Weaving Together Ecological Data with Indigenous Knowledge to Model Environmental Factors Impacting Rubus chamaemorus Productivity in Southwest Alaska

Highlights

What are the main findings?

• Indigenous Knowledge revealed environmental factors previously not explored in the literature including elevation change due to microtopographic differences in the tundra and an earlier harvest season.
• Multiple chlorophyll-based vegetation indices, MERIS terrestrial chlorophyll index, green–red vegetation index, and chlorophyll carotenoid index, were predictive of berry harvest metrics in southwest Alaska.

What are the implications of the main findings?

• Indigenous Knowledge was critical to establishing parameters of the ecological variables of the subsistence harvest metrics.
• Transdisciplinary research is a necessary element to incorporate years of observations from community experience with predictive models for difficult to access subsistence.

Abstract

The spatial distribution and productivity of subsistence resources are central to food security, nutrition, and cultural vitality in circumpolar Indigenous communities. Yet few studies incorporate Indigenous Knowledge in methodology to monitor subsistence plant species. Here, we apply participatory action research to develop a monitoring system for the culturally and nutritionally important Rubus chamaemorus (atsalugpiaq, salmonberry) near the Yup’ik village of Quinhagak in southwest Alaska. With support from community members, two ground-truth surveys assessed berry productivity at nine sites within Quinhagak’s Traditional Land Use Area. Seventeen interviews identified key themes related to subsistence harvest and highlighted winter meteorological factors important for analysis. We compiled a multi-year dataset including PlanetScope eight-band SuperDove imagery (3 m GSD); airborne LiDAR and satellite-derived DEMs; and four meteorological parameters. Linear regression and multiple adaptive regression splines were tested to evaluate relationships among vegetation health, climate, landscape features, and berry productivity. Model outputs identified chlorophyll-related vegetation indices, particularly MTCI, as strong predictors of harvest outcomes, with higher flowering-season MTCI values associated with greater berry abundance. This work establishes a foundational, scalable approach for the long-term monitoring of Arctic subsistence plants in conjunction with Arctic communities and demonstrates the value of multi-layer data integration in regions historically challenging for remote sensing and ground surveys improving outcomes for regional harvest predictions and increased understanding of possible mechanisms controlling berry productivity in Arctic regions.

1. Introduction

Rural circumpolar communities utilize subsistence plant species for food and medicine. The presence of subsistence plants in the landscape is strongly influenced by climatic variables [1,2]. While plants represent only a fraction of the dietary intake of Arctic communities, this food group is representative of strong cultural vitality and an important source of nutrients not found in other foods consumed in the region [2,3,4,5]. Knowledge and traditions of subsistence plant harvest are embedded within oral histories and place names which indicate landmarks and other location-specific resources [6]. One subsistence plant species, atsalugpiaq (i.e., Rubus chamaemorus, also known as salmonberry and cloudberry), is considered an important berry species consumed in the Yukon–Kuskokwim (Y-K) Delta, a riverine delta system that is home to Yup’ik Alaska Native communities [7,8]. Over the last three decades, Indigenous and non-Indigenous researchers have reported an increase in the variability and abundance of multiple subsistence species throughout the Arctic, including in the Y-K Delta [7,9,10].

Atsalugpiaq is a perennial forb with a circumpolar distribution predominantly found in ombrotrophic peatlands [11]. Ombrotrophic peatlands are convex raised bogs where the sides of the bog are strongly influenced by groundwater and the central part of the bog is predominantly influenced by water from atmospheric precipitation [12]. Mineral soil water is distinct as it contains elements such as iron, aluminum, silicon, and manganese not found in precipitation, and the concentration of these elements is dependent on the bedrock parent material [12]. The plant grows 10–30 cm high where 90% of the plant remains below ground as a network of rhizomes [13]. Flowers bloom from early June to early July with fruit set occurring within 35–47 days following flower appearance [14]. Atsalugpiaq is among the first plants to begin flowering after the frost and the flowering period is influenced by the abiotic conditions of the current and prior years [15]. The resulting fruit is a compound drupe with a red coloring that slowly turns to a salmon color during fruit ripening [16]. Arctic communities’ respect for atsalugpiaq gives rise to specific considerations when researching, thus emphasizing the need for enhanced co-production methodologies for the perceived success of the present study.

Indigenous Knowledge co-production (CPK) is a collaborative, dynamic, and iterative process in which Indigenous Knowledge holders and researchers work as equitable partners to generate actionable knowledge that draws on multiple epistemological frameworks [17,18]. Rather than extracting traditional ecological knowledge as raw data for integration into Western science, effective co-production centers Indigenous data sovereignty, ensures that knowledge holders direct and benefit from research outcomes, and recognizes that Indigenous and Western knowledge systems are not in opposition but operate through complementary ontologies and ways of knowing [18,19]. In circumpolar and subarctic regions, CPK has emerged as a critical method for addressing climate-related threats to Indigenous subsistence, governance, and heritage [17]. In the Y-K Delta, the collaborative scholarship of anthropologist Anne Fienup-Riordan with the Calista Elders Council has been foundational in demonstrating that Yup’ik and Western scientific frameworks engage distinct but complementary organizational and taxonomic logics [20].

This body of work gave rise to co-produced ethnobotanical surveys—including Jernigan’s regional guide developed through Elder Councils representing thirty-one Elders from thirteen villages [21] and Fienup-Riordan, Rearden, and Meade’s comprehensive record of nearly one hundred Yup’ik men and women’s knowledge of edible and medicinal plants across southwest Alaska [22]—that demonstrate how Yup’ik plant nomenclature encodes ecological and cultural information absent from Linnaean taxonomy. For example, while botany assigns the single binomial Rubus chamaemorus L. to salmonberry across its circumpolar range, Yup’ik communities in the Y-K Delta use regionally differentiated names—atsalugpiaq (Kuskokwim), naunrat (mid-coastal region), aqevyiit (Hooper Bay and Chevak), and atsat (Nunivak Island)—each grounded in local dialectal tradition and carrying contextual knowledge of habitat, harvest timing, and use [19,21]. This use-based organizational logic extends broadly across Yup’ik ecological knowledge: where Western taxonomy classifies wood by genus and species, Yupiit use terms such as kenqeggialnuq—denoting wood that does not burn well—that prioritize functional and subsistence value over phylogenetic identity [19]. Rather than competing frameworks, these classificatory systems are complementary, with Yup’ik nomenclature providing actionable, place-based knowledge that species-level binomials alone cannot capture [19,20].

Scarcity and variability of subsistence plant species are driven by environmental change [23,24]. Atsalugpiaq, in particular, is predominantly found in ombrotrophic peatlands which are characterized by an abundance of moss and lichen mats [15]. In southwest Alaska, Yup’ik communities report a higher incidence of atsalugpiaq on the sides of tundra moss mounds [25]. The biomass of this plant species and flowering period are strongly affected by the previous year’s precipitation, temperature, and snow cover as the reproductive buds form a year before emergence [15,26,27]. Permafrost thaw, soil erosion, warmer temperatures and extreme weather conditions threaten the stability of Arctic ecosystems where atsalugpiaq dominates [28,29,30]. Interannual shifts in vegetation distribution and weather variability have far-reaching consequences on the phenology and growing range of plant species important to Yup’ik subsistence practices known collectively as Yuuyaraq (trans. “Our genuine way of living”) [31,32,33]. Assessments of sphagnum moss plots reported modest (~10–20%) reductions in moss growth in response to increased temperature and decreased annual rainfall [34]. A 15-year analysis in Norwegian alpine forests found species of lichen and moss reduced in cover, while graminoids had significant increases in cover [35]. Throughout these circumpolar contexts a continual feedback exists between environmental variability, subsistence animals and subsistence plants such that a shift in one element of the cycle has multi-year impacts on each resource [36,37]. As a result of environmental change, there are also reports of Alaskan Native communities traveling longer distances to harvest subsistence plants, which results in higher costs and additional risks [7,38]. Communities, therefore, need predictive tools for enhanced monitoring and assessment to secure resources for current and future generations.

Today, widespread vegetation change is a prominent observation throughout all Arctic ecosystems [39,40,41]. Monitoring of arctic ecosystems has demonstrated changes along the coastline and interior of Alaska indicating alterations in the vegetative composition of the ecological landscape [42,43]. Extensive field surveys have been conducted, to generate species-specific spectral curves and elucidate plant presence, including subsistence plant species [44,45]. However, widespread field surveys are limited due to inclement weather, high cost, and limited infrastructure [46,47]. In comparison, remote sensing has been used across the Arctic for plant health monitoring [48,49], assessing vegetation change [50,51], and detecting landscape change from disturbances [52] in a variety of ecological modeling studies. These studies use remote sensing to reduce labor and resource inputs while increasing automation in landscape management, with work by Davidson et al. [39] and Macander et al. [40] showing that satellite-derived vegetation indices and related environmental variables vary in their ability to predict dominant vegetation classes in northern Alaska, as Davidson et al. [39] found that a single index and 2 m satellite data were insufficient for capturing tundra heterogeneity, while Macander et al. [40] demonstrated that vegetation indices were the most predictive features across random forest model iterations. However, the studies do not focus on specific subsistence resource harvest potential. Furthermore, the integration and comparison of multiple machine learning regression methodologies have the potential to provide enhanced forecasting abilities for food resources.

Species distribution modeling has been the dominant method to delineate areas suitable for plant species known to support Indigenous cultural vitality [53,54,55,56,57]. Research conducted in the Amah Mutsun stewardship area of northern California, for instance, uncovered dormant ethnobotanical knowledge of subsistence plants using spatial distribution modeling, bioclimatic data, and community science data to identify potential areas that were easily accessible to harvest subsistence plants [57]. Such studies address an important goal of many Tribal communities: to nourish the knowledge and traditions of subsistence plant harvest. Spatial distribution models of Canadian Vaccinium species revealed a reduction in habitat suitability in the southern regions but an expansion of habitat suitability in northern regions driven by soil bulk density, mean annual temperature and winter precipitation [55]. While these methods reveal the possibility of a species’ occurrence in a specific region, they do not capture abundance or productivity, which directly impacts food security for rural Arctic communities. This further highlights a lack of empirical evidence to support the food security and agency needs of communities who rely on subsistence plants for their diet.

Previous research conducted in pricky pear patches [51], forest stands [58], and cultivated crops [59,60] have collected ground sampling and remote sensing data to assess biomass, species diversity, and ultimately determine landscape features conducive to sustain productivity. However, Alaska-based studies thus far only assess spatial distribution modeling within the context of the whole state or specific regions to make predictions on how future environmental alterations could impact Alaskan subsistence berries. Baer [53] identified mean winter temperature, elevation and functional vegetation cover as important predictors of Rubus spectabilis (salmonberry) and Vaccinium alaskaense (blueberry) suitability in southeast Alaska. In southwest Alaska, Hamilton et al. [54] found that elevation, soil organic matter and bulk density demonstrated an average 10% effect on the potential occurrence of atsalugpiaq in the region. Rhodes [56] found spring snowfall, landcover, and average annual temperature to have the largest effect across all Alaskan regions in spatial distribution models of atsalugpiaq. In Quinhagak, Indigenous Knowledge has documented the decrease in berry size and quality over time [25]. In this field, there is a knowledge gap in the amount of ground-truthed data that can be directly linked to the vegetation health and productivity of individual plant species in the circumpolar north. The present research addresses this gap by collecting specific harvest metrics for atsalugpiaq and directly relating them to abiotic and biotic conditions within the data collection areas.

The overall aim of this research is to understand drivers of harvest variability in atsalugpiaq and determine which potential drivers show the strongest correlations to atsalugpiaq harvest metrics across years and locations. We applied Indigenous Knowledge and remote sensing modeling techniques to achieve this aim. There were ultimately three specific objectives: (1) identify Indigenous Knowledge trends of environmental variables that impact atsalugpiaq harvest; (2) analyze two years of spatial trends in vegetation and berry harvest potential based on climate, ground and remote sensing data; and (3) compare the predictive performance of the collected data types to estimate berry harvest metrics. This research provides the groundwork to support Arctic communities’ goals for greater food security through enhanced vegetation monitoring which is grounded in Indigenous Knowledge and practices.

2. Materials and Methods

2.1. Co-Produced Research Practices and Procedures

Effective co-production of knowledge should nourish trust and respect, relationships, empowerment, means and ability, capacity, ethics, decolonization, and sovereignty. These values need to be maintained throughout the research process while integrating Indigenous and Western science [61]. Co-produced research is not a straight path nor is there one recipe to have a successful outcome. Co-produced research relies on humility, time and dedication [62]. Quinhagak has a history of significant investment and commitment to co-producing archeological research [62], documenting Indigenous Knowledge [63], community-based remote sensing [6,64], and ethnobotanical research [22].

Nalaquq LLC, an ANCSA 14(h) subsidiary located in Quinhagak, Alaska, has a mission of community empowerment through technological solutions. Nalaquq LLC met the principal investigator (PI), Claire Friedrichsen, in winter 2023 and invited her to visit Quinhagak during Atsalugpiaq harvest season in August 2023. The initial field visit included trust and relationship building, identifying research questions, developing research methods, and identifying project outcomes that would create the largest positive impact for Quinhagak. Prior to meeting the PI, Nalaquq LLC identified through a co-production research process involving community listening sessions that variable atsalugpiaq berry harvest was a critical research priority for Quinhagak [19]. Nalaquq LLC intentionally chose to partner with United States Department of Agriculture, Agricultural Research Service, to create a positive federal research presence in the community. For Quinhagak, this project was a first collaborative research project integrating leadership from Nalaquq LLC, the Native Village of Quinhagak, and Qanirtuuq Corp (Q-Corp). The PI co-produced with Nalaquq LLC and Quinhagak leadership an interdisciplinary research plan that focused on employment of Quinhagak community members. Quinhagak community members were employed throughout the research project. Additional projects included within the co-production plant includes co-development of geo-spatial tools to monitor the health of atsalugpiaq, a short documentary on the impacts of environmental change on atsalugpiaq, food safety assessments of atsalugpiaq and oral histories of Indigenous Knowledge. This research study represents the first effort to develop a geo-spatial tool as envisioned by Nalaquq LLC to monitor the health of atsalugpiaq in Quinhagak, Alaska. Nalaquq LLC received funding from the USDA Agricultural Research Service, Northern Great Plains Research Laboratory, as a pilot project and included language on agreed-upon Indigenous data sovereignty protocols (Agreement No: 58-3064-4-004). For additional information on the co-production process see Gleason et al. [19] and Friedrichsen et al. [25].

Our co-production methodology draws on previous remote sensing projects conducted by Nalaquq, LLC (Yugtun: “it is found”), a community-based research organization in Quinhagak whose co-production framework grew out of the Nunalleq Archaeological Project. Initiated in 2009 on the request of the Native Village of Quinhagak after pre-contact Yup’ik artifacts were found eroding onto Bering Sea beaches, the Nunalleq Project established a model of collaborative inquiry and power-sharing between the descendant Yup’ik community and university-based researchers that has since extended into environmental monitoring [65,66,67]. Nalaquq formalizes this ethos as a CPK framework for combining sensor-based remote sensing with Yup’ik traditional knowledge to study ellavut (trans. “our land and weather”) [19]. Building on prior collaborative remote sensing projects in Quinhagak’s Traditional Land Use Area—including multispectral satellite and ethnographic surveys of cultural heritage sites threatened by coastal erosion [68], UAV-based heritage landscape monitoring [69], and participatory mapping of erosion threats to Yup’ik subsistence areas—Nalaquq’s approach is structured around three interconnected commitments: (1) developing local research capacity so that communities can independently collect, interpret, and own environmental data; (2) ensuring Indigenous data sovereignty through community-controlled archiving and intellectual property agreements; and (3) producing actionable, community-identified datasets that support climate change adaptation [19].

The co-production process for the present study began in 2023, when Nalaquq representatives introduced the project concept to USDA researchers at an Alaska Small Business Conference, sharing prior remote sensing outcomes and community-identified resource management needs. Monthly meetings with USDA tribal liaisons followed to secure funding and scope the project collaboratively. During this period, Nalaquq and USDA researchers worked with the Tribal council (Native Village of Kwinhagak) and the ANCSA village corporation (Qanirtuuq Incorporated) to define researcher compensation, project goals, and intellectual property rights. The project was formally presented to village members at a community annual meeting in June 2023. Nalaquq and USDA then co-drafted an intellectual property agreement specifying how data would be collected, archived, and owned by the village prior to filing a human research clearance with the University of Alaska Fairbanks (IRB #214867). In 2024, USDA co-investigators and Nalaquq representatives traveled to Quinhagak, where community meetings were hosted and local knowledge bearers identified traditional place names associated with atsalugpiaq harvest areas. Following initial site visits, Nalaquq selected five sites for analysis and organized community-based listening sessions, also hiring a local Yup’ik traditional technician to assist with data collection. These interviews were designed to document community concerns regarding atsalugpiaq harvest and to co-register local ecological knowledge with candidate remote sensing datasets. To broaden community engagement, Nalaquq and USDA co-produced a short documentary soliciting additional community input. After transcribing and thematically coding interview data, Nalaquq and USDA researchers collaboratively identified remote sensing datasets, including NOAA lidar-derived Digital Elevation Models and multi-year satellite imagery, whose spatial and temporal characteristics aligned with community-reported observations about atsalugpiaq distribution, habitat quality, and harvest location.

2.2. Study Site Description and Attributes of Participatory Research Methods

The study area, Quinhagak, Alaska, is a coastal tundra biome. Data was collected in July and August of 2024 and 2025. In the 2024 field season, five sites were defined for data collection using a Garmin 67i handheld GPS (Garmin Ltd., Schaffhausen, Switzerland). The 2024 field season was established to delineate general areas associated with atsalugpiaq harvest in relation to relevant placenames. In the 2025 field season, nine sites were delineated using an Emlid Reach RS3 Differential GPS (Emlid Tech Kft., Budapest, Hungary). Site locations varied in size and extent across years. The dimensions of each site and other relevant site descriptors are detailed in

Table 1. The coordinates of the sites selected for atsalugpiaq sampling and model development for 2024 and 2025. The number of transects collected is listed along with the site size.

2024 Sites
Site	Centroid Longitude	Centroid Latitude	Number of Quadrats	Size (Hectare)
Airport	161°52′16.31″W	59°45′23.09″N	6	1.44
Cellphone Tower	161°54′24.94″W	59°44′01.63″N	6	3.99
Dump	161°53′20.10″W	59°42′30.90″N	6	0.11
Gravel Pit	161°51′39.28″W	59°45′44.17″N	6	0.49
Old Wastewater Lagoon	161°53′25.46″W	59°44′55.84″N	6	0.70
2025 Sites
Airport	161°50′14.64″W	59°44′58.08″N	12	2.12
Dump	161°53′16.42″W	59°42′29.92″N	16	3.82
Old Wastewater Lagoon	161°53′27.61″W	59°44′55.59″N	12	1.84
Cellphone Tower	161°54′17.47″W	59°43′48.18″N	16	3.42
Misnuq Camp	161°50′51.60″W	59°42′37.34″N	10	4.04
Gravel Pit	161°51′35.40″W	59°45′44.12″N	10	2.91
Siniq Site 1	161°53′12.16″W	59°59′35.67″N	10	0.89
Siniq Site 2	161°54′30.61″W	59°59′23.19″N	10	1.10
Siniq Site 3	162°02′22.25″W	59°59′34.67″N	8	0.36

and visualized in Figure 1. The Siniq locations are 30 km north of the main township in an area which is widely used for subsistence berry harvest and is only accessible by boat. Siniq sites were only assessed in 2025 due to low water levels in the riverways connecting Quinhagak and Siniq in the 2024 field season. Siniq subsistence harvest areas are notable for their carpet-like coverage of atsalugpiaq on small but numerous islands with mounds formed by the seasonal water inundation driven by permafrost thaw which creates pockets of high and low terrain [42,70].

Sites were selected through a participatory research framework in consultation with Yup’ik leaders in Quinhagak who helped frame the research questions and site selection. In total, four members of the Quinhagak community joined a research advisory team along with researchers from Nalaquq, LLC, and Qanirtuuq Incorporated. Next, Yup’ik members of the research team selected the nine total sites within Quinhagak’s Traditional Land Use Area (TLUA), which are presently and/or historically important for subsistence berry harvest. One team member, Grace Hunter, also served as a compensated field technician employed by Nalaquq, LLC, to help non-Yup’ik researchers identify each location and co-design a methodology to collect data with respect to Yup’ik custom and belief. However, there are inherent limitations in the study, including that the number of sites visited each year was constrained due to the local conditions at the time of data collection and the number of days with which researchers had access to the field sites. The complete experimental workflow is outlined in Figure 2.

2.3. Qualitative Data Collection Methods

Seventeen oral histories were conducted in summer 2024 during atsalugpiaq harvest time. For a full overview of the methods, analysis techniques, and coding framework see Friedrichsen et al. [25]. Interview participants were selected to represent each family in Quinhagak, community respected experts in atsalugpiaq harvest, and were open to everyone who was interested in participating through announcements on social media [46]. To weave together environmental data with Indigenous knowledge, sections of the oral histories from Friedrichsen et al. [25] that were initially grouped under the theme of Indigenous knowledge of environmental relationship with berries were recoded at a finer level of detail to identify emergent codes and themes related to environmental change and its effect on atsalugpiaq. Oral histories were coded by hand, iteratively through a first and second round of inductive emergent codes. Codes were then grouped to form categories and then themes were formed by grouping categories together (

Table 2. Codebook of interviews of categories and codes under the theme of Indigenous Knowledge of environmental change impacts on atsalugpiaq.

Category	Codes
Topography	Bigger berries on sides of hill, berry quality size and density dependent on topography, cliffs protect berries from sun, berries found on banks/cliffs, reading landscape, topography impacts plant composition, grow on slopes
Weather	Sun burns berries, ripening depends on weather, too cold then no berries, berries dependent on weather, if no snow then no berries, thunderstorms
Plant ecology	Codependent plants, one edible species linked to all other edible species, moss provides water for berries, reading and understanding interaction between landscape and species
Indicators of harvest	Ecological indicators of harvest potential, cottongrass, predictable environmental indicators of harvest, grass holds snow and protects berries from weather, if cottongrass then there will be berries, snow cover is protection, predictability
Knowledge of ecosystem dynamics and change	Knowledge of observed changes, changes in ecosystem also linked to changes in animal behavior, knowledge of environmental change, indicators of change
Plant physiology	Berry physiology, 7-year growth cycle, knowing best growing conditions
Nutrient cycling	Snow carries nutrients
Geography	Ripening depends on geography (inland vs. coastal vs. north vs. south), north not as impacted by environmental change, Quinhagak is on edge of ecosystem boundary and most vulnerable, different geographies more plentiful, have to pick inland due to coastal erosion, knowledge of desirable berry picking locations, berry distribution varies with geography
Seasonality	Specific subsistence calendar and each berry has its time, environmental change is impacting seasonality
Insects	Pollination is important
Soil knowledge	Different words for different types of tundra, salmonberries depend on soil type, when there was permafrost use to use it for stage of berries, knowledge of different types of tundra
Wildlife	Follow geese to find preserved berries (mostly cranberries) under snow from last year, close relationship with wildlife

). Categories were then grouped into themes during iterative comparative review of each oral history [71,72]. Themes were formed through iterative informal oral participatory engagement over three years of field visits to Quinhagak, Alaska, during berry harvest time with community members [73]. Themes were also refined during peer debriefing occurring with the co-authors during bi-monthly meetings to refine and identify themes of high cultural consensus [72]. Important themes were collated into major variables to be considered for quantitative data collection (

Table 3. A selection of variables acquired to assess environmental drivers of berry harvest in Quinhagak, Alaska, based on both Indigenous knowledge and a literature review.

Variable	Interviews	Literature Review
Berry Size	[46]
Vegetation Cover	[46]	[44]
Elevation	[25]	[42]
Winter Snowfall	[46]
Spring Snowfall		[44]
Cottongrass	[21]	[22]

). Indigenous data sovereignty protocols outlined in Gleason et al. [19] were followed. The audio recordings and transcripts are archived at the Quinhagak Heritage Foundation at the Nunalleq Museum in Quinhagak [67], but restrictions apply to the availability of these data due to data privacy and ethics. These oral histories are, however, available from the Quinhagak Heritage Foundation upon reasonable request and with the permission of Quinhagak Heritage Foundation by writing to lchurch@nalaquq.com and with approval from the University of Alaska Fairbanks Institutional Review Board (IRB) with appropriate Human Research certification by writing to Claire.friedrichsen@usda.gov. The research was approved by Nalaquq, LLC, a subsidiary of an ANCSA 14(h) corporation, and by the IRB of the University of Alaska Fairbanks (#214867).

2.4. Weavnig Indigenous Knowledge and Western Science for Data Acquisition

Drawing from both a literature review and Indigenous Knowledge, we selected a set of predictors and response variables for this study. A short selection of these predictors is shown in Table 2. Variables such as vegetation cover and elevation appeared in both the literature and the oral histories, so they were prioritized for inclusion in the dataset. After reviewing variables found only in Indigenous Knowledge or only in the literature, we removed or adjusted them depending on whether a suitable data source could be identified. For example, daily photosynthetically active radiation (PAR) is an important moderator of plant carbon fixation and is relevant to the study species based on Indigenous Knowledge. However, the primary data source that provided temporally complete PAR data could not be resolved to a spatial scale appropriate for our study sites [74]. When possible, we identified alternative data sources or calculated estimates to approximate highly important variables. The complete list of variables used in the study is provided in Supplemental Table S1.

2.5. Ground Sampling Methods—2024

In accordance with Indigenous Knowledge which suggested that there were reductions in berry size and general changes in leaf properties. The 2024 field season proceeded with the intention to build a repository of spectral and size measurements on the target species, atsalugpiaq. As such berry size and leaf spectrometry measurements were collected. First, a 0.5 m² quadrat was placed on the ground to sample each site location. The number of atasalugpiaq berries and all present plant species within the quadrat were noted. Within each quadrat, three leaves were collected and reflectance properties were assessed using a portable leaf spectrometer (CI-710s SpectraVue Leaf Spectrometer; CID, Camas, WA, USA). In this field season, data collection began in late July to collect data on fully ripened berries. We noted that field data collection strongly coincided with subsistence harvest activities such that berries were actively being collected from sites during field surveys.

Two independent observations of berry count per site were made using a hand counter. The average between the two observations was the total berry count per site. Berry samples occupying a volume of three 100 mL bags were collected from each site. The mass of 30 berries was measured using a digital scale to calculate the average mass of a berry. Berry count was recalculated into count m⁻² for ease of comparison between sites. The count of berries m⁻² was an approximation of berry density per sampling location. Photographs of collected berries were imaged with an iPhone13 camera. Images of berries were processed in ImageJ 1.54 [75] to determine berry diameter.

2.6. Ground Sampling Methods—2025

Leaf spectrometer and berry size information acquired in the 2024 field season showed low variability and limited usefulness for the prediction of berry harvest metrics. As such, to ensure resource efficiency, methods were adjusted with coordination of the community advisory team. In the 2025 field season, researchers collected berry harvest metrics and vegetation survey data in mid-July to approximate the complete berry harvestable potential which would not be biased by subsistence activities. In 2025, a 1 m² quadrat was placed on the ground, and a photograph was taken at 1.5 m above each quadrat using an iPhone 13 camera. Within each quadrat, the presence of each plant species and the number of atsalugpiaq berries was noted. Across each survey location, berry counts were conducted in the same manner described in the 2024 ground sampling methods.

Photographs captured in 2025 were collected to estimate species-level plant fractional vegetation cover. Fractional vegetation cover and plant species presence were collected during the fruiting phenological stage for atsalugpiaq. Images were initially processed in ImageJ [75] to remove any area beyond the quadrat frame. A graphical image annotation tool, LabelMe, was used to manually annotate each image such that at least 90% of the vegetation within the quadrat was labeled [76]. Custom Python 3.7 scripts were then designed to determine site-specific fractional cover based on the number of annotated pixels per species: Species were manually coded into four functional groups (berries, shrubs, graminoids, mosses/lichen), and the functional group fractional vegetation cover was calculated per quadrat [41]. The graminoid group is not resolved to the species level as features needed to specify species were not present at the time of collection. Thus, all sedges and grasses are contained in the graminoid functional group. Grouping plants into functional types allowed for an analysis of the dominant types of species present within each site. Next, per species and functional group coverage were averaged across all quadrats at a site to determine the proportion of each species or functional group present. Proportions were adjusted into centered log ratios for modeling. The Shannon diversity index (H′) was calculated with pixel counts for each quadrat and subsequently averaged across each site as follows:

$$H^{\prime }\equiv -{\sum }_{i=1}^N{\displaystyle \frac{{\mathrm{Pixels}}_i}{{\displaystyle {\sum }_{j=1}^N}{\mathrm{Pixels}}_j}}\ln ({\displaystyle \frac{{\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_i}{{\displaystyle {\sum }_{j=1}^N}{\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_j}}),$$

(1)

where ${\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_i$ were the pixels annotated for i species and ${\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_i$ represented all annotated pixels in the image.

2.7. Remote Sensing Data Attributes and Processing

Summaries of each remote sensing data object are detailed in

Table 4. Remote sensing data attributes for modeling parameters in determining factors important to atsalugpiaq harvest metrics.

Data Type	Spatial Resolution	Temporal Resolution	Source
Weather	5 km	Monthly, October 2022 to October 2025	[77]
Elevation—LiDAR	0.05 m	Yearly, 2024	[78]
Elevation—ArcticDEM	2 m	Sporadic, 2011 to 2021	[79]
8-band satellite imagery	3 m	Sporadic, January 2024 to October 2025	[80]

. Monthly precipitation, relative humidity, snow water equivalent, and temperature data was acquired from TerraClimate from October 2022 to October 2025 [77]. Weather data were differentiated as dormant (October, November, and December of the prior year, as well as January, February, March, and April), green-up (May), flowering (June), fruiting (July and August), and defoliation (September) as estimated in the literature [14]. The average values were calculated for each year, phenological season, and parameter combination [14].

Next, PlanetScope eight-band SuperDove imagery (3m GSD) images were obtained from February 2024 to October 2025 [80]. Each scene was clipped to the extent of each individual sample site. Thirty-two vegetation indices were calculated for each date as outlined in Table 5. The vegetation indices were selected based on the ability to capture biophysical variation. Southwest Alaska is notable for its extensive cloud cover in the summer, which limits the number of available satellite images with no more than 30% cloud cover. Thus, the indices were averaged over the relevant phenological periods for atsalugpiaq as defined in the literature and from Indigenous knowledge, wherein 1 May to 5 June are the approximate days of green-up, 6 June to 10 July represent dates of flowering, 11 July to 10 August represent the emergence and subsequent harvest of fruit, and 11 August to 15 September indicate defoliation and the beginning of dormancy for species [15,81]. Vegetation indices were manually sorted into four distinct groups: greenness/chlorophyll, moisture, blooming, and structure to ensure different features of the vegetation were assessed [82,83].

Elevation data were derived from the Arctic Digital Elevation Model (ArcticDEM) mosaic, a 2 m resolution composite digital surface model generated from multiple time-stamped stereophotogrammetric strips acquired between 2011 and 2021 [79]. Each pixel in the mosaic represents a single best quality elevation estimate selected from available observations, rather than a temporal average. Within each study site, zonal statistics were used to calculate mean, minimum, maximum, and standard deviation of elevation [84]. A weighted mean elevation was calculated using the number of contributing strips per pixel to account for spatial differences in data density. Microtopographic differences at each site were evaluated by calculating the vertical range in terrain generating the elevation relief metric [84]. The vertical range was calculated by the difference in the maximum and minimum elevation within each site to estimate moundedness. Acquisition date layers were used to characterize the temporal range of observations contributing to each site but were not used to estimate elevation change.

A second set of elevation data was obtained from the National Oceanic and Atmospheric Administration (NOAA), which conducted a topobathymetric LiDAR survey of Quinhagak in 2024 [78]. LiDAR data were mosaicked with the Mosaic Raster tool in ArcGIS Pro 3.6.1. Relative elevation, standard deviation of elevation, and broad topographic position were derived from the mosaic and aggregated into a composite file. The elevation statistics derived from the LiDAR data provide an elevation dataset, which closely aligns with the perspective of berry growth based on traditional ecological knowledge. The composite was subsequently aggregated to site-level descriptive statistics. The ArcticDEM dataset represented a stable composite surface model in a landscape that is known for year-to-year variation in terrain structure [85]. The LiDAR dataset provided very high spatial resolution of current surface heterogeneity. Instead of creating redundant datasets, each elevation dataset incorporated complimentary information for downstream modeling.

Table 5. Vegetation indices calculated for atsalugpiaq harvest metrics. ρ indicates the reflectance value. B, G, G1, RE1, R, and N2 represent blue, green, green 1, red, red edge, and near-infrared spectral bands.

Vegetation Index	Equation	Reference
Greenness/Chlorophyll
Blue Normalized Difference Vegetation Index (BNDVI)	${\displaystyle \frac{\rho N2-\rho B}{\rho N2+\rho B}}$	[83]
Chlorophyll Carotenoid Index (CCI)	${\displaystyle \frac{\rho G1-\rho R}{\rho G1+\rho R}}$	[86]
Chlorophyll Red Edge (CI-RE)	${\displaystyle \frac{\rho \mathrm{R}\mathrm{E}1}{\rho R}}-1$	[87]
Green Chlorophyll Index (Clg)	${\displaystyle \frac{\rho N2}{\rho G}}-1$	[87]
Enhanced Vegetation Index (EVI)	${\displaystyle \frac{2.5\times \left(\rho N2-\rho R\right)}{\rho N2+6\times \rho R-7.5\times \rho B+1}}$	[88]
Green Chromatic Coordinate (GCC)	${\displaystyle \frac{\rho G}{\rho R+\rho G+B}}$	[89]
Green Leaf Index (GLI)	${\displaystyle \frac{2\times \rho G-\rho R-B}{2\times \rho G+\rho R+B}}$	[90]
Green Normalized Difference Vegetation Index (GNDVI)	${\displaystyle \frac{\rho N2-\rho G}{\rho N2+\rho G}}$	[91]
Green–Red Vegetation Index (GRVI)	${\displaystyle \frac{\rho G-\rho R}{\rho G+\rho R}}$	[87]
MERIS Terrestrial Chlorophyll Index (MTCI)	${\displaystyle \frac{\rho N2-\rho \mathrm{R}\mathrm{E}1}{\rho \mathrm{R}\mathrm{E}1+\rho R}}$	[92]
Modified Chlorophyll Absorption Ratio Index (MCARI)	$(\rho \mathrm{R}\mathrm{E}1-\rho R)-0.2\times (\rho \mathrm{R}\mathrm{E}1-\rho G)\times {\displaystyle \frac{\rho \mathrm{R}\mathrm{E}1}{\rho R}}$	[93]
Normalized Difference Vegetation Index (NDVI)	${\displaystyle \frac{\rho N2-\rho R}{\rho N2+\rho R}}$	[94]
NIR Green Difference Vegetation Index (GDVI)	ρN2 − ρG	[95]
Optimized Soil Adjusted Vegetation Index (OSAVI)	${\displaystyle \frac{(1.16\times (\rho N2-\rho R\left)\right)}{(\rho N2+\rho R+0.16)}}$	[93]
Photochemical Reflectance Index (PRI)	${\displaystyle \frac{\rho G-\rho G1}{\rho G+\rho G1}}$	[96]
Pan Normalized Difference Vegetation Index (PNDVI)	${\displaystyle \frac{\rho N2-\left(\rho R+\rho G+\rho B\right)}{\rho N2+\left(\rho R+\rho G+\rho B\right)}}$	[83]
Red-Edge Triangulated Vegetation Index (RTVICore)	$100\times \left(\rho N2-\rho \mathrm{R}\mathrm{E}1\right)-10\times (\rho N2-\rho G)$	[97]
Red Chromatic Coordinate (RCC)	${\displaystyle \frac{\rho R}{\rho R+\rho G+\rho B}}$	[89]
Red Normalized Difference Vegetation Index (RDVI)	${\displaystyle \frac{\rho N2-\rho R}{\sqrt{\rho N2+\rho R}}}$	[83]
Sentinel-2 Lai Green Index (SeLI)	${\displaystyle \frac{\rho N2-\rho \mathrm{R}\mathrm{E}1}{\rho N2+\rho \mathrm{R}\mathrm{E}1}}$	[98]
Blooming
Enhanced Bloom Index White (EBI_White)	${\displaystyle \frac{\rho R+\rho G+\rho B}{\frac{\rho G}{\rho B}\times (\rho R-\rho B+1)}}$	[99]
Enhanced Bloom Index Red (EBI_Red)	$(R/G)/\left(\right(G/B)\times (R-B+1\left)\right)$	[99]
Enhanced Bloom Index Yellow (EBI_Yellow)	$\left(\right(R+G)/B)/\left(\right(G/B)\times (R-B+1\left)\right)$	[99]
Normalized Difference Yellowness Ratio (NDYI)	${\displaystyle \frac{\rho G-\rho B}{\rho G+\rho B}}$	[83]
Plant Senescence Reflectance Index (PSRI)	${\displaystyle \frac{\rho R-\rho B}{\rho \mathrm{R}\mathrm{E}1}}$	[100]
Moisture
Canopy Moisture (NDMI)	${\displaystyle \frac{\rho N2-\rho \mathrm{R}\mathrm{E}1}{\rho N2+\rho \mathrm{R}\mathrm{E}1}}$	[101]
Normalized Difference Water Index (NDWI)	${\displaystyle \frac{\rho G-\rho N2}{\rho G+\rho N2}}$	[102]
Near-Infrared Reflectance of Vegetation (NIRv)	$\left({\displaystyle \frac{\rho N2-\rho R}{(\rho N2+\rho R)}}-0.08\right)\times \rho N2$	[103]
Water Band Index (WBI)	${\displaystyle \frac{\rho N2}{\rho R}}$	[104]
Urban/Wetlands Moisture Index (MNDWI)	${\displaystyle \frac{\rho G-\rho Y}{\rho G+\rho Y}}$	[105]
Structure
Modified Triangular Vegetation Index 1 (MTVI1-G1)	$1.2\times (1.2\times \left(\rho N2-\rho G1\right)-2.5\times (\rho R-\rho G1)$	[82]
Modified Triangular Vegetation Index 2 (MTVI2)	${\displaystyle \frac{1.5\times (1.2\times \left(\rho N2-\rho G1\right)-2.5\times (\rho R-\rho G1)}{{\left(2\times \rho N2\right)}^2-5\times {\left(6-\rho N2-5\times \left({\rho }^{0.5}\right)-0.5\right)}^{0.5}}}$	[82]

Table 5. Vegetation indices calculated for atsalugpiaq harvest metrics. ρ indicates the reflectance value. B, G, G1, RE1, R, and N2 represent blue, green, green 1, red, red edge, and near-infrared spectral bands.

Vegetation Index	Equation	Reference
Greenness/Chlorophyll
Blue Normalized Difference Vegetation Index (BNDVI)	${\displaystyle \frac{\rho N2-\rho B}{\rho N2+\rho B}}$	[83]
Chlorophyll Carotenoid Index (CCI)	${\displaystyle \frac{\rho G1-\rho R}{\rho G1+\rho R}}$	[86]
Chlorophyll Red Edge (CI-RE)	${\displaystyle \frac{\rho \mathrm{R}\mathrm{E}1}{\rho R}}-1$	[87]
Green Chlorophyll Index (Clg)	${\displaystyle \frac{\rho N2}{\rho G}}-1$	[87]
Enhanced Vegetation Index (EVI)	${\displaystyle \frac{2.5\times \left(\rho N2-\rho R\right)}{\rho N2+6\times \rho R-7.5\times \rho B+1}}$	[88]
Green Chromatic Coordinate (GCC)	${\displaystyle \frac{\rho G}{\rho R+\rho G+B}}$	[89]
Green Leaf Index (GLI)	${\displaystyle \frac{2\times \rho G-\rho R-B}{2\times \rho G+\rho R+B}}$	[90]
Green Normalized Difference Vegetation Index (GNDVI)	${\displaystyle \frac{\rho N2-\rho G}{\rho N2+\rho G}}$	[91]
Green–Red Vegetation Index (GRVI)	${\displaystyle \frac{\rho G-\rho R}{\rho G+\rho R}}$	[87]
MERIS Terrestrial Chlorophyll Index (MTCI)	${\displaystyle \frac{\rho N2-\rho \mathrm{R}\mathrm{E}1}{\rho \mathrm{R}\mathrm{E}1+\rho R}}$	[92]
Modified Chlorophyll Absorption Ratio Index (MCARI)	$(\rho \mathrm{R}\mathrm{E}1-\rho R)-0.2\times (\rho \mathrm{R}\mathrm{E}1-\rho G)\times {\displaystyle \frac{\rho \mathrm{R}\mathrm{E}1}{\rho R}}$	[93]
Normalized Difference Vegetation Index (NDVI)	${\displaystyle \frac{\rho N2-\rho R}{\rho N2+\rho R}}$	[94]
NIR Green Difference Vegetation Index (GDVI)	ρN2 − ρG	[95]
Optimized Soil Adjusted Vegetation Index (OSAVI)	${\displaystyle \frac{(1.16\times (\rho N2-\rho R\left)\right)}{(\rho N2+\rho R+0.16)}}$	[93]
Photochemical Reflectance Index (PRI)	${\displaystyle \frac{\rho G-\rho G1}{\rho G+\rho G1}}$	[96]
Pan Normalized Difference Vegetation Index (PNDVI)	${\displaystyle \frac{\rho N2-\left(\rho R+\rho G+\rho B\right)}{\rho N2+\left(\rho R+\rho G+\rho B\right)}}$	[83]
Red-Edge Triangulated Vegetation Index (RTVICore)	$100\times \left(\rho N2-\rho \mathrm{R}\mathrm{E}1\right)-10\times (\rho N2-\rho G)$	[97]
Red Chromatic Coordinate (RCC)	${\displaystyle \frac{\rho R}{\rho R+\rho G+\rho B}}$	[89]
Red Normalized Difference Vegetation Index (RDVI)	${\displaystyle \frac{\rho N2-\rho R}{\sqrt{\rho N2+\rho R}}}$	[83]
Sentinel-2 Lai Green Index (SeLI)	${\displaystyle \frac{\rho N2-\rho \mathrm{R}\mathrm{E}1}{\rho N2+\rho \mathrm{R}\mathrm{E}1}}$	[98]
Blooming
Enhanced Bloom Index White (EBI_White)	${\displaystyle \frac{\rho R+\rho G+\rho B}{\frac{\rho G}{\rho B}\times (\rho R-\rho B+1)}}$	[99]
Enhanced Bloom Index Red (EBI_Red)	$(R/G)/\left(\right(G/B)\times (R-B+1\left)\right)$	[99]
Enhanced Bloom Index Yellow (EBI_Yellow)	$\left(\right(R+G)/B)/\left(\right(G/B)\times (R-B+1\left)\right)$	[99]
Normalized Difference Yellowness Ratio (NDYI)	${\displaystyle \frac{\rho G-\rho B}{\rho G+\rho B}}$	[83]
Plant Senescence Reflectance Index (PSRI)	${\displaystyle \frac{\rho R-\rho B}{\rho \mathrm{R}\mathrm{E}1}}$	[100]
Moisture
Canopy Moisture (NDMI)	${\displaystyle \frac{\rho N2-\rho \mathrm{R}\mathrm{E}1}{\rho N2+\rho \mathrm{R}\mathrm{E}1}}$	[101]
Normalized Difference Water Index (NDWI)	${\displaystyle \frac{\rho G-\rho N2}{\rho G+\rho N2}}$	[102]
Near-Infrared Reflectance of Vegetation (NIRv)	$\left({\displaystyle \frac{\rho N2-\rho R}{(\rho N2+\rho R)}}-0.08\right)\times \rho N2$	[103]
Water Band Index (WBI)	${\displaystyle \frac{\rho N2}{\rho R}}$	[104]
Urban/Wetlands Moisture Index (MNDWI)	${\displaystyle \frac{\rho G-\rho Y}{\rho G+\rho Y}}$	[105]
Structure
Modified Triangular Vegetation Index 1 (MTVI1-G1)	$1.2\times (1.2\times \left(\rho N2-\rho G1\right)-2.5\times (\rho R-\rho G1)$	[82]
Modified Triangular Vegetation Index 2 (MTVI2)	${\displaystyle \frac{1.5\times (1.2\times \left(\rho N2-\rho G1\right)-2.5\times (\rho R-\rho G1)}{{\left(2\times \rho N2\right)}^2-5\times {\left(6-\rho N2-5\times \left({\rho }^{0.5}\right)-0.5\right)}^{0.5}}}$	[82]

2.8. Variable Selection Methodolgies

With <20 observations and 164 potential explanatory parameters, five methods were used in a multi-stage variable selection framework to minimize multicollinearity, overfitting and model instability. Near-zero variance variables were removed with the nearZeroVar function from the caret R package [106]. Parameters meeting a 0.9 threshold for collinearity were removed with the cor and findCorrelations functions of the caret R package [106]. With these functions, the parameter pairs were assessed for collinearity; when the correlation exceeded a 0.9 threshold, the mean absolute correlation across all parameters was calculated as follows:

$$r_i={\displaystyle \frac{1}{p-1}}{\sum }_{j\ne i}\left|r_{ij}\right|$$

(2)

where r_ij is the correlation between variable i and j, p is the total number of variables, and j ≠ i removes self-correlation metrics. The process removes the variable that is most redundant amongst all potential variables. Then correlation clusters are assessed. Variables which have a correlation coefficient of 0.8 or greater are clustered. The variable within each cluster which has the strongest correlation to the response is retained. The parameter with the larger mean absolute correlation, indicating high redundancy with all available parameters, was removed. After these steps, 33 input parameters remained which were all used for each of the following variable selection methods. Boruta, recursive feature elimination, elastic net regularization, mutual information, and hierarchical correlation clustering methods were iteratively applied to create a consensus list of variables for downstream analysis.

A random forest wrapper method, Boruta, was used to identify parameters with statistically significant importance in comparison to a set of fake variables, called shadow features, created to emulate background noise [107]. Shadow features were generated by shuffling each predictor variable value to decouple the value from the corresponding response value [108]. The expanded dataset was subjected to a random forest algorithm to determine each parameter’s relevance to the response based on variable importance. When a variable has a lower variable importance than the associated shadow feature, the variable is removed which ensures variables with discernable signal are preserved [107]. Furthermore, through a random forest algorithm, the Boruta wrapper was able to capture nonlinear effects of variables. The Boruta function of the Boruta package was used with the maxRuns parameter set to 120 [107]. From the Boruta selection method, three parameters were retained: Current Year Fruiting Season Average Soil Moisture, One Year Prior Dormant Season Average Maximum Temperature, and Bloom Defoliation Season EBI White.

Recursive feature elimination was completed with the rfe function of the caret package which recursively eliminates parameters to minimize root mean square error with a random forest-based feature ranking and prediction [106]. Recursive feature elimination was completed with a repeated five-fold cross validation over ten repeats. The number of features selected in this method started with a minimum of one and increased in increments of 1 until it reached a maximum of 30 features. The full model was also tested in this process. The algorithm retained the set of variables with the lowest root mean square error resulting in 15 retained variables.

Elastic net regularization was implemented with the glmnet package function cv.glmnet, and coefficients were extracted using the coef function from the stats R package [109,110]. Elastic net was designed for small datasets with high collinearity. The algorithm combines least absolute shrinkage and selection operator (LASSO) and ridge regression penalties to shrink coefficients associated with each parameter term to prevent overfitting. The LASSO component selects variables by forcing uninformative parameter coefficients to zero. The ridge regression component decreases the coefficients of correlated predictors to similar values. The α is a penalty parameter which balances the ridge regression and LASSO components. Here, the elastic net regularization was run with an α of 0.5, which assessed LASSO and ridge regression penalties equally. The penalty strength parameter, λ, was selected across a sequence of λ values generated automatically from the function and chosen based on the lowest cross validation error. The retained parameters are as follows: Current Year Fruiting Season Average Soil Moisture, One Year Prior Dormant Season Average Maximum Temperature, Bloom Flowering Season EBI White, and Bloom Defoliation Season EBI White.

Hierarchical clustering was performed based on the set of 33 input parameters which passed the near zero variance and multicollinearity thresholds. Hierarchical clustering offered a method to select an ecologically relevant variable from groups of variables with the goal of reducing redundancy and ensuring results are interpretable. A correlation distance was calculated as 1 − |r| using the hclust function of the R stats package [110]. Clusters are formed when sets of variables have a moderate correlation coefficient of at least 0.8. A representative parameter from each cluster was selected based on a maximum correlation with the response variable berries per m². A total of 19 parameters were retained from this analysis. Mutual information ranking is a methodology used to assess redundancy through nonlinear interactions or thresholds and not solely direct linear correlation observed in the hierarchical clustering method. Mutual information is an entropy-based filter algorithm which ranks parameter importance based on the decrease in uncertainty of the response when the parameter is known. Mutual information was deployed with the information_gain function of the FSelectorRcpp R package [111]. Twenty-three variables were retained from the mutual information method of variable selection.

Variables in at least three of the selection methods (Boruta, recursive feature elimination, elastic net regularization, hierarchical clustering, and mutual information ranking) were chosen for a consensus list of parameters totaling nine variables: Current Year Fruiting Season Average Soil Moisture, One Year Prior Dormant Season Average Maximum Temperature, Current Year Defoliation Season Average Maximum Temperature, Bloom Fruiting Season PSRI, Chlorophyll Greenup MCARI, Chlorophyll Greenup MTCI, Chlorophyll Fruiting Season CI_RE, Moisture Fruiting MNDWI, and Bloom Defoliation Season EBI_White. The environmental predictors included in this study are highly collinear, which is a common feature across ecological modeling studies. The ensemble variable selection methodology described here addressed common issues with collinearity, nonlinear association detection, and stability of selection to ensure the relevant predictors are retained. Furthermore, the integration of filter-based and machine learning algorithms reduced bias from each method to generate a set of input parameters which were likely to best approximate the response variable and simplify downstream statistical analysis.

2.9. Statistical Analysis

Mixed linear models were used to assess the impact of the presence of a plant species on berry harvest metrics. Similarly, generalized linear mixed effects models were used to ascertain the effect of percent cover of individual plant species and functional groups on berry harvest metrics. The R package lme4 was used for both assessments where site was included as a random intercept with a binomial error distribution [112]. A two-way ANOVA was used to determine the effect of year and site on berry harvest metrics. A mixed linear model was used to assess the differences in berry metrics between years and regional sites, Quinhagak and Siniq. Preliminary analysis found within-region sampling locations had no significant differences in berry harvest metrics. A Dunn–Sidak comparison was used to calculate a multiple comparison to assess region and year [113].

Two regression methods were tested to assess environmental variables’ association with berry density at each site: simple linear regression and multiple adaptive regression splines (MARS). Linear regression models were used on each input variable to assess statically significant linear relationships between each input variable and berry density. The MARS algorithm was used to develop a full model to describe berry density for the purpose of generating future predictions. Both methodologies lay the foundation to detect and monitor change in the face of anticipated environmental changes in the region. The MARS algorithm is a nonparametric machine learning regression method that statistically determines significant relationships between predictor and response variables through iterative variable selection, resulting in a non-linear function. The initial model, called a forward pass, is overfit to the data and a backward pass of the model is conducted to minimize training error and prune erroneous portions of the model to create the most parsimonious regression equation [114,115,116]. The forward model fits the data by getting all potential hinge functions with the lowest residual sum of squares. The backward pass uses a penalty function to prune terms to reduce model complexity and allow for the model to be predictive to new data. For the final model, predictor variables are split into a series of smaller regions (i.e., “hinge functions”), described by several regressions to approximate the relationship between the predictor and response variables. Each hinge function reflects the non-linearity of the predictor variables to the response variable. MARS generates the model using the form:

$$y_t={\mathsf{\beta }}_0+{\displaystyle {\sum }_{i=1}^k}{\mathsf{\beta }}_{i}B(x_{it})$$

(3)

where $y_t$ is the response variable at time t, β_i are the model parameters for the corresponding variables $x_{it}$ ranging from i = 1, …, k, and k is the number of explanatory variables. β₀ is the intercept and $B\left(x_{it}\right)$ are the hinge functions associated with each variable [115].

The advantage of the MARS regression algorithm is its flexibility to incorporate many potential variables and selectively determine which variables are most important to subsistence harvest outcomes. MARS has been used previously in other biophysical studies, including modeling relationships between soil and plant chemistries [115,117,118]. It is important to acknowledge the limitations of MARS. The algorithm arbitrarily selects one variable over the other if two or more variables are highly correlated. Hence, multicollinearity can impact final model outputs, specifically with the inspection of individual variable importance, although the predictive power of the overall model is not impacted. It is ultimately recommended to introduce a method for reducing input variables (e.g., variable selection methods described above) before running the MARS algorithm.

Each data source was aggregated at the site level as it was the most representative spatial aggregation across input data types. A complete list of all 187 variables before any filtering steps is found in Supplemental Table S1. Linear regression was implemented with the R stats package on the variables remaining after filtering for zero variance [110]. The MARS algorithm was implemented using the earth R package with default settings [119]. The MARS algorithm was used on the variables remaining after filtering for zero variance along with the reduced set of variables found after variable selection procedures.

3. Results

3.1. Indigenous Knowledge of Environmental Drivers of Berry Harvest Variability

Six themes emerged from interviews on patterns between environmental variables and atsalugpiaq harvest: (1) lack of snow and cold spells cause freezing of seedlings, (2) excessively hot and dry summers cause small or shriveled berries, (3) cold and rainy summers have limited harvest, (4) excessive wind can cause de-floriation, (5) berries are most often found on the sides of mounds where they are least affected by environment change, and (6) spring is earlier and therefore harvest is earlier.

Community members know that many environmental drivers impact berry harvest but that these common weather patterns result in a limited harvest. As Catherine Beebe stated, “If we have a lot of snow and not that many thaw cycles throughout the winter. Then as long as its not too hot in the summer then we usually do have good berry season.” Frank Mathews explains the underlying ecological impact of “when there is no snow the salmonberry seedlings, they freeze and they don’t grow,” as the snow provides insultation to the plants. Frank Mathews continues to explain that winter conditions are not the only determining factor of variable berry harvest but also if in “the summer times if it was raining too much like rain rain rain no sunshine they still freeze.” And Marget White also adds that if there is too much sun and if it is too hot in the summertime “really hot sun—cause three years ago it was really hot and there was burnt tundra and some of the berries failed” impacts harvest. One more weather variable that impacts berry harvest potential is excessive wind. Mary and Willard Church add that, “That one year when it was too windy sometimes the plants blow away and don’t grow. High wind breaks the flower and that kills the berries.” Community members know that atsalugpiaq requires just the right precipitation, sun exposure, wind exposure, and minimum and maximum temperatures each season to have good berry harvests. Additionally, the harvest period has shifted earlier in the year. Jonathon Mark shares, “We use to pick them in August. Cause our winter was over by early June back in the day. Now, it’s like March, April, May, spring.” And Frank Mathews confirms that, “We use to go in August, nowadays we go end of July picking berries we use to go in August.”

Topography is another determinate of berry density. The sides of the tundra mounds are the perfect topography to support atsalugpiaq growth. Dorthy Mark said, “especially from the sides, the sides of the hills. That is where they are picked the most… But that is how it has always been on the side of the hills and between the little, tiny valleys we call them. There is lots. On those places mostly.” The sides of the mounds have also been more resilient to the changing tundra landscape and able to continue to support atsalugpiaq growth. “On the bowl areas [of the tundra] there are not growing because the tundra is sinking (permafrost melt)… along the slopes (sides of the mounds) where it is hard soil, it’s still good. But in the marshy areas you can notice sinking and the berries around there are not growing as much as they use to,” said Jonathon Mark. The sides of the mounds can also protect berries when there is an unfavorable winter. Frank Mathews said, “Like right now if there wasn’t enough snow but the edges (sides of the mounds) of the tundra had snow they (the sides of the mounds) got berries, On the tundra there wasn’t enough berries to pick but like the sides.” Community members know that as the Tundra changes with melting of permafrost and decreased snow that the sides of the mounds are more likely to have berries.

Atsalugpiaq requires specific weather conditions to produce good harvests. Therefore if a minimum or maximum temperature is exceeded, or too much or the wrong type of precipitation occurs, the berry harvest will be impacted. As the community members indicated, additional environmental drivers of berry harvest include topography and a shifting earlier harvest season. We included these potential predictors through the use of complimentary sources of elevation data and a monthly resolved dataset of weather variables.

3.2. Variable Selection and Statistical Analysis

Linear regression was conducted on the set of input variables only excluding the variables that present with zero variance. Fifty predictor variables met the 0.1 p-value threshold (Figure 3). Of these predictors, 35 presented with R² values above 0.3. The predictors with the highest R² values were meteorological and vegetation indices-based predictors. Notably, elevation and vegetation cover did not meet the thresholds for significance.

The variable selection procedures reduced the dataset from 187 initial predictors to 9. The procedures removed highly redundant variables from the predictor set to identify the most representative set of predictors which can be used to draw conclusions about environmental variables which impact berry density. Multiple iterations of the MARS algorithm on the reduced set of variables yielded an intercept-only equation. The MARS algorithm contains internal methods to cull collinear variables thus leading to few predictors performing better than a null model. When running MARS on the set of variables excluding only zero variance variables, three predictors were selected for a model: Flowering Season MTCI, Greenup Season CCI, and Flowering Season RTVICore (Figure 4). Although Greenup Season CCI and Flowering Season RTVICore were identified in the MARS algorithm as significant modifiers of berry density, their p-values did not meet the significance thresholds (0.60 and 0.102 respectively). The univariate predictor screen indicated Chlorophyll Flowering Season MTCI was significant at the 0.1 p-value threshold with R² values above 0.3. However, the other two predictors identified in the MARS algorithm were not significant in the univariate screening.

3.3. Site Characteristics and Correlations to Berry Density

Differences among berry harvest metrics were not statistically significant between 2024 and 2025. Berry harvest metrics in 2024 showed slightly higher average berry biomass (Figure 5a). Berry harvest outcomes were biased as data collection in 2024 occurred concurrently with subsistence harvest activities leading to slight underestimations of berry metrics. On the other hand, regional differences between Quinhagak and Siniq were statistically significant. When assessing the number of berries per square meter, which is a representation of berry density, each site had similar harvest metrics, which were not significantly different (p-value > 0.57) in the two-way ANOVA. Site variation was not significant in the assessed period amongst the Quinhagak locations. Each sampling location within Quinhagak had near-identical berry density measurements, which indicate short-term stability in harvest potential (Figure 5b). The Siniq sampling sites were only assessed in 2025 and showed significantly higher berry harvest metrics (p-value < 0.05) than the sampling locations in Quinhagak for the same year. However, the Siniq berry harvest metrics observed in 2025 were not completely statistically significant from 2024 Quinhagak berry harvest metrics, as evidenced by the Dunn–Sidak multiple comparison test. Berry size was not statistically significant across locations in the 2024 dataset (Supplemental Figure S1).

Generalized linear mixed-effects models indicated the presence–absence of any plant species within a site had no significant effect on berry harvest metrics and found no significant effects (all p-values > 0.05). As presence–absence data were not a significant predictor of berry harvest metrics in the 2024 field survey, the 2025 field survey included an estimation of species-level and functional group-level fractional vegetation cover based on manual image annotations (Figure 5c,d). In 2025, graminoid and shrub functional groups had at least 60% combined presence at each location. The shrub category is dominated by Rhododendron groenlandicum (labrador tea) in the study areas. In comparison, the berry functional group and the atsalugpiaq species-specific cover were low (5–15%) across all site locations. The imbalance in plant species coverage across sites is a feature of the heterogenous landscape of the Arctic tundra. Generalized linear mixed effects models revealed that fractional vegetation cover of individual species and functional groups did not significantly predict berry harvest metrics (all p-values > 0.05). The variance associated with sampling sites, the random intercept, was 0.24 which indicates low between-site variability.

Vegetation indices such as GRVI followed a typical pattern of green-up in the spring, reaching a peak in mid-July and then slowly descending in the latter half of the growing season (Figure 6). Thus, the vegetation indices related to greenness and chlorophyll, such as GRVI, MTCI, and CCI, represent a reliable source of information for specific landscape features during the growing season. The vegetation indices within the greenness category were highly collinear. In 2024 and 2025, there were 22 and 28 satellite image acquisition dates, respectively, which met the cloud cover threshold within the growing season. This limited number of available images prevented harvest-day specific correlations between berry metrics and vegetation indices. Furthermore, not all sites are represented within each date. However, researchers were able to categorize vegetation indices to phenologically relevant periods during the growing season to make general conclusions about the behavior of atsalugpiaq over the two-year period. The phenological periods were derived from estimates of atsalugpiaq growth described in the literature. A limited number of satellite image acquisition dates were represented during the specific harvest times for each year due to the region’s high cloud cover prevalence.

Elevation relief reflected the moundedness of each sampling location. In Supplemental Figure S2, areas highlighted in yellow indicated regions slightly higher than the surrounding tundra, or a mound. Areas represented in dark blue were sections of tundra with notable depressions in the soil profile. According to traditional ecological knowledge, atsalugpiaq were abundant on the sides of mounds. Supplemental Figure S2 also highlights an important trend: the spatial variability in actual berry counts across a site. For most areas, berry count ranges from 0 to 5 within quadrats. However, small pockets within a sampling location have up to 25 berries within a quadrat.

4. Discussion

Indigenous Knowledge from the Quinhagak community is rich in nuanced data which can assist in delineating relationships between environmental change and the response of important subsistence food. Here, we utilized the oral histories of the community to inform the collection and analysis of disparate environmental data to explore and define the relationships between the sociological and ecological data sets. The themes revealed from community oral histories influenced site selection, site size and environmental variables included in the study. In two years of data collection, berry harvest metrics observed in subsistence land use areas near Quinhagak were stable and did not significantly differ. Close inspection of the timing and intensity of meteorological variables and vegetative data revealed the complimentary nature of Yuuyaraq (trans. “Our genuine way of living”) with Kangingnaurute (trans. Western science).

4.1. Co-Production Methodology Development

Oral histories contributed both unique and complimentary insights and data to the development of this exploratory research. At each step, researchers collaborated with the local research advisory team to ensure the needs of the community were met and well accounted for. This is reflected in the pivots made between the two years to assist in streamlining the research activities and reduce labor cost and time. The pivot in methodology also emphasizes the need to balance the desire of the community within a co-produced framework rather than forcing the perceived needs of data collection on the community. Furthermore, the present work has provided an additional foundation with which the Quinhagak community can begin long-term monitoring projects. Nalaquq LLC has supported workforce development in the region for the last five years. One task to support the region has been training in the use of Unmanned Aerial Systems (UAS) for search and rescue. This has invigorated the acquisition of additional UAS and UAS docking systems for the community which enhances data acquisition capabilities which will be used vegetation monitoring throughout the atsalugpiaq growing season. Additionally, the docking systems allow for a layer of automation of UAS flights. At present, at least two other communities within the Y-K Delta have requisitioned UAS for community workforce development. This provides the communities the option to expand current vegetation monitoring efforts and establish multiple long-term, multi-location comparison of vegetation change while simultaneously supporting emergent community needs.

4.2. Ground-Based Data Collection and Utility

The study presented here attempted multiple methods to build a greater understanding of variability in the atsalugpiaq plant and the surrounding microenvironment. In accordance with co-production methodologies, this necessitated a flexible strategy and data acquisition plan. The results from the 2024 field season on berry size indicated low variability and thus subsequent years necessitated a greater focus of available labor on the collection of data which would help yield more conclusive evidence of environmental impacts on berries. This further highlights not only the pressing nature of the community but also the foundational aspects of the present study in expanding both the methodological and empirical basis for similar studies in the future.

From the Alaska Department of Fish and Game, rural communities with sizable Alaskan Native populations consistently reported the ability to collect at least five gallons of berries per season [3]. When we assessed the two years of collected berry harvest data, we identified stable trends in harvest metrics. Regional variation (Quinhagak versus Siniq), rather than site specific variation (e.g., Airport, Dump, Gravel Pit), was identified as a strong predictor for berry harvest metrics (Figure 5a). In Siniq, the local community would describe the tundra as a carpet of orange [25]. The meteorological conditions which occur in Siniq, prominent, mounded areas within each location and higher snowpack in the winter, are directly linked with the oral histories present in the interview themes. When there is little to no snow in the winter, there is a lack of insulation for the ground. Reproductive tissue is formed during the growing season and freezes resulting in a lower berry harvest [32,120]. On average, our data collection supported that a higher density of berries can be found in Siniq when compared to the sites in Quinhagak. However, this result is based on two years of data and further assessments are necessary to make concrete conclusions.

Fractional vegetation cover was a labor-intensive dataset to acquire but was not significantly different between sites and regions and did not improve model performance. In a heterogeneous landscape such as that found in the Y-K Delta, this is not entirely unusual. Heterogeneous landscapes often maintain similar regional species richness and diversity, but this richness is often distributed to different species assemblages distributed across microhabitats [121]. This results in comparable functional species cover but noisy and different compositions at a species level [122,123]. Ecological observations compliment the Indigenous Knowledge which points to the mounds located within a subsistence harvest area and not the harvest area itself. Furthermore, Indigenous Knowledge indicates that the mounds are not static year-to-year [25]. As such, fractional vegetation cover was not retained in any final models generated in the analysis. The number of 1 m² quadrats, given the small size of the areas, was statistically accurate in accordance with power analyses conducted ahead of the experiment. Indigenous Knowledge suggests that future field surveys may need to focus solely on mounds and not site-wide variation to better approximate berry-specific harvest metrics. This result aligned with previous studies in arctic berries which identified elevation, vegetation indices, and winter weather parameters as important to subsistence harvest [53,54,55,56,124]. Additionally, atsalugpiaq is a perennial species where multiple years of weather can influence harvestable potential. Most studies that assess berry harvest metrics have observed harvests over at least ten years which is especially important in a subsistence plant species such as atsalugpiaq [81,125]. Future studies would greatly benefit from spreading observations over multiple subsistence seasons to ensure the full variability of harvest metrics are measured.

A key component of berry density is the pollination ecology of atsalugpiaq. Arctic berry production is influenced by pollination success and the number and distribution of male and female flowers [125,126]. In addition, arctic berry species exhibit alternate bearing, producing a high-yield harvest in one year followed by a substantially lower yield the next [124,127]. However, a significant research gap exists in understanding the primary pollinators, the percentage of successful pollination, and pollination timing of atsalugpiaq [16,128,129,130]. Capturing flower number, number of unripe berries early in the season and fully ripened berries later in the season would further ground the reliability of the harvest metrics. Furthermore, additional work is needed to ground spectral profiles of a wide variety of plant species in the heterogeneous Y-K Delta to identify species contributions to vegetation indices.

4.3. Model Performance, Indicator Selection, and Implications for the Development of Robust Prognostic Models

Model inputs were selected from the previous literature but also the oral history themes. The oral histories grounded the selection of indicators to ensure the lived experiences of local community members were captured in the model. The dataset contained more than 200 predictor variables, necessitating variable reduction procedures. Given substantial collinearity among predictors, we employed a suite of feature selection methods designed to account for correlated and potentially nonlinear relationships with the response variables. However, across methods, only a small subset of variables was consistently retained, and this subset did not improve model performance in the MARS framework relative to a null model. The outcome is consistent with prior work indicating that variable selection in high-dimensional, collinear datasets is often unstable [131,132]. By reverting to the complete set of input parameters, the MARS algorithm removed collinear variables while retaining variables with a strong correlation to the response. Two of the selected parameters were not significant in a univariate linear relationship to the response pointing. Marginal associations derived from linear methodologies may not reflect conditional relationships in multivariate or nonlinear models [133]. Due to the constraints of variable selection methods, some variables retained may not generalize in predictive contexts.

Although model performance was moderate, it was revealed that the environmental indicators used represented only one category of the features important to harvest potential. The current study identified only vegetation indices to be both significant and correlated to the response. An additional hurdle to overcome with the current dataset is the large number of predictors with a limited number of observations which restricts the ability of a model to detect signals from variables which may only confer a moderate or long-term effect on the response. Future studies would benefit from dialing into the revealed variables (chlorophyll-specific vegetation indices) as well as variables identified in Indigenous knowledge (photosynthetically active radiation and elevation) at critical times in the phenological period and dormant seasons to ascertain the specific contribution of these variables to berry success. This has been achieved in common agricultural crops such as wheat and strawberries to predict biomass and yield early in the season [60,82,134]. Collecting additional granular data within the phenological window would provide the framework to test multiple interactions of these variables in the future. Here, the goal was to ascertain whether environmental variables had a measurable impact on berry harvest metrics. This was achieved by collecting on-the-ground data on berry density present in the sampling locations and testing these harvest metrics against a suite of environmental variables.

When assessing the usefulness of data sources for weather in this region, researchers hit a significant roadblock in identifying a repository that suited the location and size of the survey locations. For example, the Quinhagak region is beyond the range of National Ecological Observatory Network (NEON) observation towers with which researchers could potentially estimate on-the-ground meteorological patterns. NASA Power, another commonly used meteorological data repository, had too high of a spatial resolution to effectively resolve individual sampling locations. As with any federally regulated airport, Quinhagak Airport (KWN) is equipped with a weathervane to track local meteorological conditions. However, the temporal coverage for this weathervane is sporadic and would also be insufficient to resolve individual sampling locations. This is one example of the limitations of working in off-road locations in the Arctic, but this has far-reaching consequences on the feasibility of addressing the research questions set forth by the communities who live in the region. Quinhagak and other areas must integrate specialized tools in a research budget to ensure variables of interest can be consistently tracked.

4.4. Usefulness of Remote Sensing Data for Monitoring Tasks of Subsistence Plant Resources

Many studies use remote sensing to reduce labor and resource inputs while simultaneously increasing automation to effectively manage landscapes. For example, Davidson et al. [39] and Macander et al. [40] identified dominant vegetation classes in northern Alaska and determined how accurate satellite-derived vegetation indices and other environmental variables were at predicting vegetation classes. Davidson et al. [39] found that a single vegetation index was insufficient to accurately predict vegetation class and 2 m spatially resolved satellite data was inadequate to represent the heterogeneity of tundra landscapes [39]. Likewise, Macander et al. [40] found that vegetation indices rather than terrestrial surveys or weather data were the most predictive features across iterations of random forest regression models [40].

MTCI and RTVICore were the dominant performance metrics found to impact berry density in the MARS algorithm. Both these metrics exhibited positive linear correlations to berry density. Additionally, MTCI was highly collinear with the fractional vegetation estimates from quadrats demonstrating chlorophyll concentration during fruiting is also related to the percentage of functional groups. Furthermore, this underscores the correlation between spectral profile and physiological health status in plants. Nutrient deficiencies and biotic–abiotic stressors impact the cellular integrity and alter plant pigment expression leading to detectable optical differences from healthy plant stands. Anderson et al. [120] noted that berries are often found on mounds. Oral history themes highlighted that the distinctive feature is the side of the mound which holds a unique profile influenced by soil type and how the side of the mound offers protection from the wind and assists in stabilizing temperature [12,135]. Elevation and snowpack have been strongly associated with atsalugpiaq harvest potential according to Yup’ik science [136]. The amalgamation of dissimilar environmental gradients gives rise to stress responses in the subsistence plants such as photobleaching of leaves as a result of high light irradiance but cooler temperatures in any given year. On the other hand, atsalugpiaq is known to present with high anthocyanin in leaves under high irradiance and temperatures. The distinct appearance of leaves with specific environmental conditions provides a pathway to detect vegetation of this species from remotely sensed tools. Furthermore, a closer inspection of vegetation color of the key species throughout the season can capture early indicators of plant performance.

Remotely sensed data is a useful tool in hard-to-access areas to approximate changes in vegetation and vegetation functional groups [137]. Increased automation and monitoring for subsistence berries is a present need. Access to subsistence use areas has been an important feature of cultural vitality for Alaskan Native communities [138]. Prior work has integrated unmanned aerial vehicles (UAVs) with traditional knowledge to identify and monitor important cultural sites in Quinhagak [69]. The connection between vegetation indices and on-the-ground conditions plots a potential avenue for future subsistence use harvest plans wherein community members can acquire UAV or satellite data at critical timepoints during the growing season of atsalugpiaq to ascertain which areas indicate high berry density.

5. Conclusions

The current study achieved an assessment of the relationship between ground-based, remotely sensed data and atsalugpiaq berry (Rubus chamaemorus) harvest in southwest Alaska. The present study contributes to the growing body of literature pointing to the adaptive capacity of Alaskan Native communities through remote observation of the environment while maintaining cultural stewardship practices. Machine learning algorithsms with multiple chlorophyll-based vegetation indices, MERIS terrestrial chlorophyll index, green–red vegetation index, and chlorophyll carotenoid index predicted berry harvest metrics. This is the first study conducted in the region that assesses berry harvest metrics with these data types. Future studies would greatly benefit from evaluating larger sites with known records of high and low berry density over a decade-long temporal scale. Two years of data proved to be foundational to begin the work to capture the variability of berry harvest metrics and their relation to important environmental variables.