour distance (r = 0.65, p = 0.042), where most aggregations were located at areas with low SST, ranging from -2 to 5°C, except for Russian locations ( Figure 3A ). The averages of temperature ( Figure 3B ) also highlighted how the Arctic Sea, including the Russian and North Canadian locations, are increasing SST mean records along the last decades. Our results highlighted that SST is expected to increase at least one degree before 2050 in the three scenarios, but will increase from 1.5 to 3°C from 2050 to 2100, depending on the RCP scenario ( Figure 3 ), with northern areas comprising the most affected under the three RCP scenarios ( Figure 4 ; Supplementary Figure 3SI ).

Here we employed satellite imagery and Deep CNN to identify individual beluga whales and assess their patterns, obtaining accurate results (Wu et al., 2022). However, not in all bays the accuracy was good, and we proceeded to make a manual effort to obtain the final results, using a lot of effort and time. The lesson learned from this attempt to automate the detection of beluga whales with satellite images would be that the spatial resolution of the free images used is still low for detecting beluga whales (5m long without submerging). In addition, individual automatic detection is easily confused with icebergs, with the same color and sometimes size or shape. Finally, we are optimistic for future work to increase the accuracy of beluga whale detection using other strategies such as the detection of sets of whales instead of isolated individuals, as they are often found in groups where they could be better identified.

Our results highlighted that the rapid changes the Arctic region is now facing due to climate change (Poloczanska et al., 2016) emphasizing the importance of learning more about beluga whale spatiotemporal distributions and latitudinal habitat use shifts. (We detected herein a positive and significant relationship between SST and whale aggregation, suggesting a higher number of beluga whale aggregations in colder waters. However, the correlation value could be influenced by the distinctive characteristics of two Russian populations with low aggregation of whales ( Figure 3A ) compared to other populations. Overall, this relationship may be due to the fact that many areas, such as Baffin Bay, have experienced significant sea ice changes, with sea ice retreating 7 days earlier and advancing 5 days later per decade (Laidre et al., 2015). In West Greenland, there was a clear correlation between decreasing sea-ice cover and increasing distance from shore for beluga sightings. Despite their relatively restricted distribution close to the coast and in limited open areas in the early 1980s, when sea ice cover was higher (Heide-Jørgensen et al., 2010). The northern archipelagos in the Barents Sea have also been recording annual warming of up to 2.7°C per decade, with a maximum up to 4°C observed in some periods (Isaksen et al., 2022). The higher records also observed in the Russian locations may be due to the Arctic Sea experiencing the highest SST ever recorded since increasing SSTs in the Northern Pacific over the past decades, with associated changes in extreme marine heatwaves (Carvalho et al., 2021). Current Arctic sea-ice coverage levels are the lowest since 1850 for both annual mean and late-summer values and the region will become practically ice-free by the end of the 21^st century in both low and high greenhouse gas emissions scenarios such as SSP2-4.5, SSP3-7.0, and SSP5-8.5 (high confidence) (Fox-Kemper et al., 2021) (IPCC, 2021).

The ocean absorbs almost 90% of the Earth’s energy imbalance (Von Schuckmann et al., 2020), which is accelerating ocean warming due to climate change (IPCC, 2021). We downloaded the IPCC SST projections at the eleven study locations to highlight whether those beluga whale populations could potentially be affected by sea surface warming in the next decades, and our findings indicate that all the locations might experience a one degree temperature increase from 2015 to 2050 under the three RCP scenarios. However, the warming pattern diverges from 2050 depending on the RCP scenario, demonstrating how, depending on the emissions path we choose, the SST will increase at those locations from 1.5°C to 3°C, with northern locations more affected than southeast ones ( Figure 4 ). Consequently, if our SST IPCC scenarios are confirmed, northern populations could be experiencing deep sea thermal changes in their distribution areas from the 50s decade ( Figure 5 ). This could, in turn, push their distribution patterns towards cooler southern areas, overlapping with northward distribution shifts of other marine mammal species (Thorne et al., 2022), increasing habitat and resource competition or even leading to drastic population depletion, depending on species ability to adapt to new climate scenarios.

Overall, our study with a low-cost tool to monitor cetaceans using freely available satellite imagery and artificial intelligence demonstrated how beluga whales aggregate less in higher SST and how their habitats are predicted to warm throughout the 21^st century. Spatial beluga whale population patterns are directly associated to sea temperature and climate change will certainly affect them in the future decades. However, the chosen IPCC greenhouse emission path will determine the extent and speed of SST increases for the rest of the century. Under this scenario, we recommend a prompt and global reduction of gas emissions and the protection of beluga whale territories from marine traffic if we wish to preserve this vulnerable species. In this regard, this study, along with future research to detect thermal shelters in the Arctic may comprise valuable tools in highlighting potential climate change impacts on beluga whale populations.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding author.

Ethical approval was not required for the studies on animals in accordance with the local legislation and institutional requirements because the study was observational only.

MR: Writing – original draft, Writing – review & editing. EG: Conceptualization, Formal analysis, Writing – review & editing. ZO: Formal analysis, Supervision, Writing – review & editing.