Assessing Relationships Between Land Cover and Summer Local Climates in the Abisko Region, Northern Sweden

Abstract

Climate warming impacts arctic and subarctic lands, subjecting it to a generalized rise in soil temperature and causing changes in the surface cover. Land cover is a key control parameter for soil hydrothermal states, and its study by satellite imagery is necessary for monitoring boreal surface changes over time at large scales. Understanding the links between land cover and environmental conditions is also crucial to anticipate the impacts of atmospheric changes on continental surfaces. Sentinel-1 and Sentinel-2 data combined with a field campaign in July 2024 were used to produce a 10 m spatial resolution land cover map in the Abisko region, northern Sweden, covering 2180 km² and including three watersheds with an overall accuracy exceeding 94%. In parallel, temperature and precipitation fields were statistically downscaled at 100 m spatial resolution using topography, ordinary kriging based on weather stations and reanalysis. The relationships between surface areas and average summer temperature–precipitation clusters reveal that the vegetation distribution closely reflects the recent atmospheric conditions with the treeline following the 10.2 °C July–August isotherm in the considered area. This study provides a spatial basis for investigating the complex atmosphere–surface interactions and for assessing the sensitivity of boreal landscapes to ongoing climate warming.

1. Introduction

Due to global warming and Arctic amplification [1,2,3], at the origin of a temperature increase about 3.8 times higher than the global average since 1979 [4,5], arctic and subarctic areas are subject to radical changes: snow cover disturbances [4,6,7,8], plant diversification [9,10,11,12,13,14], variations in soil–microbiome assemblies [15], permafrost thaw [16,17], hydrological changes and thermokarstification [18,19,20]. In this setting, the vegetation cover plays a major role in the surface, soil heat and water balance in such a way that the difference between two different vegetation types can be of the same order of magnitude as the difference between glacial and vegetated soil [21]. Likewise, the presence of vegetation affects soil albedo and surface water fluxes [22], and it also interacts with snow cover [23,24,25]. Thus, there is a need for precise and reliable land cover maps whether used as surface state indicators or as input data in predictive models [26].

In order to meet this need, satellite imagery provides data coverage of large continental areas at varying levels of temporal and spatial resolutions [27], including the polar regions [28] to monitor the polar climate [4], ice caps [29,30,31], sea ice [32], snow cover [28,33], lakes [34], permafrost [26,35] and vegetation [36,37,38,39]. In these environments subject to heavy cloud cover [28,40,41], particularly in summer when average monthly cloud cover can be as high as 85% [42], microwave remote sensing, such as synthetic aperture radar (SAR), is of primary interest [26,43,44]. Additionally, the development of machine learning methods for studying land cover means that large volumes of data can be processed efficiently [45] with a high-quality training dataset required [46]. These algorithms are widely used for image processing and supervised classification tasks. Random forest is one of the most widely used, efficient and reliable of these algorithms [47,48].

On the other hand, the vegetation itself is controlled by climatic conditions. The Arctic appears to be subject to global greening [49], leading to an increase in vegetation, mainly shrubs, expansion toward high latitudes and altitudes [9,12,13,50,51,52]. Experiments and observations have suggested that air temperature is the main driving force of these changes ahead of precipitation and soil water content [9,12,51]. This air warming could stimulate the advance of the forest in northern Fennoscandia and parts of the Russian Arctic on about one half of the tundra areas [53]. Thus, these interactions between the surface and climate change are leading to a redefinition of northern landscapes, such as the case of “borealization” of the tundra [54]. Understanding how weather and climate conditions affect the distribution of surface land cover ensures better anticipation and understanding of the changes underway.

Nowadays, pan-Arctic land cover maps, such as the Circumpolar Arctic Vegetation Map (CAVM) [39] and, more recently, a circumarctic land cover map at a high spatial resolution of up to 10 m [55] have been produced. While these products are essential for global-scale studies, their generalist aspect can limit their ability to reproduce finer variations, particularly with regard to locally specific vegetation types, with the same robustness as regionally calibrated approaches. Furthermore, although some land cover classifications have been carried out in the Abisko region using, for instance, aerial color and infrared imagery, airborne laser scanning, LiDAR and Sentinel-2 [56,57,58,59], few studies have combined SAR and optical imagery to produce high-resolution land cover maps over entire watersheds and have investigated their relationships with micro-climatic and topographical drivers. Hence, there is a need to explore the local relationships between surface cover and its interactions with its environment.

The purpose of this study is to describe precisely the land cover characteristics of the Abisko region, in Arctic Sweden, through three watersheds—Abiskojokka, Miellajokka and Stordalen—and to associate them to local morpho-climatic conditions. This work is restricted to the July–August period to ensure a minimal snow and ice cover [60], and to correspond to the peak growing season in the area [61,62]. A new land cover map based on Sentinel-1 (S1) and Sentinel-2 (S2) data was produced at 10 m spatial resolution using a Random Forest classifier, resulting in a comprehensive surface map, because radar images are unaffected by clouds or topographic shadows. In a second step, 100 m spatial resolution air temperature (T) and precipitation (P) average maps were derived from reanalysis using regional topography, ordinary kriging based on local meteorological stations and a K-means clustering algorithm. The treeline, which is a good spatial marker of ground vegetation distribution, was delineated through the use of a logistic fitting method, and its relations with T and P were explored. Topographic influences were examined, and a second Random Forest was trained based on the T, P, slope and northness to explore the sensitivity of the vegetation to environmental conditions and support the analysis.

2. Materials and Methods

2.1. Study Area, Survey Dataset and In Situ Measurements

The region of interest (ROI) for which the maps are derived (Figure 1A)—a ~2180 km² area—consists of three contiguous catchments, Abiskojokka, Miellajokka and Stordalen [57,63,64]. The two lasts are positioned around ten kilometers east of the town of Abisko (68.35° N, 18.83° E), Sweden. The catchments are located on an area of discontinuous permafrost [65]. The vegetation consists of shrub species (Betula nana, Vaccinium myrtillus, Salix, Empetrum nigrum). There are also numerous trees (Betula pubescens ssp. tortuosa), taller shrubs (Salix glauca, Salix lanata) and diverse wetlands (dominated by Sphagnum spp.) and lakes [59]. Abiskojokka is the largest watershed studied here, with an area of 565.3 km², it covers a valley surrounded by mountains, notable lakes like Ábeskojávri or Čuonjájávri, as well as the beginning of the well-known hiking track Kungsleden, “King’s path”. Miellajokka covers an area of 51.5 km², from the characteristic mountain Nissončorru in the south to Lake Torneträsk in the north, with altitudes ranging from 383 m to 1731 m [56,63]. The Stordalen site covers 16 km² to the northeast of Miellajokka with an altitude ranging from 380 m to 770 m [57,66].

According to Köppen–Geiger Climate Classification, the whole region is located in a Dfc-subarctic climate area, characterized by a cold and moist climate with long winters [60,67], inside the Scandinavian mountain birch forest and grasslands [68,69]. The territory is crossed by the road linking Abisko to Kiruna and the Iron Ore line linking Narvik (Norway) to Riksgränsen (Sweden). The sampling area (Figure 1B) covers ~487 km2 with altitudes ranging from 370 m to 1964 m [66].

The ground truth dataset is the result of a field mission carried out from 15 to 24 July 2024 across parts of the Abiskojokka, Miellajokka and Stordalen catchments, in collaboration with the Abisko Scientific Research Station (Figure 1). Polygons representing ground observation were generated in Qfield (https://qfield.org/) using two Samsung Galaxy Tab S6 Lite tablets (Samsung Electronics, Seoul, South Korea) and located thanks to GPS, GLONASS, Beidou and Galileo navigation systems. Land cover characteristics were collected during walking transects visual interpretation. As shown in Figure 2, eight classes were defined: Rock, Dry Heath, Mesic Heath, Wetland, Alpine Willow, Mountain Birch, Water and (human) Infrastructure, following the classes of [56] according to [59]. Three more classes have been defined as masks and are not part of the machine learning process: Snow, Shadows and Clouds.

The dataset consists of 966 georeferenced polygons, each of them associated with a land use class (

Table 1. Land cover classes and associated polygons/surfaces of the field mission dataset. Star (*) indicates presence of photo-interpreted pixels. 1 pixel is equals to a surface of 100 m².

Class Name	Polygons	Surface (km2)	Class Name	Polygons	Surface (km2)
Rock	108 *	0.54	Alpine Willow	106	0.15
Dry Heath	164	1.26	Mountain Birch	230	1.73
Mesic Heath	74	0.16	Water	87 *	1.68
Wetland	129	0.23	Infrastructure	68 *	0.08

), covering a total surface of 5.8 km2. No snow patches were encountered, and only the highest peaks were slightly covered. Some photo-interpreted areas were subsequently added by analyzing true-color S2 images.

For the climatic analysis, air temperature (T) and precipitation (P) data from eight meteorological stations located within the study area were used (

Table 2. Meteorological stations and their data (accessed on 15 October 2025).

Station Name	Altitude (m asl.)	Latitude (°N)	Longitude (°E)	T	P	From	To
Abisko a	388	68.3555	18.8211	☑	☑	1913	2024
Alesjaure b	750	68.1465	18.4537	☑	☑	2013	2020
Almbergasjön a	380	68.3319	19.1542	☑	☑	2017	2024
Katterjak	514	68.4202	18.1680	☑	☑	2008	2024
Latnjajaure a	982	68.3585	18.4951	☑	☑	2018	2024
Miellejohka Alpine a	685	68.3117	18.9152	☑	☑	2019	2023
Miellejohka Subalpine a	383	68.3460	18.9550	☑		2019	2024
Nuolja a	702	68.3621	18.7387	☑	☑	2019	2023

^a Abisko Research Station (www.polar.se/en/research-support/abisko-scientific-research-station/). ^b Tarfala Research Station (www.su.se/english/divisions/tarfala-research-station).

). Datasets were obtained from the Swedish Meteorological and Hydrological Institute (SMHI, https://opendata.smhi.se (accessed on 15 October 2025)) and from the Swedish Infrastructure for Ecosystem Science (SITES, https://data.fieldsites.se/portal (accessed on 15 October 2025)).

2.2. Remote Sensing, Topographic and Reanalysis Data

Four kinds of spatial data were used: Sentinel-1 and Sentinel-2 satellite data, the digital elevation model ArcticDEM, and the reanalysis Nordic Gridded Climate Dataset.

Sentinel-1 is a platform orbiting at an elevation of 693 km and carrying a SAR (C-SAR) with a 12.3 m long antenna. Its orbital repetition cycle is 12 days. C-SAR operates in dual VV/VH or HH/HV polarization. The products used for this work are VV/VH images derived from the Interferometric Wave (IW) acquisition mode, consisting of a 250 km swath at Level-1 and in Ground Range Detected (GRD) format. The resolution is 20 m × 22 m (high-resolution mode). The C-band (5.4 GHz) in which C-SAR operates enables the surface to be studied with little penetration of vegetation. The absolute location accuracy is 7 m. Seven images covering the study area (Appendix A), both ascending and descending orbits, were retrieved from the GEODES platform of the Centre National d’Études Spatiales (CNES) between 3 July and 17 August 2024.

Sentinel-2 is a mission comprising two twin visible/near-infrared multispectral imaging satellites, 180° out of phase and orbiting at an altitude of 786 km. Their swath is 290 km wide. The multispectral instrument (MSI) comprises thirteen spectral sensors with different spatial resolutions (10, 20 and 60 m) and vertical viewing. One image from the 33WXR tiles acquired on 15 July 2024 with less than 2% of cloud cover and no atmospheric correction (Level-1C) was downloaded (Appendix A).

ArcticDEM [66] is a digital surface model (DSM) of the entire Arctic (above 60° N), which was projected in the Sea Ice Polar Stereographic North system. The data were generated using stereo imagery from a combination of the Worldview-1,2,3 satellites and a collaboration with the University of Illinois, the US-NGA and the Polar Geospatial Centre. The images have a horizontal resolution of 2 m and a vertical resolution of less than 0.5 m. For the purposes of this work, 41 tiles in GeoTiff format were downloaded. Slope and aspect products were derived at the same spatial resolution using QGIS.

Air 2 m above ground temperature (°C/K) and precipitation (mm) are obtained from the Nordic Gridded Climate Dataset (NGCD), which is a 1 km spatial resolution reanalysis dataset covering Finland, Sweden and Norway since 1961 from in situ measurements provided by the National Meteorological and Hydrological Services of each of the above-mentioned countries [70,71]. Each gridded map, made available to users by applying a deterministic interpolation approach (Type 1), from 1 July to 31 August for each year between 2015 and 2024 were used.

2.3. Land Cover Classification

2.3.1. Classification of Remote Sensing Images

S1Tiling (v1.0.0) [72] was used to generate time series of calibrated, ortho-rectified and filtered Sentinel-1 images. The process is based on the SAR ortho-rectification application from the Orfeo Tool Box [73], and the resulting images are registered to Sentinel-2 L2 optical geographic reference grid. To reduce speckle while maintaining the native spatial resolution of radar images, a multitemporal filter [74] based on Quegan’s method [75] was applied. Thus, seven projected images (VV and VH) were created over the 33WXR S2 tiles.

In order to remove any false or vacant spectral signature due to the topography, we used a method inspired by the concept of ray tracing [76] and approaches to this issue [77]. Each pixel is characterized by a zenith angle $\theta$ and an azimuth angle $\beta$, depending on the satellite (S1) or sun position (S2), retrieved from the image metadata. For S1, $\beta$ is set according to the orbital trajectory under consideration, i.e., descending (~104°) and ascending (~258°). Let $x$ be the ground projected distance between the target pixel and another pixel between the target and the satellite nadir position; then, we define $h(x, \theta )$, the relative altitude of the ray using the equation below:

$$h\left(x,\theta \right)=x/tan\left(\theta \right)$$

(1)

For each point in the image, every pixel between it and the source is tested according to its relative height $A_r$. If $h<A_r$, the target pixel is inside a topographical shadow. For radar, all the pixels in the images used are merged after this process using the mean of the values, and pixels in shadows are excluded from the calculation. The image thus obtained takes into account several points of view of the same scene and is therefore richer in information. Other minor contrast corrections are applied (Appendix B).

Sentinel-2 Level-1C images were corrected from atmospheric effects with the Sen2Cor Process [78] based on the default settings [79,80]. The cloud masking algorithm Clouds2Mask [81], based on a deep-learning library trained on multiple large global datasets, was applied using the default settings. A simple snow detection based on a Normalized Difference Snow Index (NDSI) threshold set to 0.41 [82,83] was performed. Snow is considered a “near-class” because it is much more temporary.

For the classification task, a set of 78 S1 features and 18 S2 features has been defined (Appendix C). The features are grouped into categories: polarimetry, colorimetry, texture, geometry and structure. Polarimetry uses the VV and VH polarimetric amplitudes. Color describes the value of pixels in different colorimetric bases as well as the distribution of these colors around them. Texture [84] aims to transcribe the statistical distribution of the image grain in value space. Geometry looks for recurring patterns in the local environment. The value of each feature was cleaned up from extremes and then normalized using minimum and maximum values, after which missing values were replaced by the median. A feature selection was applied in order to reduce the number of features with a procedure of relevance-redundancy selection [85] (Appendix C).

A Random Forest (RF) classifier was carried out [86]. The RF, as defined by [87], is an ensemblistic learning algorithm composed of a large set of decision trees. It is particularly well suited to multiclass classification, large datasets and small training sets [88,89]. In order to increase the model’s performance, the hyperparameters were tuned [90,91] using a Random Search algorithm with a hundred random combinations and evaluated with an accuracy-score criterion on cross-validation (Appendix D). Only 80% of polygons are used as training data with a stratified split; the remaining 20% is used for a fully independent randomized testing step. In order to improve the predictive qualities of the classification, ten iterations were carried out with training of several models.

Post-classification spatial smoothing was performed for all classifications based on a 3 × 3 contextual sliding window [92] using the window’s center-distance-weighted probability maps given by the RF. The combination of S1 and S2 (S1 and S2) was computed from the S2-based product and corrected for pixels classified as cloud, the cloud’s shadow or the topographical shadow with S1-based product. The proportion $\alpha$ of radar data origin in the final synergistic classification was used to compute the final confusion matrix and associated scores. The S1 and S2 confusion matrix has been defined as the sum of both S1 and S2 weighted by $\alpha$ and $\left(1-\alpha \right)$, respectively.

2.3.2. Quality Assessment of the Land Cover

Table 3. Quality of the land cover classifications with evaluation metrics per class (Sensibility, Precision and Critical Succes Index). Snow, shadow and cloud are not represented.

Class Name	Sensitivity (%)	Precision (%)	CSI (%)
Rock	99.50	98.61	98.12
Dry Heath	88.71	95.56	85.20
Mesic Heath	0.90	3.33	0.70
Wetland	60.17	92.43	57.35
Alpine Willow	72.97	97.30	71.52
Mountain Birch	98.26	84.56	83.31
Water	99.98	99.50	99.93
Infrastructure	96.56	97.40	94.12

shows global metric scores and the sensitivity or recall (class detection capacity), accuracy (result relevance rate) and Critical Success Index (CSI) scores for each class. These scores are defined for a class $i$ in Equation (2), where TP is True Positive, FN is False Negative and FP is False Positive:

$$Sensitivity\left(i\right):{\displaystyle \frac{T{P}_i}{T{P}_i+F{N}_i}} | Precision\left(i\right):{\displaystyle \frac{T{P}_i}{T{P}_i+F{P}_i}} | CSI\left(i\right):{\displaystyle \frac{T{P}_i}{T{P}_i+F{N}_i+F{P}_i}}$$

(2)

Thus, for S2 (resp. S1), the Overall Accuracy is 94.19% (resp. 82.12%), the Balanced Accuracy is 77.00% (resp. 44.05%) and the weighted F1-score is 93.80% (resp. 79.52%). The S1 and S2 Overall Accuracy is 94.33%, the Balanced Accuracy is 77.13% and the weighted F1-score is 93.91%. More generally, an analysis of the proportion of radar-derived data contributing to the synergistic classification shows that the Overall Accuracy is greater than S2-alone classification for $\alpha$ in [0.01%, 1.54%], and it is a maximum of 94.36% for $\alpha$ near 0.01% (Appendix D). Here, $\alpha$ = 0.80% over the whole region of interest.

The Water and Rock classes have very high scores. Concerning vegetation, Mountain Birches and Dry Heaths show very high CSI scores too with 83.31% and 85.20%, respectively. Globally, scores are lower for other vegetation classes, excepting the high precision for alpine willows. The lowest scores (particularly sensitivity and CSI) are for the Mesic Heath class, which is very poorly detected. The overall quality test of the classification is summarized by the confusion matrix (

Table 4. Confusion matrix of S1 and S2 classifications. Snow, shadow and cloud not included.

True labels	Predicted Labels
	Class Name	Rock	Dry Heath	Mesic Heath	Wetland	Alpine Willow	Mountain Birch	Water	Infrastruct.
	Rock	2196	5	0	0	0	5	0	1
	Dry Heath	27	2278	27	10	6	220	0	0
	Mesic Heath	0	15	1	5	0	93	0	0
	Wetland	0	27	0	281	0	157	2	0
	Alpine Willow	0	15	0	3	216	62	0	0
	Mountain Birch	0	44	2	5	0	2940	0	1
	Water	0	0	0	0	0	0	4235	1
	Infrastruct.	4	0	0	0	0	0	0	112

). This latter shows that rocks are mainly confused with dry heaths, which also seem to be confused with mountain birches. The Mountain Birch class presents also a lot of confusion, as shown by the numerous false positive cases with other vegetation classes, leading it to be the second lowest precision.

2.4. Meteorological Parameters Maps

2.4.1. Downscaling of Coarse Reanalysis

To produce fine-scale daily summer air temperature (T) and precipitation (P) maps from coarse-resolution datasets, we applied a statistical downscaling approach using a kriging interpolation of residuals with external drift (KED) [93,94] and a kriging interpolation of residuals (KR) [95], respectively. This method was implemented for each July–August date from 2019 to 2023 using observation from local meteorological stations (Table 2), corresponding to a maximized time period for data availability. Coarse-resolution temperature maps, $T_{coarse}$ and $P_{coarse}$, were first reprojected and resampled by bilinear interpolation onto a 100 m ArcticDEM grid. As T is a continuous field varying with altitude [96,97], a linear regression was fitted to $T_{coarse}$ at 1000 m resolution such as

$$T_{coarse}\cong aDE{M}_{1000}+b \& T_{coarse}^{alt}=aDE{M}_{100}+b$$

(3)

For each date, the daily mean for T- or cumulative for p-values were extracted, and the residuals are computed as

$$T_{corr}=T_{obs}-T_{coarse}^{alt} \& P_{corr}=P_{obs}-P_{coarse}$$

(4)

Residuals were spatially interpolated using ordinary kriging with a spherical variogram model as $R_{kriged}$. This interpolation accounts for the spatial autocorrelation structure of the residuals, allowing for spatially coherent local adjustments. The final downscaled field $T_{fine}$ was obtained as

$$T_{fine}=T_{coarse}^{alt}+R_{T-kriged} \& P_{fine}=P_{coarse}+R_{P-kriged}$$

(5)

Model performance was assessed using a leave-out cross-validation by randomly selecting one to three independent station and evaluated with mean error (ME), root mean square error (RMSE), and coefficient of determination (R²). Each average map was subsequently debiased using the mean error (ME) and testing using a leave-one-out cross-validation procedure.

To characterize the spatial variability of T-P conditions, a K-means clustering algorithm (K = 3) was applied using the debiased T and P maps as inputs. The number of clusters was selected empirically to balance interpretability and intra-cluster variability, yielding three distinct climatic regimes—Cool–Wet, Warm–Dry, and an intermediate class—that capture the dominant thermal and hydrological gradients across the region. The distribution of temperature and rainfall is shown in Figure 3. The Cool–Wet cluster presents an average temperature of 7.9 °C, and rainfall is centered around a main mode of 260 mm of cumulative rainfall each summer with a global average of 254.3 mm. Warm–Dry is defined with temperatures ranging from 9.8 °C to 11.8 °C and with a global average of 11.3 °C, while precipitation indicates an average of 142.2 mm. This typology is completed by the presence of a third T-P area, Intermediate, which is characterized by a T and P distribution centered around the median and with respective average values of 9.5 °C and 211.6 mm.

2.4.2. Quality Assessment of T and P

Table 5. Comparison of daily-based interpolation methods for meteorological parameters, ±std. among dates (A) and average maps over the 2019–2023 time period (B) at 100 m spatial resolution.

(A) Daily Parameters	ME|Bias	RMSE	R2	Unbiased RMSE
T (°C)—Interpolation	−1.42 ± 15.85	2.76 ± 15.69	0.62 ± 0.25	0.77 ± 0.27
T (°C)—KR	−0.01 ± 0.66	0.89 ± 0.40	0.73 ± 0.30	0.62 ± 0.35
T (°C)—KED	0.05 ± 0.59	0.72 ± 0.40	0.78 ± 0.30	0.49 ± 0.29
P (mm)—Interpolation	1.44 ± 5.69	3.76 ± 5.40	0.34 ± 0.33	1.95 ± 2.22
P (mm)—KR	0.64 ± 2.22	1.95 ± 2.53	0.61 ± 0.37	1.31 ± 1.78
P (mm)—KED	0.05 ± 2.47	1.82 ± 2.84	0.45 ± 0.37	1.13 ± 2.01
(B) Average maps	ME|Bias	RMSE	R2	Unbiased RMSE
T (°C)	0.01	0.26	0.95	0.26
P (mm)	−26.50	28.22	0.96	9.71

(A) displays the performance assessment of the spatial interpolation methods (bilinear interpolation, simple kriging of residuals and kriging with external drift) at 100 m resolution for the daily-based downscaled T and P. The simple kriging method shows a higher absolute accuracy for both T and P compared to the simple interpolation method with a large decrease in unbiased RMSE. Also, the R² correlation values with observations are better than the ones obtained by interpolation, but the increase is slightly lower for T than for P. The KER method shows an improvement in all metrics for T only, particularly for the unbiased RMSE. For P, the KED method decreases the correlation score of results. Table 5 (B) presents the results of average temperatures and cumulative summer precipitations for the period 2019–2023 based on KED and KR methods for T and P, respectively. Both T and P long-term maps exhibit high-quality scores.

2.5. Delineate the Treeline and Birch Distribution

As the conceptual definition of the treeline varies in the literature [53,98], we adopted an operational approach, defining the treeline as a boundary between the mountain birch and the heaths classes. The distribution of trees was modeled as a logistic function, linking the elevation or temperature to birch presence, which is consistent with previous studies [99]:

$$P\left(x\right)={\displaystyle \frac{1}{1+\exp \left(-\left({\beta }_0+{\beta }_{1}x\right)\right)}}$$

(6)

Coefficients ${\beta }_0$ and ${\beta }_1$ are calculated from the LogisticRegression module of Scikit-Learn [100] trained on ArctiDEM or Temperature fields and the land cover classification upsampled to 100 m spatial resolution. The treeline line limit $x_{TL}$ is defined as the altitude or temperature at the predicted probability of birch occurrence declines below 0.25, representing a transition from a forested to a heath-dominated region:

$$x_{TL}=\left(\ln \left({\displaystyle \frac{0.25}{0.75}}\right)-{\beta }_0\right)/{\beta }_1$$

(7)

Uncertainty was estimated from a 50-iterations Monte Carlo process combining a bootstrap resampling of trees and classification-error propagation based on a replacement of pixels according to their misclassifications scores (i.e., 1-CSI—Equation (2)): 0.167 for birches and 0.058 for all other classes. The final estimate treeline is reported as the mean ± std across simulations. The slope (S), T and P cross-effects on the birch distribution are studied through another logistic regression, which is here presented for T-S:

$$P\left(x\right)={\displaystyle \frac{1}{1+\exp \left(-\left({\beta }_0+{\beta }_{1}T+{\beta }_{2}S+{\beta }_{3}T.S\right)\right)}}$$

(8)

Secondly, the classification map was resampled to 100 m spatial resolution using a land cover majority condition and a second model. The optimized RF classifier was trained with the average T and P, slope and northness (cosine of aspect) maps as predictors to study the climatic control. The analysis focused on the main vegetated areas surrounding the treeline. The dataset was randomly split into 80% training/20% testing subsets with stratified sampling. The same method was applied with both T and P alone.

3. Results

3.1. Land Cover Spatial Patterns and Their Morpho-Climatic Characteristics

Figure 4 shows the land cover (A) and the T-P areas (B) in the studied region, while

Table 6. Extent and mean ± std morpho-climatic characteristics for the eight environmental classes.

Class Name	Extent (%)	Altitude (m)	Slope (°)	T (°C)	P (mm)
Rock	33.14	1233.03 ± 171.27	16.24 ± 10.41	8.13 ± 0.73	228.04 ± 23.76
Dry Heath	38.39	937.99 ± 161.23	12.13 ± 8.97	9.38 ± 0.69	204.63 ± 24.04
Mesic Heath	0.28	719.35 ± 135.35	8.79 ± 6.26	10.31 ± 0.58	179.11 ± 18.87
Wetland	1.75	724.57 ± 202.70	5.38 ± 6.95	10.29 ± 0.86	182.10 ± 30.37
Alpine Willow	4.19	777.64 ± 92.01	13.03 ± 8.31	10.06 ± 0.39	191.35 ± 17.43
Mountain Birch	15.22	621.36 ± 134.10	9.76 ± 7.42	10.73 ± 0.57	166.55 ± 21.32
Water	3.78	876.13 ± 249.17	2.23 ± 5.25	9.64 ± 1.06	205.89 ± 35.03
Snow	1.23	1319.24 ± 128.85	18.29 ± 7.69	7.76 ± 0.55	246.83 ± 22.39
TOTAL|AVERAGE	97.98	901.16 ± 159.33	10.73 ± 7.66	9.54 ± 0.68	200.56 ± 24.15

summarizes the morpho-climatic characteristics of the eight environmental classes. The railway line and the main road are clearly visible in the upper third of the Miellajokka and Stordalen catchments. The global landscape is characterized by very specific ecological patterns, following the altitudinal gradient: a birch-dominated area first, with some wetland clusters, as in the Stordalen Mire (68.36° N, 19.05° E), a low-vegetation tundra zone and a bare/rocky soil at the highest altitudes. Ecological indices summarize the sampling area as an intermediate heterogeneous landscape (Shannon index = 1.57, Evenness = 0.68, Simpson diversity index = 0.74) with a partial equilibrium among classes and a clear dominance of certain types (rocks, dry heaths and birches).

The Rock class is the most extensive non-vegetative category, representing 33.1% of the total area. Some misclassifications are evident, including abnormally large infrastructure-like polygons mapped in high elevated rocky terrains in the northwestern mountains. The highest actual occupation class is the Rock class (1233.0 ± 171.3 m), originating mainly from the mountainous regions. The region’s famous lakes (Ábeskojávri and Čuonjájávri) are well represented, and numerous smaller lakes are scattered throughout the landscape—many at high altitude in topographical basins. The water altitudinal distribution is the most widespread with 876.1 ± 249.2 m. Some snow patches are detected in high elevation areas in the southern part of the map, most of which are associated with small water ponds. Rocky areas and snow-covered surfaces have the highest average slopes, with 16.2 ± 10.4° and 18.3 ± 7.7°, respectively, while the lowest values (2.2 ± 5.2°) correspond to water areas.

The vegetation is dominated by Dry Heath (38.4%), which is followed by Mountain Birch (15.2%). Three classes are very poorly represented: Mesic Heath (0.3%), Wetland (1.7%) and Alpine Willow (4.2%). Willows seems to be distributed near streams and ponds at moderately high altitudes and also at the forefront of the Mountain Birch treeline. Abiskojokka forms a green corridor at the bottom of the valley due to the predominance of Mountain Birch along the watercourse leading to Lake Ábeskojávri. On the other hand, Miellajokka has a relatively similar structure with birch forest dominating in the north at low altitudes and then transitioning to tundra and rocky areas as it rises towards Lake Čuonjájávri. In terms of altitude, there are two major ecological levels: Mountain Birch at low altitude (mean at 621.4 ± 134.1 m) and Dry Heath at higher altitude (mean at 938.0 ± 161.2 m). The Mesic Heath and Wetland classes are distributed between approximately 400 m and 1100 m altitude. Vegetation classes have average slopes between 5.4° (Wetland) and 12.1° (Dry Heath). The northeastern quarter bordering Lake Torneträsk, a well-known peat bog area, has a fairly high concentration of the Wetland class, at the foot of the slopes, in the flat areas. Temperature and precipitation distributions follow opposite altitudinal zonation patterns. The lowest average temperature and highest precipitation correspond to bare soil (8.1 ± 0.7 °C and 228.0 ± 23.8 mm), whereas birch trees show the opposite pattern (10.7 ± 0.6 °C and 166.5 ± 21.3 mm).

The spatial distribution of the summer T-P areas across the three catchments (Figure 4B) follows more or less the local topography, according to the interpolation method used for T. The precipitation pattern seems also to follow this altitudinal gradient. Cool–Wet is found at high altitudes, in the mountainous areas bordering the Abiskojokka basin and south of Miellajokka. Warm–Dry is located in the valley bottoms, north of Miellajokka and over the entire Stordalen area, which has a relatively uniform, flat topography at low altitude. Note the presence of an intermediate area corridor in the high-altitude valley floor of Čuonjávággi (68.27° N, 18.96° E) and the existence of a Warm–Dry zone anomaly above the 10.2 °C isotherm line in the south part of Stordalen. The same climatic–ecological pattern is observed all over the area with the warm regions associated with the vegetated classes and the cold regions dominated by the bare soils. This distribution of the land cover according to the three T-P areas is summarized in Figure 5.

Rock is virtually absent from the Warm–Dry region, whereas Mountain Birch is absent from the Cool–Wet zone. Dry Heath is slightly more prevalent in Warm–Dry than in Cool–Wet areas (~18% and ~13%, respectively). As expected, snow occurs exclusively in the coldest regions. Trends for the remaining land cover classes are less distinct: water surfaces occur in roughly similar proportions across all T–P clusters, and the apparent presence of infrastructure in cold regions results from misclassification in the highlands.

3.2. The Mountain Birch Forest Distribution

The treeline is located at 745.2 ± 0.06 m across the entire study area, specifically 752.3 ± 0.5 m for Abiskojokka and 703.28 ± 0.5 m for Miellajokka. The thermal approach places this limit at the 10.2 °C isotherm with very little variation from one watershed to another. In contrast, precipitation is more variable across the areas around the treeline with 190.9 mm for Abiskojokka and 167.4 mm for Miellajokka. Above the 10.2 °C isotherm, logistic regression linking slope and precipitation shows that the probability of birch occurrence decreases with increasing P, and this decline is even stronger on gentler slopes, as indicated by the interaction coefficient of −0.1. Conversely, temperature increases the likelihood of birch establishment, and this positive effect is amplified on steeper slopes, as reflected by the interaction coefficient of +0.5. Overall, slope is consistently associated with a higher probability of birch presence (Appendix E).

Local birch-canopy anomalies are nonetheless observed. For instance, over some east–west oriented valleys, while the south-facing slope is largely covered by forest from the valley floor up to the 10.2 °C isotherm, the north-facing slope can be almost completely devoid of birch trees and dominated by dry heaths and some alpine willow patches. Using the Potential Incoming Solar Radiation module of SAGA-GIS on a representative case in northern Abiskojokka, the cumulative incident solar radiation for July–August was estimated at 7.5 kWh·m⁻² for the south-facing slope and 6.7 kWh·m⁻² for the north-facing slope. Additional unforested patches, including areas with dead birch, were identified through visual interpretation of true-color satellite imagery. Several of these patches occur in zones characterized by high precipitation and low slope, which is consistent with the patterns described in the previous paragraph (Appendix E).

The T-P-Slope-Northness-based binary classification (Figure 6) from the second trained RF presents a global accuracy of 91.7% and a precision (resp. sensitivity) of 0.95 (resp. 0.94) for Dry Heath and 0.84 (resp. 0.85) for Mountain Birch. The feature importance analysis reveals that the air temperature contributed 54.9% of the model’s explanatory power, while the precipitation accounted for 28.7%. The birch probability increases significantly around the 9.3 °C isotherm, exceeding 0.5 at 10.0 °C. The overall distribution of birch trees is similar to that obtained with the first map of the region. The 10.2 °C isotherm is respected and clearly delineates the altitudinal limit of tree presence. Structural anomalies associated with P and topography are found—e.g., north-facing slope effects—but not all, as the effects of soil type are not implemented. The classification based on T and P alone leads to a satisfactory result (accuracy of 89.4%) but with a more uniform birch distribution. The T-autonomous classification also produces a similar distribution of birch trees, but without anomalies, while the P-autonomous version is unable to reproduce a satisfactory map, excepting some above-treeline anomalies (Appendix F).

4. Discussion

4.1. Technical Analysis and Limitations of Products

4.1.1. Land Cover Mapping

The results demonstrate the effectiveness of combining radar and optical data, combined with a robust machine learning process, to produce high-quality land cover maps. Given that Sentinel-1 data are available regardless of lighting and cloud cover conditions, they provide a valuable complement to Sentinel-2 images for long-term and/or a high temporal resolution monitoring of land cover changes. The synergy between optical and radar data enhances the reliability of land cover classification, while radar data help compensate for the topographical shadows and cloudy pixels present in optical images, and they even improved the land cover quality. This apparent paradox arises from the non-linear nature of the quality metrics computed from the combined confusion matrix. When merging both sources, errors are not simply averaged but can compensate for each other across classes. For instance, radar data provide a slightly higher precision for water areas, which reduces omission errors in this class without strongly affecting other categories. As a result, the overall accuracy and F1-score of the S1 and S2 product slightly increases even if the global performance of S1 taken alone remains lower.

Some misclassification errors are nevertheless observed. Several pixels have been labeled as Mountain Birch in areas where this vegetation type is not typically present (personal communication). Most of these are false positives, some resulting from confusion with alpine willows whose foliage may appear similar in satellite imagery. It is also possible that the Mountain Birch class itself exhibits substantial within-class variability or co-occurs with other vegetation types within the same pixel, which could further contribute to misclassification. A poorly dense foliage could explain confusions with the heath-dominated zones. Similarly, the Mesic Heath class is frequently confused with Dry Heath, which is likely because the height of low vegetation cannot be accurately resolved from the available imagery. In addition to its limited spatial extent in the study area, there are considerably more Dry Heath training samples than Mesic Heath, which may have biased the RF Classifier toward the former. Field-based distinctions between these vegetation types are also subject to uncertainty, which may explain further confusion with Dry Heath—especially given that this class sometimes exhibits radiometric characteristics similar to rocky terrain. Finally, wetland detection is challenging either because vegetation can grow sufficiently tall to obscure water-saturated soils or because the wetland class encompasses a high diversity of plant communities, increasing the within-class variability.

The case presented here is quite exceptional in that there were no clouds over the study area on 15 July 2024. Under cloudy conditions, S1 becomes the primary data source, but several limitations may arise. First, the C-band used by C-SAR can partially penetrate vegetation depending on leaf density [101], meaning that some low-stature plant communities may be difficult to distinguish. Moreover, the backscatter coefficient is strongly influenced by vegetation and soil moisture with higher water content increasing backscatter values [102]. Very small landscape features may also remain undetected at this spatial resolution. Higher-resolution SAR systems have already been used for similar purposes, including C-band RadarSat-2 [103], X-band TerraSAR-X [104] with resolutions down to 2.1 m × 2.3 m resolution, and L-band with ALOS-PALSAR [105,106] with resolutions of around 10 m × 10 m. Incorporating Single Look Complex (SLC) imagery could further improve classification, as many polarimetric variables—requiring all four polarization channels and phase information—offer enhanced discriminatory power [38,103,107,108,109].

4.1.2. Meteorological Mapping

Large-scale meteorological products for temperature and precipitation rarely provide spatial resolutions finer than 1 km. In this study, we downscaled these products by a factor of ten. The moderately high density of meteorological stations around Abisko made such downscaling feasible, although the temporal coverage of many stations is limited. With the notable exception of the Abisko station—which has operated continuously since 1913—several datasets originate from relatively short measurement campaigns (e.g., Miellajohka, Nuolja). A spatial resolution of 100 m was selected as an optimal compromise between accuracy, local representativeness, and the risk of overfitting. Incorporating in situ data to compute kriging residuals for T and P substantially improves the quality of the downscaled products. However, the spatial distribution of meteorological stations is uneven across the study area, with most located in the southern portion of the catchments, which may introduce subtle, non-visible biases in the interpolated fields. The scores also exhibit significant daily variability, which could be improved.

Average maps of T and cumulative P derived from the interpolated daily products show high R² values, reflecting the smoothing of short-term variability through temporal averaging. Incorporating DEM data improves the mean air temperature fields by capturing the elevation-dependent lapse rate. Although precipitation exhibits a similar elevation-related pattern, applying a simple linear regression does not substantially enhance the quality of the final P product. A non-linear approach or the integration of additional terrain variables such as slope or aspect may help overcome this limitation. Moreover, the use of debiased products—obtained by subtracting the mean error (ME)—reduces systematic biases and increases the reliability of the resulting maps.

4.2. Ecological Analysis of the Abisko Region in Morpho-Climatic Context

The land cover map captures the main ecological structures and topographic controls of the Abisko landscape. The moderate diversity indices (Shannon index of 1.6 and Simpson index of 0.7) further characterize this area as a heterogeneous but partially balanced subarctic mosaic, which is typical of mountain ecotones where tundra, birch forest and rocky surfaces coexist. Water bodies are the most well-detected landscape features, including small high-altitude ponds corresponding presumably to late snowmelt accumulation zones. Patches of snow are indeed also regularly detected around these flooded areas. The lowest slope values in the classification correspond either to permanent lake surfaces, to topographic depressions temporarily flooded during snowmelt or even to places with wetlands related to a tendency of rainfall accumulation, leading to soil saturation such as, for instance, in the Stordalen Mire (68.36° N, 19.05° E). Thus, the heterogeneous distribution of wetlands suggests local topographic and hydrological influences rather than purely climatic drivers or misclassifications. Conversely, the highest slopes are associated with heaths or rocky areas, reflecting their transitional nature, at the boundary between stable vegetated surfaces and steep, sparsely vegetated terrain and often linked to areas of significant runoff.

The resulting land cover classification highlights a clear ecological zonation along the altitudinal gradient, which is consistent with previous descriptions of the Abisko region [110,111]. The dominance of Mountain Birch at low elevations and Dry Heath at higher altitudes reflects the strong thermal and edaphic control over vegetation distribution. Local average T–P conditions are closely linked to specific vegetation cover types, and in general, warmer climates support taller dominant plant species. At broader scales, plant diversity typically declines with increasing altitude, mirroring gradients in temperature and precipitation [112]. Note also that willows are often found just above the birch forest areas, suggesting that they are the first line of plant colonization toward higher altitudes.

The presence of clear altitudinal patterns in plant communities is well established in the literature and has been widely discussed [112,113,114]. The treeline is thus defined as the ecological limit beyond which trees are no longer able to develop due to environmental conditions. Our results show a treeline at around 745.2 m asl, which is consistent with the ranges found by several other studies in the region (650–750 m asl. for [115] and between 600 and 800 m asl. for [116]). The air temperature–precipitation analysis indicates that the treeline seems to be limited by the thermal conditions of the atmosphere at the 10.2 °C July–August isotherm more than the elevation itself. Indeed, even if Abiskojokka and Miellajokka show distinct elevations and conditions of precipitations, the treeline stands along the same air temperature boundary. This assertion is reinforced by the results of the coarser classification based on topographic and climatic variables, which demonstrates the prevalent importance of air temperature in relation to the presence of birch trees.

Obviously, the coarser land cover does not fit perfectly the satellite-based one, and local anomalies are not all well represented, and they are even less so without topographical parameters. If summer air temperature is recognized to be the main controlling driver of the treeline [99], monitoring over time and estimating the progress of the birch limit cannot be accomplished based on T-P parameters alone, as local factors such as topography play a contrasting role depending on the regions studied [117]. Thus, slope has been shown to play a secondary role in the importance of temperature and rainfall on the probability of birch occurrence. Areas with moderate to heavy rainfall are even less favorable to tree establishment when the topography is flat. This phenomenon may be due to the fact that these environments are conducive to water stagnation, leading to soil water saturation.

Another case, found over the region, can be linked to the fact that plant growth is also constrained by the orientation of the slope (northness), which modulates the incident radiation as shown in northern Abiskojokka by the difference of 0.81 kWh·m⁻² in summer insolation between two slopes of the same valley. Other types of treeless areas throughout the landscape could be linked to periglacial features as solifluction landforms [118,119] or cryoturbated soils which contribute to reduce the above-ground biomass [120]. Additional factors as such as soil properties [121,122], wind stress [123], reindeer grazing [124], or even defoliation from insect outbreaks [125,126] might also impact the resulting tree distribution, but these fields are beyond the scope of this study.

Finally, the treeline itself has multiple definitions [53,98], and its delineation may vary depending on the criteria used. To ensure broad applicability—especially in the context of climate change—this conceptualization of the treeline should be tested and adapted across different environmental settings. It is also important to note that this distribution of mountain birch based on recent climate (five years) is only intended to highlight the ecological niche compatible with forest colonization. The establishment of birches in a tundra zone cannot instantly follow local climate variations on such short timescales. Nevertheless, the 10 °C isotherm of the warmest month, the “Köppen Line”, is well known to be the Arctic limit of the forest and, in the Abisko region, it has been shown that the treeline was closely related to the 10 °C July isotherm limit [127]. This leads us to believe that the distribution in 2024 is close to equilibrium with the climatic conditions of 2019–2023 presented here.

4.3. Global Change Implications and Environmental Feedbacks

The close relationship observed between different land use classes and local climatic conditions provides strong evidence of the potential impact of Arctic amplification on ecosystems and biodiversity. Based on spatial associations, which do not imply a direct quantification of temporal ecosystem change, we can discuss potential implications.

Previous studies have shown that the local treeline has shifted upward over the past century in response to climate warming, as illustrated through a transition from krummholz forms to erect trees of intermediate size [128]. In parallel, a significant increase in shrubs has been detected on tundra plots next to Abisko over recent three decades [111]. On a larger scale, this greening is neither uniform nor temporally stable, and some regions are subject to browning events, as it has been shown for the Scandinavian Arctic [125].

Since vegetation type is an important factor in understanding energy balances in Arctic regions [21], complex and heterogeneous impacts of global warming on the subsoil have been highlighted depending on the different plant communities of the tundra [129,130], the boreal forest [131], or simply the surface vegetation density [132]. Vegetation is therefore an important mediator of soil temperature and water content [133], particularly in Arctic and alpine environments such as the Abisko region [134]. In this context, considering the importance of land cover for soil thermal regime, accurate land cover monitoring is of paramount importance for the study of climate warming-induced changes in the Arctic and subarctic regions [135,136,137], as well as permafrost systems, which have warmed by an average of 0.29 ± 0.12 °C over the period 2007–2016 [138], including in subarctic Sweden [139,140].

In Abisko, the distribution of surface cover in high-latitude ecosystems is also closely linked to carbon fluxes [57]. Also, permafrost lands store almost 1500 gigatons of carbon, representing around twice the current atmospheric carbon stock [18,141], the release of which into the atmosphere in the form of CO₂ and CH₄ is triggered by rising temperatures [142]. Surface dynamics, particularly those of vegetation, thus impact not only the surface carbon cycle but also the volatility of underground stocks. However, counter-mechanisms exist, and if the carbon cycle is altered by the widespread warming of the tundra, increasing vegetation cover could act as a net sink for additional emissions [143].

These changes also affect ecosystem interactions. In particular, the availability and accessibility of food resources, as lichen, for reindeer grazing are also constrained by the structure of the surface and the climatic conditions [144]. A treeline rise could spatially constrain the tundra pastures necessary for local livestock farming practices. This consideration could be taken into account in the management of livestock farming practices coupled with other measures such as the management of the forests themselves [145].

Nevertheless, this shrub colonization is limited by numerous non-climatic factors, such as the availability of nutrients—particularly nitrogen [146]—semi-domesticated reindeer husbandry practices [147] or even topography, which can stabilize any advance of the treeline [148]. Ecosystem responses to warming, such as greening, are varied in nature and generally do not consist of the forest colonizing the tundra but rather specific adaptations depending on the species [149]. Global phenomena such as widespread greening cannot always account for all local variations, and the mechanisms at play are not yet fully understood [150].

5. Conclusions

We produced a new 10 m spatial resolution land cover map for a 2180 km² region near Abisko, Arctic Sweden, using synergistic Sentinel-1 and Sentinel-2 data and an optimized Random Forest classifier trained on field samples from July 2024. The resulting map shows high reliability (Global Accuracy > 94%, F1-score > 93%). We also downscaled temperature and precipitation fields for July–August (2019–2023) by combining Nordic Gridded Climate Dataset reanalysis, ArcticDEM, and local meteorological stations, yielding accurate 100 m spatial resolution climatic surfaces (R² is equal to 0.95 for T and to 0.96 for P). K-means clustering of these products identified three distinct summer climate regimes: Cool–Wet, Intermediate, and Warm–Dry. Topographic factors, particularly elevation and slope, influence vegetation patterns and the occurrence of water-saturated soils. However, summer climatic conditions exert the strongest control on land cover: warmer and drier areas consistently support greater vegetation cover. This climatic gradient is especially visible in the position of the treeline, which broadly follows the recent 10.2 °C July–August isotherm. Local deviations from this pattern likely reflect soil moisture, solar exposure, substrate properties, or other environmental constraints.

This paper demonstrates an efficient approach that integrates high-resolution open-access radar and optical satellite imagery with model outputs and field observations to generate robust land cover maps and relate them to local climatic conditions. The method remains effective under cloud cover and variable illumination, offering strong potential for repeated ecological monitoring and for assessing how surface properties respond to environmental change. This approach enables several key applications. The treeline as the birch forest distribution can be monitored through time to assess their sensitivity to shifting summer climate conditions. The strong link between vegetation type, soil temperature and soil moisture allows the land cover products to serve as proxies for spatial and temporal patterns of soil physical properties. In addition, the high accuracy of water-body detection makes it possible to relate pond dynamics to changing temperature–precipitation regimes. Altogether, the framework provides a scalable basis for long-term environmental monitoring and for analyzing interactions between climate, vegetation, and soils in subarctic landscapes.