Urban Land Cover Mapping Enhanced with LiDAR Canopy Height Data to Quantify Urbanisation in an Arctic City: A Case Study of the City of Tromsø, Norway, 1984–2024

Highlights

What are the main findings?

• LiDAR-derived canopy height information provides the largest and most consistent accuracy gains for LULC mapping in small Arctic cities, particularly when using Landsat.
• Landsat augmented with canopy height achieves practical equivalence to spectral-only PlanetScope in the overlap years, enabling more consistent long-term monitoring.

What are the implications of the main findings?

• Long-term, district-scale LULC monitoring in small Arctic cities can be achieved with reduced reliance on continuous high-resolution commercial imagery.
• The proposed workflow offers a reproducible basis for identifying green-to-artificial conversion hotspots to support evidence-led urban planning.

Abstract

Intensifying urbanisation in the Arctic, particularly in spatially constrained coastal and island cities, requires reliable information on long-term land-use/land-cover (LULC) change to assess environmental impacts and support urban planning. However, multi-decadal, high-resolution LULC datasets for Arctic cities remain limited. In this study, we quantify LULC change on Tromsøya (Tromsø, Norway) from 1984 to 2024 using a Random Forest classifier applied to multispectral satellite imagery from Landsat and PlanetScope, complemented by LiDAR-derived canopy height models (CHM) and building footprints. We mapped LULC change trajectories and examined how these shifts relate to district-level population redistribution using gridded population data. The integration of a LiDAR-derived CHM was found to substantially improve the accuracy of Landsat-based LULC mapping and to represent the dominant source of classification gains, particularly for spectrally similar urban classes such as residential areas, roads, and other paved surfaces. Landsat augmented with CHM was shown to achieve practical equivalence to PlanetScope when the latter was modelled using spectral features only, supporting the feasibility of scalable and cost-effective long-term monitoring of urbanisation in Arctic cities. Based on the best-performing Landsat configuration, the proportions of artificial and green surfaces were estimated, indicating that approximately 20% of green areas were transformed into artificial classes. Spatially, population growth was concentrated in a small number of districts and broadly coincided with hotspots of green-to-artificial conversion The workflow provides a reproducible basis for long-term, district-scale LULC monitoring in small Arctic cities where data constraints limit the consistent use of high-resolution image.

1. Introduction

Intensifying urbanisation is a globally prevalent phenomenon, which also occurs in the Arctic, despite generally slower population growth [1,2,3]. Small- and medium-sized Arctic cities experience infrastructure development, densification, and increasing connectivity with regional and global networks [4,5,6]. In some cases, notably that of Tromsø, Norway, these changes occur within highly constrained coastal and island environments, where development competes directly with limited green spaces and sensitive ecosystems [7,8,9,10,11]. As such, high-quality information on the long-term evolution of land-use/land-cover (LULC) is critical for analysing the environmental consequences of Arctic urban growth and for supporting evidence-based planning [12,13]. However, multi-decadal, high-resolution LULC datasets for Arctic cities remain extremely limited.

Existing global and circumpolar LULC products, such as CORINE, ESA WorldCover, and CALC-2020, can provide valuable baselines but have important limitations in Arctic urban settings [14,15,16,17,18,19,20,21,22]. Their spatial resolution (10–100 m), relatively coarse class definitions, and low temporal frequency restrict their ability to represent heterogeneous built-up patterns, green–urban mosaics, and narrow coastal margins [14,16,18,20,22].

While very high-resolution commercial imagery has been available for earlier years, its use for multi-decadal monitoring is often constrained by acquisition costs and incomplete temporal coverage for many locations. PlanetScope provides more systematic, high-frequency, high-resolution observations, but only from 2017 onwards [23,24,25]. Consequently, no Arctic city currently has a continuous, multi-decadal LULC record at spatial scales appropriate for neighbourhood-level analysis. Landsat is the only publicly available satellite archive that spans four decades with consistent multispectral observations across the circumpolar region [26,27]. However, its 30 m spatial resolution poses substantial challenges in high-latitude urban environments, where buildings, roads, tree stands, and shoreline features frequently occupy areas smaller than a single pixel [28,29,30].

Recent advances in data fusion suggest that structural information derived from airborne LiDAR, and particularly canopy height models (CHM), can provide substantial benefits in discriminating spectrally similar urban classes and improving the delineation of built-up areas [25,28,29,30,31,32,33,34,35]. Although CHM-enhanced classification has been widely studied in forested and temperate urban environments [30,31,32,35,36,37], its performance in Arctic cities has not been systematically evaluated. Likewise, the potential of OpenStreetMap (OSM) building footprints to enhance LULC classification, particularly for small cities with mixed-density urban morphology, remains under-explored in high-latitude contexts [14,15,38,39].

An additional methodological gap concerns the operational equivalence of different sensor configurations for Arctic urban environments. It is currently unknown whether Landsat imagery, when supplemented with CHM and building context, can achieve classification performance comparable to high-resolution PlanetScope imagery [28,40,41,42,43]. This question has direct relevance for monitoring Arctic cities, where high-resolution imagery is often incomplete, costly, or unavailable for earlier decades [3,20,21,22,43]. Demonstrating that Landsat integrated with auxiliary data can approximate the performance of commercial sensors would offer a more accessible and scalable solution for long-term Arctic urban monitoring.

To address these gaps, this study develops and evaluates a multi-decadal LULC monitoring framework for Tromsøya (Tromsø, Norway) covering the period of 1984–2024. Three feature configurations are tested for both Landsat and PlanetScope: (i) spectral reflectance only; (ii) spectral reflectance combined with LiDAR-derived canopy height; and (iii) spectral reflectance, canopy height, and rasterised OSM building footprints. This design allows us to (a) quantify the improvement in classification accuracy attributable to CHM and OSM features; (b) assess whether Landsat combined with CHM can deliver non-inferior performance relative to PlanetScope spectral-only data; and (c) evaluate cross-sensor consistency in class-area estimates using the overlap years 2017, 2020, and 2024.

Using the best-performing Landsat-based configuration, we then generate the first continuous 40-year LULC dataset for Tromsøya. This time series is used to quantify the dominant trajectories of change, including transitions between green and artificial surfaces, and modifications to the island coastline. Finally, we integrate LULC transitions with gridded population data (1985–2025) to examine how long-term patterns of land transformation relate to neighbourhood-level population redistribution—an essential but understudied dimension of Arctic urbanisation.

To our knowledge, this is the first study to (i) produce a multi-decadal, district-scale LULC dataset for an Arctic city; (ii) systematically evaluate the added value of CHM and OSM features for LULC mapping in Arctic urban environments; (iii) test practical equivalence between Landsat combined with CHM and PlanetScope spectral-only classifications; and (iv) link four decades of LULC change to population redistribution at district-scale. The workflow provides a reproducible and cost-effective approach to long-term urban monitoring in high-latitude regions and offers new insights into the spatial structure and drivers of Arctic urban expansion.

Overall, this study seeks to advance long-term urban LULC monitoring in Arctic cities by combining multi-source remote sensing data within a consistent analytical framework. By evaluating the contribution of structural information to classification performance and linking long-term land transformations to population dynamics, the study aims to improve understanding of the processes and spatial patterns underlying Arctic urbanisation.

2. Materials and Methods

2.1. Study Area

Tromsø is located along the Tromsø Fjord in Troms og Finnmark county, Northern Norway (69.6° N, 18.9° E). Tromsøya is an island in Tromsø Municipality, Troms, situated between Kvaløya to the west and the mainland to the east. It is the largest urban area in Northern Norway and the fourth largest city north of the Arctic Circle [44]. Because of its geographic position and role in Arctic logistics and research, Tromsø is often described as a gateway to the High Arctic [2,3,44]. Tromsø is the 12th most populous municipality in Norway, with 78,745 residents, a population density of 31.2 inhabitants per km², and approximately 10% growth over the previous ten years. The island covers an area of 21.7 km² with 42,782 inhabitants in 2024 [45]. The urban area is situated within a narrow, ice-free fjord environment, bordered by steep mountain slopes. Land-cover information indicates that the Tromsø municipality comprises a heterogeneous mosaic in which urban fabric is a substantial component alongside extensive forest cover [19,46]. The local climate is relatively mild for its latitude due to the influence of North Atlantic Ocean circulation [47]. Previous assessments further suggest that coastal areas in Northern Norway exhibit moderate-to-high vulnerability, with high-to-very-high vulnerability reported for the island of Tromsø [48,49]. In this study, the analysis is restricted to the main island, as shown in Figure 1.

The primary reason for using Tromsøya as a case study is to test whether LiDAR-derived CHM can improve urban LULC mapping from multispectral imagery in high-latitude settings where medium-resolution pixels often mix buildings, roads, vegetation, and shoreline. Norway has exceptional coverage of high-resolution airborne LiDAR, providing a strong basis for deriving CHM and better separating built and green surfaces in compact Arctic cities such as Tromsø [50]. Combined with Tromsøya’s constrained island geography and coastal mosaic, this makes it an ideal location for a transferable long-term monitoring workflow.

2.2. Data

2.2.1. Multispectral Imagery Features

As the multispectral information base for LULC classification, surface reflectance products from Landsat and PlanetScope were used to support multi-decadal change analysis and high-resolution urban mapping for Tromsøya.

PlanetScope was selected as the high-resolution reference primarily because it provides near-daily multispectral observations at around 3 m spatial resolution, and analysis ready, orthorectified surface reflectance scenes are delivered in four-band and eight-band configurations at Level 3B. This combination is well-suited to quantitative LULC classification. Although very high-resolution commercial imagery has been available earlier and offers finer spatial detail, its use for systematic multispectral LULC mapping across multiple dates at city-scale is often constrained by acquisition strategy and inconsistent temporal coverage, and such imagery frequently provides a more limited set of spectral bands for analysis [23,24].

Landsat Collection 2 surface reflectance products were acquired and processed via Google Earth Engine platform. Imagery from the Landsat 8–9 Operational Land Imager and Landsat 4–5 Thematic Mapper sensors were utilised [26,27]. These sensors were selected because they provide consistent, high-quality multispectral data across a multi-decadal time, enabling robust long-term LULC change analysis.

The characteristics of the selected scenes and a summary of band correspondence are presented in the Supplementary Materials, Tables S1–S3.

2.2.2. Structural Height and Building Footprint Features

To test whether structural information can improve the separation of spectrally similar urban classes, height and building footprint features were added as auxiliary predictors. Height information was derived from very high-resolution airborne LiDAR products obtained from the Norwegian national elevation data portal Høydedata [50]. Raster tiles covering the entire island of Tromsøya were downloaded, merged, and processed to derive a digital terrain model and a digital surface model, from which a canopy height model was calculated. Although the LiDAR-derived CHM represents discrete temporal snapshots (2006, 2014, and 2022) rather than a temporally continuous variable, it is assumed to capture stable structural characteristics of the built and vegetated environment that remain largely invariant over longer timescales. When integrated with the Landsat and PlanetScope time series, the CHM is, therefore, treated as a static structural constraint, and its contribution is interpreted in terms of improved class separability rather than temporal change. Details of the LiDAR data used in this study are reported in the Supplementary Material, Table S4.

Building footprints were obtained from OpenStreetMap and used to derive building-context layers for classification [14,15,39]. Footprints were filtered by building types where relevant and were rasterised to the common analysis grid to produce binary predictor layers indicating the presence of residential and industrial buildings. Data were downloaded from https://www.openstreetmap.org/#map=14/69.68109/18.95931 (accessed on 30 September 2024).

2.3. Land-Use/Land-Cover Classification

Training and validation samples were prepared to support LULC classification in a heterogeneous urban environment. Mapping classes followed CORINE Land Cover (CLC) Level 3 nomenclature and codes [19] and included water bodies, marsh land, grassland, coniferous forest, other trees, residential area, industrial area, roads and paved surfaces, and bare rocks. Class labels were assigned using the official CLC class characteristics, supported by the authors’ field observations in Tromsø in June 2025. Samples were labelled by visual interpretation of reference data, with recent years primarily supported by Google Maps imagery and earlier periods informed by a georeferenced topographic map (Topographic map 1990, M 711, 1534 3, 1:50,000) [51]. The full class rationale and local interpretation for Tromsøya are provided in the Supplementary Material, Table S5.

Significant classification errors may arise if training and validation data overlap or exhibit strong spatial autocorrelation [25,28,29,30,31,32,33,34,35,52]. To mitigate over-optimistic accuracy estimates, classification was performed using a Random Forest algorithm [52,53,54] with spatial block cross-validation implemented in Google Earth Engine. The study area was tessellated into 1 × 1 km grid cells, each assigned a unique fold identifier [29,55]. For each iteration, one block acted as the validation fold, while all remaining blocks formed the training set, ensuring that model evaluation was always conducted on spatially independent areas and preventing data leakage. The block size was selected to exceed the spatial range of local autocorrelation in the predictor stack while retaining sufficient samples per fold. To avoid class imbalance effects, training samples were stratified to ensure comparable representation across land-cover classes. In total, 36 folds were evaluated, yielding a spatially cross-validated overall accuracy of 86.79%.

As additional quality control, median class spectral profiles and within-class variability were examined for Landsat and PlanetScope. Details and plots are reported in the Supplementary Material, Figure S1.

Three feature configurations as models are tested for both Landsat and PlanetScope: (M1) spectral reflectance only; (M2) spectral reflectance combined with LiDAR-derived canopy height; and (M3) spectral reflectance, canopy height, and rasterised OSM building footprints. These models are applied to both Landsat imagery (2005–2024) and PlanetScope imagery (2017–2024), enabling assessment of the classification performance across several years.

Let p denote a pixel on the common analysis grid with spatial resolution 3 m, x(p) denotes the feature vector (predictor vector) assigned to pixel p, S denote the set of surface reflectance bands used for a given sensor and year. SR_s(p) is the surface reflectance value of bands s ∈ S at pixel p. In this study S varies across sensors and PlanetScope product versions. For PlanetScope 2024 (8 bands), S_PS = Coastal Blue, RGB, Green I, Yellow, Red Edge, and NIR. For PlanetScope 2020, 2017 (4 bands), S_PS = RGB, and NIR and for Landsat 2024–1984, S_L = RGB, NIR, SWIR1, and SWIR2.

The three models are defined as

$${\mathrm{x}}_{M1}\left(p\right)={\left[{SR}_s\left(p\right)\right]}_{s\in S}$$

(1)

$${\mathrm{x}}_{M2}\left(p\right)=({\mathrm{x}}_{M1}\left(p\right),{\mathrm{C}\mathrm{H}\mathrm{M}}_0(p\left)\right)$$

(2)

$${\mathrm{x}}_{M3}\left(p\right)=({\mathrm{x}}_{M2}\left(p\right),B_{\mathrm{r}\mathrm{e}\mathrm{s}}\left(p\right),B_{\mathrm{i}\mathrm{n}\mathrm{d}}\left(p\right))$$

(3)

here CHM₀(p) is the non-negative LiDAR-derived canopy height (in metres) aligned to the 3 m grid:

$${\mathrm{C}\mathrm{H}\mathrm{M}}_0\left(p\right)=\mathrm{m}\mathrm{a}\mathrm{x}({\mathrm{C}\mathrm{H}\mathrm{M}}_{\mathrm{r}\mathrm{a}\mathrm{w}}\left(p\right),0)$$

(4)

B_res, B_ind rasterised binary building-context layers derived from OSM building footprints, aligned to the same 3 m grid. Both masks were produced by painting the filtered OSM footprint polygons into an empty image and applying unmask(0) so that background pixels are explicitly set to zero.

Model performance was evaluated using the confusion matrix (including producer’s and user’s accuracy per class), per class, macro- and weighted F1-scores, and balanced accuracy [55,56,57,58]. LULC classification was performed in Google Earth Engine, and the full processing scripts are available via the links provided in the Supplementary Materials.

To quantify categorical change between consecutive reference years, a vector-based post-classification comparison workflow was applied. The resulting intersection polygons contained both from and to class attributes, enabling the computation of transition matrices and area statistics for each interval [59,60,61]. This workflow was used to quantify the transitions from green to artificial surfaces and the mapping of change hotspots.

2.4. Population Change Analysis

Residential population was derived from the Global Human Settlement Layer population grid (GHS POP, R2023A), which reports the number of residents per grid cell and provides estimates for 1985 to 2025 at five-year intervals, with projections for 2030 [62,63]. To align with the LULC control years, population surfaces for 1985, 1990, 1995, 2000, 2005, 2015, 2020, and 2025 were extracted, clipped to the Tromsøya boundary, and harmonised to the study coordinate system [59,61,63]. Tromsøya was subdivided into analysis districts based on postcode district boundaries, which were used as zoning units for district-level aggregation. Each zone was assigned a conditional District ID (1–14) for visualisation and comparison. District population totals were computed by zonal summation and converted to population shares by normalising each district total by the Tromsøya total for the corresponding year. Changes in the population share were then calculated to identify districts with relative gains or losses in residential population.

3. Results

3.1. Effect of Height and Building Footprint Features on Land-Use/Land-Cover Classification Accuracy

3.1.1. Overall Performance Across Sensors and Years

Given these sensor specific differences within class variability, we next evaluate how adding height and building footprint information translates into measurable improvements in LULC classification performance across sensors and years. We compared three feature configurations (M1, M2, M3) to quantify the marginal contribution of CHM and OSM information and to test operational equivalence between sensors (

Table 1. LULC classification performance by year and sensor (Landsat, PlanetScope). Balanced accuracy (BA, %) for models M1, M2, M3 (Equations (1)–(3)).

ΔOSM	ΔCHM	BA, %			Sensor	Year
		M3	M2	M1
+1.22	+8.52	81.96	80.74	72.22	Landsat	2024
+1.15	+3.32	85.43	84.28	80.96	Planet
+1.8	+6.47	79.56	77.76	71.29	Landsat	2020
+0.94	+2.86	82.05	81.11	78.25	Planet
+1.16	+3.77	78.33	77.17	73.4	Landsat	2017
+0.02	+4.15	82.15	82.13	77.98	Planet
	+6.99		75.81	68.82	Landsat	2014
	+6.48		72.89	66.41	Landsat	2005/2006

).

According to Table 1, CHM is the principal driver of accuracy gains, with median ΔBA(M2-M1) of 6.48 for Landsat versus 3.32 for PlanetScope, indicating a stronger compensatory effect for the medium-resolution sensor. OSM provides a small but generally positive gain. For 2017–2024, M2 and M3 based on Landsat are non-inferior to Planet spectra-only within a margin [−0.81 to +1.31], indicating practical equivalence given the cost and availability.

3.1.2. Class-Level Gains

Beyond overall accuracy, per class mean F1 trajectories provide insight into where CHM and OSM reduce confusion among LULC categories, particularly within artificial surfaces and green surfaces classes (Figure 2).

According to Figure 2a, Landsat with M2 delivers the largest gains for residential area (25.8) and roads and paved surfaces (15.1). M3 yields only a small additional improvement for residential areas (4.7). For industrial areas, adding CHM has little positive effect, whereas adding OSM (3.0) sharpens the delineation of wide roofs. Natural classes such as coniferous forest and other trees also improve under M2, while M3 provides no meaningful additional benefit for these classes. Grassland benefits from CHM, whereas marsh land shows only a weak response to CHM. For PlanetScope (Figure 2b), despite the higher spatial resolution, CHM still has a strong effect for residential areas (13.5) and roads and paved surfaces (9.7). Industrial areas improve only slightly. Coniferous forest and other trees exhibit minor fluctuations, and the combined effect of M2 and M3 is close to zero. Marsh land and grassland show moderate gains with M2. Water bodies and bare rocks change minimally for both sensors because these classes are spectrally well-separated.

3.1.3. Models’ Behaviour Through Class-Specific Errors and Visual Validation

Class-specific error analysis indicates that the main classification challenges are concentrated in a small set of class pairs with similar spectral characteristics or mixed urban morphology. The averaged class-to-class misclassification matrix, expressed as the percentage of the true class by the sensor and model, is shown in Figure 3, while detailed producer and user-accuracy values are reported in Supplementary Materials, Figure S2.

For Landsat, the baseline configuration shows the strongest confusion between residential area and roads and paved surfaces, with a substantial share of residential samples being mislabelled as paved classes and additional confusion between residential and industrial areas (Figure 3a). After adding CHM, the dominant residential-to-roads confusion is markedly reduced, whereas residential-to-industrial confusion remains the principal residual challenge. Similarly, for roads and paved surfaces, the primary error arises from mixing with residential areas, which is substantially reduced when CHM is included (Figure 3b), while the additional contribution of OSM for this class is minimal (Figure 3c). For industrial areas, height information alone does not consistently separate industrial and residential built forms. Instead, OSM provides the most evident incremental benefit by reducing industrial-to-residential mixing and increasing class purity (Figure 3c). The most persistent error-pattern among natural classes is the mutual confusion between coniferous forest and other trees, observed for both sensors, together with leakage of grassland into tree classes. Additional confusion between grassland and marsh land is observed, consistent with closely related spectral signatures and a weaker response to structural predictors. For PlanetScope, the overall error structure is similar. CHM most strongly reduces residential to roads confusion (Figure 3e), but residential-to-industrial mixing partly persists even in M3 (Figure 3f).

According to Figure 4, using the 2024 scenes as examples, we can visually assess the classification accuracy and class delineation across all models and both sensors for Tromsø. The PlanetScope scene and the national orthophoto are used as high-resolution, up-to-date references, complemented by our field surveys conducted in June 2025. For water bodies, the main lakes such as Prestvannet, Langvannet, Lillevannet, and Rundvannet are correctly mapped by all models and both sensors, as expected for a highly spectrally distinctive class. Notably, the Landsat-based models classify bare rocks along the western coastline more reliably (Figure 4d–f), plausibly due to the presence of the SWIR band. By contrast, PlanetScope models (Figure 4a–c) tend to confuse bare rocks with marsh land and industrial and residential area.

Despite spectral proximity to other artificial surface classes, M1 on both sensors (Figure 4a,d) separates large roads and paved surfaces units, runways and aprons at the airport, and shopping-centre car parks and other broad asphalted areas reasonably well. Street networks within neighbourhoods, aprons, and major car parks are visibly more coherent in M2 (Figure 4b,e), reducing mixing with industrial and residential areas. This effect is especially clear in the Landsat classifications (Figure 4d,e). M3 (Figure 4c,f) further regularises the boundaries between adjacent artificial classes.

For the industrial area dominated by large warehouse roofs in the east, the shopping complex, and parts of the UiT, The Arctic University of Norway estate, M1 performs acceptably, with a clear advantage for PlanetScope (Figure 4a,d). M2 yields the most faithful interpretation on both sensors (Figure 4b,e). Some confusions with roads and paved surfaces persist along the aprons, and the boundaries are best damped in M3 (Figure 4c,f). The residential area shows the largest step-change from M1 to M2, and built-up blocks follow the real neighbourhood outlines, particularly for the Landsat-based maps, while PlanetScope M1 is already adequate. M3 mainly adds boundary regularisation, reducing stray green pixels within the built fabric. Substantial pixel mixing among artificial surface classes remains visible in all models.

The natural classes, such as coniferous forest and other trees, are spectrally well-defined on both sensors and are generally represented consistently across models (Figure 4). Narrow urban shelterbelts remain challenging, especially for Landsat where M1 often underestimates them, and M2 only partly recovers these strips. Confusion between coniferous forest and other trees is substantial, peaking for Landsat M1. M3 reduces this uncertainty. Across both sensors, other trees and coniferous forest mixing also increases when moving between models. Our June 2025 field observations confirm the presence and distribution of coniferous forest patches across the island. For grassland (open lawns and verges), M2 and M3 improve overall accuracy, yet residual mixing with residential areas and other trees is common inside green neighbourhoods. The greatest confusion for grassland occurs with other trees on both sensors under M1. Marsh land is mapped well by all models, particularly in the Northern concentrations. Misclassifications are modest with grassland and M2 and M3 reduce isolated false detections along shadowed forest edges by a small margin.

3.1.4. Cross-Sensors Agreement in Class Areas

To characterise how differences in class areas arise, having established the most accurate configuration (M3), we next examine whether the resulting class area estimates are consistent between sensors in the overlap years. For transparency, the operational island extent in this analysis is defined by the classification mask, which includes inland water bodies, the intertidal and bare rock fringe, and land along the shoreline (Figure 5).

Figure 5a indicates that the most consistent agreement is observed for bare rocks and water bodies, while other trees and industrial areas also show generally stable differences between sensors. In contrast, the largest and most persistent discrepancies occur for residential areas, with Landsat systematically mapping a larger residential extent than PlanetScope under the common mask. Roads and paved surfaces also show substantial instability, and the direction of the sensor difference varies between years. Among natural classes, coniferous forest and grassland exhibit moderate, year-dependent variability rather than a consistent bias, while marsh land occasionally shows pronounced outlier differences.

The normalised transition matrix (Figure 5b) indicates that the strongest diagonal agreement is observed for other trees (84%), roads and paved surfaces (78%) and industrial area (67%), suggesting relatively stable class allocation between Landsat and PlanetScope for these categories. Moderate agreement is found for marsh land (56%), residential area (57%), coniferous forest (54%), grassland (51%), and water bodies (48%), implying more frequent cross-class reallocations. The most prominent systematic reallocations include coniferous forest to other trees (30%), grassland to other trees (19%), marsh land to grassland (26%), and residential area to roads and paved surfaces (18%). Bare rocks show the lowest diagonal agreement (24%), with substantial reassignment to industrial area (22%) and marsh land (20%), indicating limited cross-sensor consistency for this class within the coastal fringe.

3.2. Quantitative Analysis of Long-Term Land-Use/Land-Cover Changes in Tromsø from 1984 to 2024

The transition diagram (Figure 6) shows a clear long-term shift towards a larger share of artificial surfaces on Tromsøya.

In the early period (Figure 5a), when LULC was represented by four aggregated classes, artificial surfaces increased from 31% to 42%, while green surfaces (including grassland and trees) declined from 64% to 54%. Over the same interval, water bodies remained stable at about 3–4%, indicating that the main changes are driven by urban expansion rather than changes in the inland water extent. From 2005 onwards, the analysis is reported using the expanded nine class scheme, which becomes feasible once height information is incorporated, improving separation of structurally similar urban and vegetation types. This reveals how the urban fraction is partitioned across specific built-up categories (Figure 6b). Roads and paved surfaces form the single largest component throughout 2005–2024 (about 32–38%). The built-up fabric becomes more differentiated over time. In particular, industrial area increases from 4% (2005) to 8% (2024), while residential area remains a comparable magnitude but shows a gradual decline from 11% to 9%. Green surface area remains substantial and relatively stable in proportional terms, with other trees around 17–19% and grassland around 14–17%, while coniferous forest constitutes a smaller but persistent share (about 4–5%). Marsh land decreases from 8% in 2005 to 4% in 2024, suggesting either contraction or systematic reallocation of marginal wetland areas to adjacent vegetated classes. The direct 1984–2024 comparison (Figure 6c) confirms sustained urbanisation, driven mainly by the expansion and persistence of transport-related artificial surfaces, alongside increasing industrial land-use and declining shares of green surfaces and wetland covers.

Using a pixel-based change-detection analysis, we can clearly identify spatial hotspots of the largest changes in LULC class areas (Figure 7). Changes are evident across the whole island, but they are most pronounced in Northern Tromsøya, particularly around Hamna. Shoreline change is also apparent, in the north-eastern part of the island, the coastline has shifted mainly due to the conversion of parts of the water area into paved surfaces (Figure 7a).

The maps in Figure 7 reveal a pronounced expansion and consolidation of artificial surfaces between 1984 and 2024, with the island-wide balance shifting from 64% green and 31% artificial surfaces in 1984 (Figure 7c) to 44% green and 55% artificial surfaces in 2024 (Figure 7b). Overall, 20% of the green surface area transitions to artificial surface classes, and these conversions are spatially clustered with the largest hotspots associated with the main built-up and transport-related surfaces, consistent with long-term urbanisation across Tromsøya (Figure 7a).

3.3. Population Patterns in Relation to Land-Cover Change

This subsection examines how residential population redistribution across Tromsøya relates to long-term changes in the balance between green and artificial surfaces. Population redistribution is spatially uneven (Figure 8).

The largest relative gains in population share are concentrated in District 8 (133%) and District 11 (121%), with a secondary hotspot in District 12 (60%). The embedded LULC bars indicate that districts with the strongest population-share gains generally coincide with a larger shift towards artificial surfaces between 1984 and 2024. Spatially, these population-share gain districts broadly align with the main green-to-artificial conversion hotspots mapped in Figure 7a, particularly in areas characterised by recent development footprints.

4. Discussion

4.1. Classification Uncertainties and Their Influence on the Results

Some uncertainty in the mapped urban LULC patterns is likely to arise from differences in the native spatial resolution of the input datasets and the need to harmonise all predictors to a common 3 m analysis grid. Resampling Landsat to the PlanetScope grid improves geometric alignment for comparison but does not add spatial detail, so boundaries, narrow linear features, and fine-scale urban mosaics may remain affected by mixed-pixel effects in medium-resolution imagery [32,33,34,35,36,37,38]. Additional uncertainty may be introduced by the aggregation of LiDAR-derived canopy height information from sub-metre products to 3 m, which can smooth small objects and dampen local height extremes in heterogeneous neighbourhoods [40]. Residual confusion among built classes may also reflect heterogeneous materials, shadowing, and limitations in vector to raster alignment for building footprint layers, rather than height information alone [41,42,43]. Finally, it is important to consider temporal misalignment of ancillary layers with specific mapping years. For example, LiDAR is available for discrete years, while OSM reflects the state of the built environment at the time of contribution and updating. This may locally affect class estimates in rapidly changing areas, but it does not change the overall interpretation of the dominant trends supported by accuracy metrics, error structures and map-based evidence.

Some of these issues were addressed by using spatial block cross-validation, testing multiple feature configurations, and interpreting change patterns in the context of class-specific performance diagnostics and map-based evidence. Future work could further reduce uncertainty by incorporating temporally matched structural data where available, applying additional shadow and illumination normalisation, improving footprint quality control, and expanding reference data with independent ground truth or higher frequency validation for rapidly changing urban areas.

4.2. Comparison with Other Similar Studies

The regularities obtained in this study are consistent with previous work on the fusion of optical data and LiDAR, where height features (CHM) systematically improve thematic classification quality, especially for structurally complex urbanised classes. In particular, in studies [13,28,29,30,31,32,33] the authors note that spectral features are insufficient for classifying trees in a city because of the high spectral similarity of species, whereas integrating CHM substantially increases accuracy, and CHM itself acts as one of the most influential predictors. Thus, for the task of mapping shrub communities [35,41] and vegetation area [29,53], the authors applied models based on Landsat predictors, SPOT and PlanetScope that were trained on CHM, and which effectively distinguish shrubs, young forest, and other vegetation from other classes. In study [13,38], the authors used the integration of vegetation indices and CHM and improved the separation of built-up and vegetated types, refining change analysis in the context of urbanisation and greening. This directly supports the interpretation of our results, since Landsat in combination with CHM compensated for the mixed-pixel effect (Table 1) and increases separability of structurally different but spectrally similar classes. In particular, the grassland class was well-separated from the coniferous forest and other trees classes (Figure 3). Unlike many previous studies that evaluate fusion mainly within a single year or a single sensor [28,29,30], in our work the CHM effect is traced across several years and linked to the practical issue of the reproducibility of long-term monitoring. It is shown that Landsat configurations with CHM in the overlap years provide practically equivalent quality to the spectral PlanetScope model, which strengthens the argument for reliable mapping under conditions of limited availability of high-resolution imagery. In addition, the analysis of cross-sensor area consistency and the normalised transition matrix allow interpretation not only of metric gains, but also of the nature of systematic reallocations of area between classes, which is less often considered in studies focused only on accuracy metrics.

It is also important to note that a wide range of ready-to-use LULC products is currently available at global and regional scales, including datasets that cover the Arctic. For instance, the OSM Landuse/Landcover dataset [14,15] provides an OSM-informed classification framework broadly comparable to Level 2 of the CORINE Land Cover legend and can be readily integrated into standard GIS workflows. It is derived from Sentinel-2 imagery using a deep-learning model trained on OSM land-use and land-cover features. It is a classification of Sentinel-2 imagery using a deep-learning model trained on OSM land-use and land-cover features [15]. The ESA WorldCover product [16,17] offers global land-cover maps for 2020 and 2021 at 10 m spatial resolution, developed and validated in near-real time using Sentinel-1 and Sentinel-2 observations. The Sentinel-2 10 m Land Use/Land Cover product [18], generated using artificial intelligence and large-scale human-labelled training data, provides predictions for nine classes and includes an annual time series for 2017–2021 available via the Living Atlas. In Europe, the CLC inventory provides a pan-European land-cover and land-use dataset with 44 thematic classes and is updated at six-year intervals, with the most recent update reported for 2018 [19]. For the circumpolar domain, the Arctic LULC product for circa 2020 (CALC-2020) integrates Sentinel-1 polarisation, Sentinel-2 surface reflectance, and ArcticDEM topographic variables within a machine-learning framework to map dominant biophysical components at 10 m resolution using ten generic classes [20,21]. In addition, recent work has demonstrated Landsat-based machine-learning mapping of urban change in Tromsø, including impervious surface dynamics over 1993–2023 [22].

These products provide valuable broad-scale context, but within the scope of this study they are best viewed as complementary rather than as a substitute for local mapping. They are typically optimised for large-area applications and differ in nomenclature, analysis masks, and temporal focus, which complicates direct comparisons in a small, highly heterogeneous urban setting. For Tromsøya, narrow transport elements, fragmented built fabric, and mixed urban materials (Figure 2, Figure 3 and Figure 4) can lead to generalisation or heightened sensitivity to shadowing and mixed pixels in products with coarser thematic or spatial representation. Moreover, most ready-made maps focus on single contemporary years or relatively short time series [14,15,16,17], whereas reconstructing 1984–2024 dynamics requires a consistent retrospective framework. In this sense, combining the long Landsat archive with height information enables the interpretation of long-term transitions between green and artificial surfaces and supports finer differentiation of urban components in recent years (Figure 6 and Figure 7).

A relevant local comparison is the study [22], which mapped urban change in the Tromsø region using a binary impervious and non-impervious Landsat-based approach for 1993–2023. Despite differences in approach and spatial extent, both analyses indicate sustained growth of artificial surfaces and a stronger change signal in the later period, consistent with our trajectories (Figure 6) and spatial patterns of conversion (Figure 7). However, our multi-class LULC nomenclature enables urbanisation to be decomposed into roads and paved surfaces and residential and industrial areas, allowing a more detailed interpretation of the processes underlying change (Figure 7). In addition, the district-level population analysis shows that population-share gains are concentrated in a small number of districts (Figure 8) and align with hotspots of green-to-artificial transitions (Figure 7a), extending ISA-based interpretations by linking LULC change to demographic redistribution at an urban scale.

Although this study focuses on Tromsøya, the proposed workflow is readily transferable to other small- and medium-sized Arctic cities that face similar constraints, including limited data continuity and spectrally complex urban environments. The reliance on long-term Landsat archives, complemented by structurally informative datasets such as LiDAR-derived CHM and open building footprint data, makes the approach adaptable to contexts where high-resolution imagery is temporally sparse or economically prohibitive. As LiDAR coverage and ancillary geospatial datasets continue to expand across the Arctic, the framework provides a scalable basis for comparative analyses of urban change across high-latitude regions.

4.3. Population Redistribution and the Shift from Green-to-Artificial Surfaces

Comparing the pixel-based hotspots of LULC change (Figure 7a) with district-level demographic redistribution (Figure 8) reveals a broadly consistent spatial signal of urbanisation. Areas with the most intensive transitions from green-to-artificial surfaces in Figure 7a are concentrated in the north of the island and along key development and transport corridors, and these same districts in Figure 8 tend to show the largest gains in population share, suggesting a link between expansion or intensification of the built environment and the concentration of residents in a small number of districts. At the same time, some pronounced LULC transitions may be only weakly reflected in demographic change, while population gains can also arise through densification within already urbanised areas without large net losses of green space. Together, these aligned and partly decoupled patterns indicate that urbanisation on Tromsøya is shaped by two parallel processes: (i) spatial expansion of artificial surfaces at the expense of green areas in distinct hotspots, and (ii) demographic redistribution between districts that does not necessarily require proportional growth in built-up area but may reflect changes in housing density and LULC function.

5. Conclusions

This study developed and evaluated a transferable remote-sensing workflow for multi-decadal, district-scale urban LULC monitoring in small Arctic cities. LiDAR-derived canopy height models provide the most robust and consistent improvement in urban LULC classification, enabling Landsat-based mapping to achieve performance comparable to spectral-only PlanetScope and supporting reliable long-term monitoring with reduced dependence on continuous high-resolution commercial data. Cross-sensor comparisons further showed that agreement in mapped class areas is class-dependent, underscoring the importance of accounting for sensor-related effects when interpreting long-term trajectories.

Applied to Tromsøya, the resulting time series indicates that long-term, neighbourhood-scale urban change in Arctic cities can be quantified more consistently when multispectral archives are strengthened with structural and contextual predictors. The integration of Landsat with LiDAR-derived canopy height enables more reliable delineation of green-to-artificial conversions, infrastructure-related expansion, and localised shoreline modification than spectral information alone, particularly in heterogeneous coastal settings. In addition, the workflow supports reproducible hotspot-based monitoring, providing an operational evidence base for identifying where land transformation is concentrated and for prioritising planning responses such as protecting and reconnecting green corridors, targeting compensatory greening, and managing densification pressures using comparable indicators across years rather than single-year map interpretation.